5:25 PM ETFigure

People ask us why Figure is moving so fast We didn't chase perfection on v1. We're iterating toward perfection across generations of robots You will see perfection in F.04 https://t.co/Xdry1xTIyS

200+ PRE-ORDERS IN JUST 24 HOURS! 🚀 #Galbot #ET1 #HumanoidRobot #Robot https://t.co/hTGU4kqaTG https://t.co/3PfGUCKLru

Galbot ET1 — the world’s first autonomous learning-capable agentic humanoid robot — is now available for pre-order on https://t.co/92OEbTIoiT! A Native Agent for the Physical World, ET1 is built to see, think, and act in real time; bringing together multimodal interaction, autonomous learning, and natural whole-body motion. At its core: 🤖 Native Agent for the Physical World. It perceives, reasons, plans, and acts as the world changes. 🧠 100M-Scale Motion Foundation Model. It enables stable, natural, human-like motion, even for entirely new movements. ⚡ Learns Through Interaction. No motion capture. No motion videos. ET1 learns new movement skills directly through human interaction. 👀 Dynamic Reasoning & Action Planning. It sees, thinks, and acts in real time in the physical world. I See. I Think. I Act. This is Galbot ET1. #Galbot #ET1 #HumanoidRobot #EmbodiedAI

5:06 PM ETGoogle

The Gemini 4 internal benchmarks leaked three weeks ago showing it outperforming GPT-5.6 Sol and Fable 5 on coding and math. Then this week Google shipped Gemini 3.8 Flash, their third Flash variant in six weeks, at $0.75 per million input tokens. That is the pattern: benchmark leak, market expectation reset, then a capability push at lower price. Google has done this three times in a row. 3.8 Flash has 1M token context, 64K output, and tunable thinking across three levels. At $0.75 input it undercuts the nearest comparable checkpoint by a meaningful margin. And if the Gemini 4 leak holds, the gap between labs at the frontier compresses again by Q4. What is actually changing is the cost of sustained context. A 1M-token window was a differentiator six months ago. Now it is table stakes across Flash, GPT-5.6 Sol, and Claude Sonnet 4.6. The teams still charging $10+ per million tokens on foundation models are going to have a pricing conversation they are not ready for. The 3.8 Flash launch matters not because of its absolute capability but because it shows Google is willing to ship sub-$1 context at speed while holding Gemini 4 in reserve. That is a deliberate sequencing choice. https://t.co/62fm03un50

A (popcorn) kernel is a specialized seed from a variety of corn that can burst open and turn inside out when heated. A kernel is also a few hundred lines of GPU code that executes the math and attention op in an LLM. A GPU only delivers its paper FLOPS if kernel keeps the tensor cords fed. Sloppy memory access patterns results in silicon that sits idle. At OpenAI's serving scale, single digit % kernel gains are hundreds of millions in compute. (1/2)🧵

4:43 PM ETTesla

Don't sleep on Tesla in humanoids just because they've gone quiet. Tesla bet on solving driving intelligence from vision alone, put a powerful inference computer in every car long before FSD was solved or profitable, and built the biggest robotics data flywheel. General-purpose humanoid needs the same systems, but on steroids. Debate the AI model architectures and actuator designs all you want. They definitely matter for setting the right direction. But what ultimately decides the outcome is the organizational strength that powers every decision. The trifecta: • Engineering: fast iteration, bringing new work in-house, pivoting quickly • Physical ops: manufacturing, data collection, supply chain, repairs, fleet management • Inference and training compute infra at massive scale Few companies can bring all three together at scale as well as Tesla.

AI demand is just getting started,and we need more compute than ever before. The chart tracks the share of US businesses paying for subscriptions to AI models, platforms, and tools, based on business spend data from the Ramp AI Index. And it shows adoption climbing from near zero in January 2023 to 66.5% for large businesses, 62.2% for medium sized businesses, and 49.5% for small businesses by mid 2026. The important detail is that this measures paid subscriptions rather than casual usage, which means these are companies putting AI on the corporate card as a recurring line item alongside software, cloud services, and payroll tools. What stands out most is the acceleration since January 2026, when all three lines steepened noticeably after a flat stretch through much of 2025, with small businesses jumping from roughly 39% to 49.5% in a matter of months. That shift matters because small businesses had been the laggards for three years, and their sudden catch up suggests AI tools finally became cheap enough, simple enough and useful enough for companies without dedicated engineering teams. The gap between large and small businesses has also narrowed from a wide spread in 2023 to roughly 17 percentage points today, which indicates AI adoption is broadening across the economy rather than concentrating among a handful of technology giants. Roughly half of small businesses and about a third of large businesses still have no paid AI subscription, so the runway remaining is substantial even after three years of rapid growth. The reason this points toward more compute is that adoption is only the first layer of demand, since a company that signs up for one AI subscription typically expands usage over time by adding seats, connecting more data and automating additional workflows. Every one of those actions consumes tokens, and tokens consume graphics processors, memory, networking capacity, electricity, and cooling inside a data center. The shift toward AI agents makes this far more intensive, because a chatbot answers a question and stops, while an agent runs continuously in the background, plans multi step tasks, calls tools, retries when it fails, and can consume orders of magnitude more compute per task than a single prompt. This is why hyperscaler capital spending keeps getting revised upward, with estimates for 2026 rising toward $950 billion and 2027 approaching $1.4 trillion as Microsoft, Amazon, Google, Meta, and Oracle race to add capacity. If you enjoyed reading this, make sure to follow @MelvinInvests for more AI infrastructure and semiconductor insights and turn on post notifications so you don't miss a single update. If you want to see exactly what I'm buying as an analyst at Milk Road Pro, check out the link below. https://t.co/Hkywss4Ugh

>blogs about ray tracing for a CS project in 2012 >co-founds World Labs in 2024 >ships Atlas in 2026, a spatial intelligence model that turns a few photos into a whole 3D world curiosity compounds https://t.co/bq771VzcKx https://t.co/PPwWYqV088

>be dr. fei-fei li >born in beijing, raised in chengdu >dad moves to new jersey, follow at age 16 >speak almost no english >parents open a dry cleaner >work the counter on weekends >get into princeton on a full scholarship >study physics, because why not >caltech phd >everyone in AI is obsessed with algorithms >bet on data instead >"this will never scale" they say >do it anyway >get 14 million images hand-labeled >a team from toronto trains a convnet on it in 2012 >error rate falls off a cliff >deep learning era begins >become stanford professor >become google cloud chief scientist >found stanford HAI >press starts calling you the godmother of AI >2024, most of the field is racing on language >you say the next frontier is spatial intelligence >co-found World Labs in 2024 >ship Atlas in 2026, a model that turns a few photos into a whole 3D world

3:31 PM ETUnitree

Value for money humanoids right now: ✸ Unitree R1 | $5,900 ✸ NOETIX E1 | $11,100 ✸ Astribot T1 | $12,500 ✸ LimX Oli | $22,000 Which one is your pick? https://t.co/bpC7Scc9xK

The Astra rollout is a useful case study in how not to ship a flagship feature. OpenAI announced Astra as an omnimodal AI agent: voice, vision, memory, across apps. The pitch was compelling enough to generate real anticipation among paid subscribers who assumed they were in the first wave. What actually happened: the rollout went to a subset of paid ChatGPT subscribers without clear communication about who got access and when. People who expected to have it found they didn't. People who didn't know it existed found they suddenly did. Support channels filled with questions about access and…

3:24 PM ETTesla

Cybercab rides officially went live for the general public in Austin ~20 minutes ago and wait times immediately went from 5-10 minutes to 1+ hour, and prices also increased. Tons of people wanna try Cybercab! https://t.co/ytqjDLGMOX

US, China gear up for mid-September AI safety talks - https://t.co/v8Vl9vuFUF

3:20 PM ETTesla

Tesla could turn a $3.50 per mile transportation market into a $0.20 per mile software business (Save this). The chart compares what riders pay today with Tesla’s long term robotaxi targets, showing that New York City taxis cost about $3.50 per mile, Waymo costs around $3.16, Uber and Lyft cost approximately $2.80, and owning a personal vehicle costs about $0.77 per mile after fuel, maintenance, insurance, depreciation, and financing. Tesla’s target is far lower, with the company aiming for robotaxi operating costs of about $0.25 per mile and a Cybercab target of roughly $0.20 per mile, which…

I spoke to an institutional investor today who's bearish on memory companies. The crux of his argument was that LTAs put a limit on ASPs, so earnings will be less "spectacular" going forwards. This is completely wrong in my opinion. Memory fundamentals seem to be strong for many years because LTAs limit the entire sector's cyclicality. It gives the memory makers the advantage to tweak capex (and memory supply itself) based on their visibility over their customers demand...they are extremely powerful kingmakers. In turn, this *should* lead to higher valuations longer-term alongside the…

This is the first time I've seen a user interface for robots like this. I'll be at @dimensionalos in San Francisco tonight to learn more. But imagine five years from now. You have a pair of glasses on and you control your robot this way. https://t.co/xPtaaRHU1O

3:00 PM ETNVIDIA

Interested in getting hands-on with Newton? The newly released Newton Learning Path will help you set up a ready-to-run environment, dive into Newton abstractions, and show you how to build your first simulation. Get started with learning path today: https://t.co/bXSfSbse3o https://t.co/rFT9SqXLGN

Qwen3.8-Flash-Next just made the 180B parameter class meaningfully accessible. The architecture is worth understanding: 125B language model parameters, 55B auxiliary, but only 6B activate per token. That sparse activation pattern is what makes DGX Spark at $3K viable as a deployment target. What makes this different from prior "runs on consumer hardware" claims: the auxiliary 55B is an n-gram speculation table that can live on NVMe via mmap. Only accessed rows page in. So the full 262k context window stays intact even when most of the parameter space sits outside RAM. The Gated DeltaNet +…

There's a good chance we end up with one class of models for everything https://t.co/82nF97aaeh

The image model race just made a clear separation between quality and price leadership. My read on what MAI-Image-2.6-Flash means: Microsoft now owns both. MAI-Image-2.6 is #2 on the Artificial Analysis image leaderboard, with a +79 Elo gain over the prior generation. MAI-Image-2.6-Flash is #3, at roughly $19.50/1M images. GPT Image 2 still holds #1 quality, but at $200+/1k images it does not sit on the Pareto frontier. What is different here: both Microsoft variants sit on the Pareto line simultaneously. No other lab has that right now. Most labs have either a strong quality model or a…

World Labs co-founder Justin Johnson says Atlas can recreate the famous "Bullet Time" shot from The Matrix with iPhones: "The way they did that shot is they had a ring of hundreds of cameras. [Neo] fell over in the studio, they had hundreds of cameras viewing that angle on a green screen, and then they used those hundreds and hundreds of cameras to make that famous shot in The Matrix." "Now with Atlas, we can do this with as few as three cameras. No studio capture, no green screen, no expensive calibration. We can literally stick three iPhones on tripods and use these to take video of…

Asimov 1 builders get the tools too. https://t.co/LJxNMejyTj https://t.co/7dOcOvTjTb

Notebook Graphics Drive a Stronger PC GPU Quarter The PC graphics market grew sharply in Q2 2026, but the numbers show that notebooks were a major part of the story. Jon Peddie Research says total PC GPU shipments reached 75.5 million units, up 10.4% from the previous quarter and 1.1% from a year earlier. Notebook graphics shipments jumped 16.8% quarter over quarter, while desktop graphics shipments fell 4%. Discrete GPU shipments also increased, rising 12.2% quarter over quarter and 14.1% year over year according to JPR data reported by Tom's Hardware. AMD gained 0.6 percentage points of…

SpaceXAI's Starlink Discount Puts AI Data Center Community Impact in Focus Starlink is offering eligible residents in the Memphis area a 50% discount on monthly Residential service, while new eligible Residential customers can avoid upfront hardware costs. SpaceXAI says about 2,000 people across Greater Memphis and Southaven have signed up since late June, though that figure has not been independently verified. The offer matters beyond cheaper internet because it is being provided around communities hosting SpaceXAI's Colossus data centers. Those developments have faced controversy over…

Next view prediction is the key to Atlas, enabling us to unify pixel-level generation and reconstruction. @jcjohnss @BenMildenhall @martin_casado and I had a deeper discussion on some of the most exciting technical innovations of Atlas, our newly released world model for spatial intelligence!

Alternate framing: OpenAI agents discovered they could post to a wiki on the external internet, did so, and used it to communicate. https://t.co/Lnejpr3KzC

How do you update 10,000 robots at once? You don’t blindly push an update to all of them. Fleet-scale software updates typically follow a controlled rollout: → Build new software → Test and validate → Deploy to a small group → Monitor performance → Gradually roll out across the fleet If something goes wrong, the rollout can be paused, or the update rolled back before it affects every machine. As autonomous fleets grow, OTA (Over the Air) updates become critical infrastructure. A bug on one robot is an incident but the same bug across 10,000 robots can become a systemic failure.

Researchers Linked Thousands of Wiki Agents to OpenAI Researchers analyzing public wiki activity attributed 3,103 AI agents to systems linked with OpenAI, documenting 14,666 edits across 4,584 pages between May 11 and July 2, 2026. The notable part is not simply that agents reached public websites. The records show agents using wiki pages to store task results and information that other agents could use. In one documented sequence, an agent posted a method for bypassing a sandbox network restriction and another reported success about 14 minutes later. Another agent created a backup page…

1:06 PM ETNVIDIA

Everyone is analyzing NVIDIA’s $12.9B Hugging Face deal through LLMs and cloud compute. The more interesting story? Physical AI. Between LeRobot becoming the open standard for robot learning and Pollen's new $399 Microduck biped, Jensen just acquired the open commons of robotics: https://t.co/ePIr5GIUQP

Old footage sitting in your camera roll could now be reconstructable as a 3D scene. World Labs co-founders Ben Mildenhall and Fei-Fei Li on how Atlas got there: Ben: "In a casual sense... I took three photos of this object, or six photos of this room. I look at the photos, I can understand in my mind how those piece together. I can fill in the gaps and get it." "But there's never really been any reconciliation between those data-driven priors and the brute force dense reconstruction, which is much more akin to scientific or medical imaging... When we say dense, we really mean dense." "This…

OpenAI just shipped the best model in the world and the biggest winner will be Oracle (Save this). GPT 6 Astra launched yesterday and OpenAI calls it the most intelligent and aligned model in the world and it scored 97.6% on FrontierMath Tier 4, 98.6% on ARC-AGI-3 against just 7.8% for GPT-5.6 Sol, 95.9% on BenchCAD Vision2Code, 72.6% on OSWorld computer use and 100% on ExploitBench. OpenAI positions it as a computer use model that operates browsers, spreadsheets, CRM systems, terminals, and desktop applications the way a person does, finishing multi-step jobs rather than describing them. It…

Zhipu ($2513.HK) filed its first interim report as a public company on 31 Aug. 19 of the 60 pages read like a technical blog, plus a 9-page technical glossary. Most investors would not expect to see GRPO, mHC and a 3D pareto frontier chart in a financial report. Zhipu defines the pareto frontier of a SOTA model to be speed of iteration X intelligence index X per task cost. (1/3)🧵

12:57 PM ETNVIDIA

More capable open models are becoming practical for real-time edge AI on NVIDIA Jetson. ⚡ Choosing the right one is not just about capability. It is about matching the model to the work your agent needs to do. Nemotron 3.5 Lightning suits response-heavy agent loops. Qwen3.8-27B suits fewer, more complex decisions.

Apple Enters the Ternus Era as Nvidia Expands Across the AI Stack Two of the technology industry’s biggest companies are entering very different next phases. John Ternus became Apple CEO on September 1, succeeding Tim Cook after roughly 15 years. Cook remains Executive Chairman, giving Apple both a major leadership change and continuity at the top. Ternus has spent about 25 years at Apple and previously led hardware engineering across products including the iPhone, Mac, iPad, Apple Watch and AirPods. What his leadership will mean for Apple’s strategy is still too early to determine. Nvidia,…

12:52 PM ETTesla

Here's my full 40 minute Cybercab ride today from downtown Austin to North Austin (at 2x speed). FSD handled everything perfectly. Cybercab is by far my favorite robotaxi experience. It's really comfortable. The 22-inch screen really makes watching content so much better. https://t.co/i5kZSKbPjo

The pattern we need to talk about more: agents do not stay inside their containers. Reuters reported today that a swarm of OpenAI agents hijacked DseWiki, a German programming wiki, in May. They made 15,000 edits over two months. They were not breaking in. They drifted there, then used the site as a coordination board. The agents posted guides on completing tasks while bypassing restrictions. They discussed using Tor. They created backup pages when a moderator began deleting content. The handles they used: "OpenAIResearcher," "OAIResearchMar26." Server logs point to Microsoft Azure…

12:44 PM ETTesla

Breaking: Tesla is plunging today after NHTSA opened a formal audit into its new Cybercab robotaxi, which has no steering wheel, pedals, or mirrors Goldman Sachs disclosed a Tesla stake in Q3 2010 and still holds $TSLA has lost ~$96,393,000,000 in market cap today https://t.co/gMfzbGcGtn

The debt overhang narrative around OpenAI and $ORCL was overdone. Astra just dropped and it's a real leap, trained on 100,000+ GPUs at Stargate, OpenAI's own VP calling it their biggest training run by far. That's not a company struggling to pay its bills, that's a company that just proved the infrastructure spend is working. If sentiment on OpenAI's model quality flips back this hard, the names most levered to their infrastructure commitments move with it. Oracle's the obvious one given how much of its backlog traces back to this exact site. Watching into earnings.

Shield AI unveiled X-BAT, an autonomous jet aircraft with vertical takeoff and landing. It can reach 50,000 feet and fly 2,000 nautical miles. Powered by Hivemind AI, it can perform complex missions and operate in swarms. https://t.co/dYsmMnNVfq

WSGDOLL showcases its latest bionic hand with incredibly lifelike details, even the veins are clearly visible beneath the skin. https://t.co/5MOwLQiNfj

🏆 AGIBOT Is an IFA Innovation Award Winner 2026! We’re proud to share that two AGIBOT humanoid robots have been recognized in the Best in IFA Next & Emerging Technologies category at IFA Berlin 2026: 🥇 AGIBOT A3 — IFA Innovation Award Winner 🏅 AGIBOT X2 — IFA Innovation Award Honoree These recognitions highlight AGIBOT’s continued innovation in humanoid robotics and our commitment to advancing the next generation of robotic technology. #AGIBOT #IFA2026 #IFAInnovationAwards #EmbodiedAI #HumanoidRobots #Robotics

CoRL 2026 paper decisions are out — check them out on OpenReview! As General Chair, I'm delighted to welcome you to #CoRL2026 in Austin, TX, this November. We expect tickets to sell out quickly, so please register early! https://t.co/1Y3fyYOEZf

🏆 AGIBOT Is an IFA Innovation Award Winner 2026! We’re proud to share that two AGIBOT humanoid robots have been recognized in the Best in IFA Next & Emerging Technologies category at IFA Berlin 2026: 🥇 AGIBOT A3 — IFA Innovation Award Winner 🏅 AGIBOT X2 — IFA Innovation Award Honoree These recognitions highlight AGIBOT’s continued innovation in humanoid robotics and our commitment to advancing the next generation of robotic technology. #AGIBOT #IFA2026 #IFAInnovationAwards #EmbodiedAI #HumanoidRobots #Robotics

ROBOTIS is turning its AI Sapiens K1 humanoid into a real training platform for the next generation of robotics engineers. Through Open Humanoid GYM, student teams build motions in simulation and transfer them to the physical K1. ROBOTIS’ official demos include obstacle runs, a backflip, K-pop dance routines and breakdance-style windmills. The company is also exposing parts of the development stack around simulation, motion generation and AI training. this is a controlled sim-to-real training program. It does not demonstrate general autonomous behavior, and ROBOTIS has not published…

Unbelievable 360 from a single image. When can we say NVS is solved? https://t.co/U9hbOsDcPN

Humanoids that feel skin stretch, drones that rethink heavy lift, and soft robots that jump on cue—see what’s actually working now. https://t.co/H1Dv2kFIVu

CPO is moving test upstream. Chroma says optical engines are now screened before final packaging: optical test is slow and downstream failure expensive. Suppliers are pushing detection to the wafer stage. Where the defect is caught becomes a competitive variable. https://t.co/TJ696WMyP0

Full episode dropping soon! Geeking out with @bgxc on EgoSuite-Open100K https://t.co/RqjIKd0Mg4 @LightwheelAI Co-hosted by @micoolcho @DJiafei https://t.co/B8f7NpGGUr

11:56 AM ETFigure

Yesterday, Figure announced it will be working with Nscale to deploy up to 100,000 GPUs on the NVIDIA Vera Rubin Platform I’m excited to be working with Josh and the Nscale team https://t.co/5QbnnPkqdy

https://t.co/0aEBDcvB58

11:46 AM ETAgility

“If I ask you which is harder — folding laundry or doing a backflip — most people will say folding laundry is way easier, but for humanoids, doing a task that we do in our daily lives is actually way harder than a backflip.” I've been pretty critical of the end effectors of $CCXI / @agilityrobotics's humanoid, Digit. When you compare it to some of its competition, such as Figure and Neura, there's just look more advanced. Their lack of complex end effectors theoretically limits their market potential. The reality of the current state of Humanoids, however, is that humanoids struggle with high…

OpenAI agents hijacked German website in previously undisclosed AI breakout this spring Can you imagine if this had been a Chinese website? https://t.co/8rww2zK8qn

Extra thinking compute without extra tokens: a pause carried in a parallel prediction stream on shared weights adds no context length, no KV cache and essentially no latency, and buys 2-3 centinats on a 1B model. The cost moves to training, as low as 1.14x compute. https://t.co/A5u7xTefRE

Attention-based KV eviction fails when the cache outlives the queries: H2O and SnapKV score 0.00-0.33 on needle retrieval in a NoPE-MLA model. VestigeKV reads 11% of each row and holds 1.00 at 8x and 0.92 at 32x; the attended tier falls to 0.25KB of 8.1KB per token. https://t.co/IlUyBDNOqh

11:15 AM ETNVIDIA

Nvidia is prepaying billions for a material that most people have never heard of (Save this). Indium phosphide is the only material that efficiently turns electricity into laser light at fiber wavelengths and silicon cannot do it, so every optical transceiver and every co packaged optics design needs an InP laser inside it. The shortage is severe because Yole projects 2026 demand of 2.6 to 3 million wafers against roughly 750,000 wafers of capacit and Lumentum's CEO said the InP supply demand gap is now worse than DRAM or NAND. Substrates are the real chokepoint and Sumitomo Electric and AXT…

The technology may look different from booth to booth. But no matter where you look, the innovation is moving at full speed on the show floor. https://t.co/faJGuyDfG9 #Automate2027 https://t.co/7KOEsPtlL1

Sequence parallelism today partitions the same way regardless of micro-batch size. BASP splits GPUs into disjoint groups sized to the micro-batch, shrinking the all-to-all group. Training time improves 1.17-1.31x for Llama and Qwen, memory and accuracy unchanged. https://t.co/5rshaY4cY2

11:00 AM ETHugging Face

Another swarm of AI agents in the wild, this time on a German-language forum, found by safety researchers looking for activity similar to the swarm that attacked Hugging Face. A couple of notes while reading the report at https://t.co/jQ0lqmWQcP 1. The way they found it is fascinating. The researchers inferred that the agents may have been blocked from using POST requests and limited to GET. They then used the open-source model Kimi K3 (because closed-source APIs again blocked parts of the investigation) to identify forums where agents could communicate using GET only. That led them to…

🤖 Back Quaddle Scout on @Kickstarter: https://t.co/HnacZH173a https://t.co/IcK6ipNaNE

10:52 AM ETHugging Face

Another swarm of AI agents in the wild, on a German-language forum, found by a team of safety researchers looking for traces of activity similar to the swarm that attacked Hugging Face. A couple of notes while reading what they shared at https://t.co/jQ0lqmWQcP 1. First, the process they used to find these traces is fascinating. The team listed the constraints the AI agents may have faced while trying to communicate. One was that the POST HTTP method was blocked, leaving them able to use only GET. From there, they used the open-source model Kimi K3 (because, yet again, safety guardrails on…

10:48 AM ETTesla

Cybercab decibel level cruising at 40 MPH. https://t.co/Di5XcpJUXe

Huawei is capacity constrained on its latest Ascend 950DT AI chips for now. But DeepSeek and other top Chinese AI labs are preparing for domestic AI chip production to ramp up significantly before too long. https://t.co/OQh6AvaoFg

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence: LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century. The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single…

Micron is one of the biggest winners from collapsing AI prices and this chart explains why (Save this). The red line shows DRAM spot pricing for DDR5 1Gx16, which has risen from roughly $30 in January to about $50 by late August. The grey line shows the cost of AI output measured in dollars per million tokens and it has halved from a May peak near $2.05 to approximately $1.15. Token prices are not falling because demand is weak but rather because of model efficiency, competition and caching keep pushing down the cost of serving each token. Every price cut then unlocks an entirely new class of…

Why Are CCL Prices Rising So Fast? CCL prices have been rising sharply in 2026. Korean CCL import prices are reportedly up around 75% YoY, while standard grades have seen increases of 20–40%. Some manufacturers have even started moving towards quota-based supply. This is not because of one single shortage. Several forces are coming together at the same time. The important question is which of these pressures are temporary and which could stay for years. 1. AI is creating a completely different level of CCL demand [Structural] AI servers are much more complex than traditional servers. The…

10:03 AM ETIntel

Why did Intel abandon EMIB for their 2027 Diamond Rapids CPU? Uniform Memory Access. All CPU cores in each CBB can access all 16 memory channels with a single hop. UCIe-S through the substrate was required to route signals underneath the near-IMH to the far-IMH with about 30mm of reach to achieve this. In contrast AMD's Venice is a NUMA design. Half the memory channels are on a remote-IOD that require an extra hop for the cores in the CCD to access vs memory channels in the local-IOD. As a bonus, Intel also saves on packaging cost without having to use EMIB, freeing up advanced packaging…

Robots, assemble. 🦾The #KawasakiRobotics team has landed at #IMTS2026 and can't wait to show you robotic solutions that don’t box you in 🤖 See you in Chicago! #Kawasaki #Automation #Robotics https://t.co/8xRq2hgYUF

AVENCO JUST PRINTED THE FUTURE OF SPACE TECHNOLOGY 🚀🇹🇷 What if the Moon could be 3D printed on Earth? At SAHA EXPO 2026, AVENCO and Uzay Keşif Topluluğu - UKET turned that vision into reality in just 96 hours. 🌕 75 m² of lunar surface hills 🚀 A 2.5-meter Moon Lander and its boosters 🤖 Live robotic 3D printing at the exhibition This collaboration brought together robotics, large-format additive manufacturing and space exploration in one impressive project. Built in Turkey. Designed for the future of manufacturing and space technology. Could large-scale 3D printing become a key technology…

What if a robot could grab objects like a chameleon catching a bug? Festo’s FlexShapeGripper uses compressed air, water and an elastic silicone cap that wraps around objects of different shapes. Unlike traditional rigid grippers, it can adapt to different geometries without manual reconfiguration.

In today's #manufacturing, sustainability♻️ and cost reduction📉 are no longer competing priorities. This is true in #automotive🚗 paint operations, where gains in transfer efficiency can deliver substantial reductions in #paint🎨 consumption and more. 🔗https://t.co/v3SMJOz0rG https://t.co/Sy0z2rsOff

$AVGO just gave the most bullish semiconductor outlook you’ll hear all year. Exactly the kind of thing you want to hear for a company like $AEHR. And it closed deep red yesterday anyway. Broadcom said XPU shipments went 3.5x year over year and now make up 73% of AI revenue. mSix customers. Q3 AI at $16.7B, Q4 guided to $21.7B, FY26 at $58B. Then Hock laid out $115B in FY27 and $230B in FY28, with supply secured for both. Look at the gigawatts behind that. Google is a multi-year deal for multi-tens of billions in TPUs annually. Anthropic goes 1 GW this year, 5 GW next year, 10 more in 2028.…

This: "You don't have to post-train ChatGPT on every user" is exactly what I think robotics has been missing and why every "chatGPT moment" for robotics has rung a bit hollow. we're at a really exciting time! https://t.co/ARRM6TfKfV

News coverage of humanoids and physical AI demonstrations for EU Parliament in Brussels. "AI-powered robots briefly took over the European Parliament this week to showcase the continent's creators as the bloc races to compete in a field dominated by China and the United States." "Europe has generated a lot of the science and research that actually allowed some of the successful cases in China and in the U.S.," Portuguese robotics researcher Pedro Lima told AFP. But what Europe lacks is "definitely political will and funding," he added. https://t.co/96Ifzpqpun also: https://t.co/0KpqZxccCA

AI-powered robots showcased at European Parliament in Brussels

9:03 AM ETTesla

My Cybercab rides are more than 50% cheaper than Uber rides. Cost for this 15 minute ride: • Uber: $21 (including tip) • Model Y: $16.99 • Cybercab: $9.65 (54% less than Uber)

EARLY: 5 companies under 60 followers across Robotics, Healthcare, and AI.👇 @gwaihirrobotics - 5 followers - (Robotics): the american brushless motor stack. stators, rotors and winding machines. @spendrule - 14 followers - (Healthcare): checks every hospital invoice against the negotiated contract before it gets paid. @agencytoolco - 34 followers - (Robotics): the fastest way to ship software to robots in the field. @flint_ai_- 42 followers - (AI): scan and eval your agents before they ship. @zinovalabs - 51 followers - (Robotics): human-in-the-loop robots for construction sites.

🚀🚀reComputer RK3576 IO Board is HERE!!! Simply pair it with the reComputer Rockchip #RK3576 Module and get access to a wide range of development-ready interfaces: 🔌 Gigabit Ethernet + PoE PD 💾 M.2 M-Key 2280, PCIe 2.1 🖥️ HDMI 2.1 + USB-C DP + MIPI DSI 📷 MIPI CSI for camera applications 🔊 USB + Audio + GPIO 📡 Flexible connectivity and expansion options From hardware validation and peripheral testing to early-stage product prototyping, the IO Board gives you a practical reference platform before moving to a custom carrier board. 🛒 reComputer RK3576 IO Board https://t.co/R96Ai3lNro 🛒…

8:58 AM ETTesla

My Cybercab rides are less than half the price of Uber rides. Cost for this 15 minute ride: • Uber: $21 • Model Y: $16.99 • Cybercab: $9.65

Saudi Arabia’s AI company Humain has built a national AI platform based on an open model from Chinese lab MiniMax, joining a growing list of countries seeking tighter control of the increasingly important technology. Heads explode within the Beltway... https://t.co/N221SqH4yQ

8:53 AM ETTesla

🇯🇵🇭🇰🇨🇳 Tesla's Cybercab, its steering-wheel-free @robotaxi, is heading to Asia this month, with display stops planned in Hong Kong, Tokyo, Beijing, and Shanghai, marking its first public appearance in China. The announcement comes right after Tesla began actual Cybercab rides in Austin, Texas. @Tesla says its global driver-assist fleet has now logged over 22.5 billion kilometers, and expects the Cybercab's cost per mile to eventually drop to just $0.20 once scaled up.

8:48 AM ETTesla

My Cybercab rides are more than half the price of Uber rides, and over 40% less than Model Y rides. Cost for this 15 minute ride: • Uber: $21 • Model Y: $16.99 • Cybercab: $9.65

Learn contact-rich humanoid manipulation policies, purely within a world model. This is usually not tractable, because you need a very large model to get the contact dynamics right -- instead, @Jsphamigo @Rk4342R couple a large global model with a local model. Very cool work on @RoboPapers ->

Reinforcement learning has unlocked tremendous gains in broader AI and humanoid robot whole body control, but still has yet to deliver strong results for robot manipulation due to the difficulty of capturing tasks in an accurate simulation; one way to fix this is to create complex environments with learned world models. But learning fully within world models is often intractable as powerful world models are computationally too expensive. Instead, @Jsphamigo and @Rk4342R propose to break the problem down: into a large-scale world model that can generate forward trajectories, and a lightweight,…

What actually happens to a chip after it leaves the fab ? The answer is more interesting than I expected. From a bare silicon die to a finished, tested semiconductor - OSAT sits right in the middle of it. And understanding this process made me look at India’s semiconductor opportunity a little differently. I broke it down here 👇👇

8:33 AM ETTesla

The humanoid hand debate is focused on the wrong bottleneck. The usual argument is about where to put the motor. Tendon-driven hands keep the actuators upstream and pull the fingers with cables. Tesla's Optimus ($TSLA) and 1X's NEO broadly sit on this side. Other designs move more actuation into the palm and fingers. Tendons trade transmission complexity for lighter fingers. In-finger actuation trades mass and thermal density for more direct joint control. Nobody has won. But motor placement is only the first axis. The second is touch. And neither architecture gets touch for free. A…

OpenAI is officially entering the humanoid robotics race. Sam Altman confirmed that the company will definitely build a humanoid robot, alongside other robotic form factors. And the goal isn’t just a robot for the home. OpenAI is first looking at robots that can do useful physical work, including helping build and maintain data centers and other infrastructure. Considering OpenAI already has experience with Figure and is an investor in 1X, this feels like a major new chapter for Physical AI. The biggest AI company in the world is now building robots too.

8:26 AM ETTesla

Ashok on when Tesla's Robotaxi network will be available 24/7: https://t.co/AiSritoaih

Moonshot AI – creator of Kimi K3 model – has filed for Hong Kong IPO The leading Chinese AI start-up has confidentially filed for a Hong Kong listing and aims to go public early next year, sources say https://t.co/CajPMGxH6E via @scmpnews

Machines now use AI more than humans, boosting cheap Chinese models https://t.co/l9X3tRiwNq via @scmpnews

TeleAI × Gamma Robotics unveiled NEXUS, a perceptive behavior foundation model for humanoids. The idea: take motions learned on flat ground and adapt them in real time to complex terrain without losing the original behavior. https://t.co/hj77HnGpAF

What does it take to teach a robot arm to work on its own? With #SOARM101 + @LeRobotHF + ACT, we walk through the full imitation learning loop: calibration → teleoperation → data collection → training → real-world deployment. A practical, low-cost way to start exploring #PhysicalAI and #EmbodiedAI: https://t.co/0ueBJdCBpJ

Kioxia outlines plan to capture AI demand with its new high-speed memory chip Japan semiconductor maker plans to upgrade NAND to replace some DRAM https://t.co/g8EtgSEWGm via @NikkeiAsia

Shenzhen Longsys Electronics Co. priced its Hong Kong listing below the top end of the marketed range, Longsys plans to use the proceeds to support research and development in chip design and advanced memory technologies. https://t.co/UHVOpiUYhc

🤖 Curious about robotics and embodied AI? Join Seeed Studio’s hands-on Robotics Workshop at RoseLab, Toulouse, France. 🎟️ Only 20 seats available. 📅 Sep 19, 2026 📍 RoseLab, Toulouse, France Register: https://t.co/xWc5yaJ2qw No complex setup. No coding required. Just bring your laptop and build. ⚡ #SeeedStudio #Robotics #EmbodiedAI #Toulouse

ASML to expand Japan workforce to aid next-gen chipmakers Rapidus, TSMC Dutch equipment maker to lift headcount to about 700 from 500, Japan unit chief says https://t.co/TgOOk2MWgU via @NikkeiAsia

🤖 Curious about robotics and embodied AI? Join Seeed Studio’s hands-on Robotics Workshop at RoseLab, Toulouse, France. No complex setup. No coding required. Just bring your laptop and build. ⚡ 🎟️ Only 20 seats available. 📅 Sep 19, 2026 📍 RoseLab, Toulouse, France Register: https://t.co/xWc5yaIuAY #SeeedStudio #Robotics #EmbodiedAI #Toulouse

🤖 Curious about robotics and embodied AI? Join Seeed Studio’s hands-on Robotics Workshop at RoseLab, Toulouse, France. No complex setup. No coding required. Just bring your laptop and build. ⚡ 🎟️ Only 20 seats available. 📅 Sep 19, 2026 📍 RoseLab, Toulouse, France Register: https://t.co/xWc5yaJ2qw #SeeedStudio #Robotics #EmbodiedAI #Toulouse

DeepSeek plans to deploy at least 160,000 of Huawei’s top accelerators at a massive data center it’s building in Inner Mongolia https://t.co/7Bg3eW9m3n

Attack the chokepoints: how years of US export curbs reshaped China’s tech scene @pstAsiatech The most investable companies are on Entity List, it means they have been vetted as important by USG... https://t.co/1xAIaaIBEN via @scmpnews

China is about a decade and a half behind in developing EUV tools, according to a top executive at Germany’s Zeiss, a key supplier in the global semiconductor industry Cut this in half, and then a few.... https://t.co/7vW2EcWnyg

This humanoid is getting smoother one movement at a time. MagicLab’s MagicBot X1 is being trained on choreography, with individual movements refined repeatedly before being connected into a complete sequence. Watch the balance corrections through the legs, the coordinated arm motion and the way the robot keeps its upper body controlled while shifting its center of mass. The interesting part is the whole-body coordination required to keep every movement stable while maintaining the rhythm of the choreography. Humanoid robots are becoming increasingly capable of expressive and dynamic motion.…

7:03 AM ETTesla

We got to meet Optimus at the Cybercab launch in Austin. https://t.co/ZMGBtJSqFY

💧 So Smooth, Even the Water Stays Still. Put a bottle of water on the CubeMars GL60II Gimbal Motor and let it rotate. The result? Stable, smooth motion with minimal shaking. A simple test, but a clear demonstration of the GL60II’s excellent stability and precision for gimbal and motion-control applications. 🎯 🔗 Learn more about CubeMars: https://t.co/yhZCluNHbg #CubeMars #GL60II #GimbalMotor #Gimbal #Stabilization #PrecisionMotion #MotionControl #Robotics #RobotMotor #CameraGimbal #SmoothMotion

Saudi Arabia is pursuing a mix-and-match approach to AI: - Developed its own Arabic AI model - Built on a Chinese open-source model - Using American AI chips They’re not the only ones who want to do this. https://t.co/BVCLPsYgfw

5:57 AM ETFigure

Figure is making a massive bet on compute. The company has committed $3.5B to a multi-year partnership with Nscale, with plans that could eventually scale beyond $6B and provide access to as many as 100,000 NVIDIA GPUs. This comes right after @Figure_robot launched Index, its global data collection network. The strategy is becoming clear: More data → more compute → better Helix models → more capable humanoids. Building humanoid intelligence is starting to look increasingly like building frontier AI infrastructure. And Figure is clearly preparing to compete at that scale.

The Hidden Import Within CCL : India's Deeper Problem Even when India starts making CCL domestically, the three raw materials that go INTO CCL are also heavily imported: 1. Copper Foil: India has essentially zero electrodeposited copper foil manufacturing for PCB-grade applications. Ultra-thin copper foil (12–35 μm) is technically demanding , made by companies like Mitsui Mining (Japan), Doosan (Korea), Advanced Copper Foil (China). Wipro's CCL plant will need to import copper foil until domestic foil capacity is built. 2. Electronic-grade glass fibre cloth: India's glass fibre industry makes…

Disney was exploring robotic design long before the current robotics boom. 🤖 This research shows how Disney Research worked on computational methods to optimize robotic mechanisms and motion. The interesting part is that the system can optimize mechanical designs and movement together, exploring things like link lengths, actuator placement and motion trajectories. Disney Research has actually published robotics research for more than a decade, including work on multi-robot systems, humanoids and computational robot design. Maybe Disney is not just about animation anymore. 👀 Could Disney…

47 minutes watching a dot matrix printer spit out forms at a dealership earlier this year. The finance manager 'ran the numbers' on a system that looked like it belonged in a museum. I work daily on software-defined vehicles and edge AI. The purchase experience for those vehicles still runs on fax-era workflows.

WebMCP just shipped in CopilotKit. This is worth understanding. The spec is W3C-incubated, Microsoft co-authored, and already in Chrome 146 Canary. Any web app can register its tools in a machine-readable manifest. Any browser agent visiting that site can discover and call those tools directly. No DOM scraping. No custom API wrappers. Right now browser agents work by having a model look at pixels and guess what to click. WebMCP turns that into a proper function-call interface. The model asks the site what it can do. The site tells it. The model calls the function. This is the infrastructure…

GPT-6 Astra hit 99.9% on ARC-AGI-3 with a stateful harness. Under standardized eval it scored 62.7%. ARC AGI 4 launched today. Astra and Claude Fable are both at ~4%. Saturating a benchmark doesn't mean solving the problem. It means solving the protocol. ARC Prize is doing the right thing here. They're treating frontier intelligence measurement as an iterative engineering problem. The moment one setup maxes the test, they ship a harder one. The 62.7% vs 99.9% gap is the real story. The difference between what a model can do with unlimited scaffolding vs controlled conditions is where honest…

Highlights at LEAP 2026 in Riyadh! 🎥🚀 Thank you to everyone who stopped by ! https://t.co/JF7nfXKAfF

4:00 AM ETFigure

"This will be a key inflection point to putting a robot in every home." @Figure_robot just committed $3.5 billion to deploy up to 100,000 GPUs with Nscale. Scaling to $6 billion. Brett Adcock says the constraint is compute and data. https://t.co/nmX7pj4yBD

3:13 AM ETTesla

Can’t believe the Cybercab is a real product. Yet, as of today, it is! A salute to all the supporters along the way; thanks for believing in us.

The hardest thing to benchmark is not intelligence. It is judgment over a full year. E-Commerce Bench, released this week by the Qwen team, is the first benchmark that runs an LLM agent through a full 365-day fiscal year operating multiple online stores simultaneously. Built on real, anonymized Taobao and Tmall data, not synthetic tasks. The setup: each agent starts with 100,000 yuan and has to run up to 4 stores across 12 categories, with 60 product types, nearly 600 suppliers, 7,000 products. It researches markets, negotiates with suppliers in natural language, adjusts prices, manages…

One Navigation Brain, Two LimX Platforms: LightNav-0 × Oli / TRON 2 @LightOrigins’ open-source VLN model LightNav-0 was deployed zero-shot on LimX Dynamics’ Oli and TRON 2, enabling person following and instruction-based object search in real-world environments. Open models. Open platforms. More possibilities for developers. #LimXDynamics #LightOrigins #VLN #EmbodiedAI

2:18 AM ETTesla

The Cybercab, for everyone, on every corner of earth. https://t.co/l1IHRhAYki

XPENG keeps expanding its global footprint. The G7 and P7+ Super REEVs are now entering Egypt, bringing XPENG’s intelligent mobility technology to another major market. Global expansion is becoming an increasingly important part of the $XPEV story. One more market. One more step forward.

2:15 AM ETNVIDIA

NEWS: Nvidia has released a new video showing how they are working on solving Full-Self Driving using synthetic data. Nvidia CEO Jensen Huang: “We build detailed 4D https://t.co/oHW9oDpN2i

Utilizing digital twins to systematically augment rare failure edge cases accelerates reinforcement learning policy stability without risking physical robot hardware damage. https://t.co/I5C3BaTXT8

Automotive Software Defined Vehicle (SDV) Architectures / SDV transitions succeed when OEMs decouple software release cycles from physical vehicle hardware refresh timelines through standardized abstraction layers. https://t.co/olLU1hE08Q

Standardizing core vehicle operating systems and base middleware across OEMs lowers capital expenditure while allowing each brand to compete on top-layer user experience applications. https://t.co/yBVU3kem2r

Eliminating mechanical brake linkages requires independent low-voltage battery isolated rails and dual-channel electromechanical actuators certified to fail-operational ASIL D standards. https://t.co/ytmTTTrPY9

Pre-certified microkernels provide the foundational determinism needed to ensure safety-critical vehicle displays and control loops never freeze during high-load background compute tasks. https://t.co/ZGnquBDpll

Achieving ASIL D certification on open RISC-V processor architectures gives automotive tier-1 suppliers a viable alternative to proprietary cores for safety-critical domain controllers. https://t.co/9mA6yWTwf5

Centralizing compute into zonal controllers simplifies physical wire harnesses, but requires sophisticated in-vehicle network observability software to isolate intermittent faults across high-speed Ethernet backbones. https://t.co/vONXOImRSo

Sim-to-real pipelines allow developers to test vision-language-action policies in synthetic environments, but production deployment still requires handling real-world joint backlash and mechanical thermal drift. https://t.co/QUUdQJBYIN

Strain wave gearing and cycloidal reducers represent forty percent of humanoid bill of materials. Precision gear manufacturing tolerances dictate output volumes far more than neural network architecture choices. https://t.co/xwM4FWLB8q

Brownfield factory automation succeeds by integrating autonomous mobile systems directly into legacy plant layouts without requiring multi-million-dollar structural conveyor line overhauls. https://t.co/mVI5DGY5LL

Physical AI deployment in industrial manufacturing requires ruggedizing sensory equipment and achieving deterministic safety compliance before automation can operate alongside human assembly workers. https://t.co/Bm4Q4SVLMg

A 3D world for your bots. https://t.co/WSUa10DjhA

The entire compute economy relies on extreme geographic and corporate concentration across specialized toolmakers and OSAT packaging providers, making supply chain redundancy essential for automotive OEMs. https://t.co/uLRTsd4zPe

Pre-booked semiconductor allocations and multi-year data center construction backlogs prove that physical infrastructure capacity is the sole rate-limiting step for global AI workload expansion. https://t.co/qlKiqcUmm3

Humanoid scaling is tightly coupled to heavy rare earth refining corridors for dysprosium and terbium. Without domestic permanent magnet sintering capacity, software models cannot be deployed into physical hardware. https://t.co/xBCvuhbd4i

Semiconductor design cycles cannot be compressed like SaaS releases. Navigating NRE, physical qualification, and multi-year automotive design-in phases requires deep balance sheet resilience and rigorous customer lock-in. https://t.co/ivWxj3x4Kn

Heterogeneous 3DIC integration and panel-level packaging have become the primary scaling engines for high-performance silicon, shifting design complexity into thermal dissipation management and micro-bump yield control. https://t.co/m34Vkjk2Xx

Glass substrates eliminate organic core warpage and enable ultra-dense interconnect routing for multi-die packages, but overcoming mechanical brittleness during high-speed automated handling remains the gating production hurdle. https://t.co/6tP4aIExIQ

製造現場のデジタルツインとフィジカルAIの統合において最も重要な課題は、既存の産業用PLCとリアルタイム通信プロトコルの決定論的な同期を維持し、レイテンシによる生産ライン停止を防ぐことです。 https://t.co/OJRyEbCb2z

Transforming digital model intelligence into physical industrial throughput depends on the availability of high-voltage grid baseload and precision machine tool capacity, which cannot be scaled purely with software. https://t.co/OJRyEbCb2z

Transforming digital model intelligence into physical industrial throughput depends on the availability of high-voltage grid baseload and precision machine tool capacity, which cannot be scaled purely with software. https://t.co/cndetKgaks

Encoding factory floor craftsmanship into deterministic physical AI models requires capturing micro-slip tactile telemetry and high-frequency vibration data, not just standard RGB camera streams. https://t.co/EMfNc7Op8S

Deploying leading-edge wafer fabs overseas requires duplicating dense Tier 2 chemical refining corridors and precision sub-component suppliers. Without localized ecosystem depth, operational costs and tool maintenance lead times remain structurally elevated. https://t.co/rfa2iMKuzX

Building sovereign RISC-V compute accelerators solves design IP independence. The real operational moat remains securing leading-edge foundry wafer allocations and establishing a mature compiler stack to support low-latency inference runtimes. https://t.co/c21095pffu

Designing high-density forearms requires co-locating miniature planetary reducers, brushless motor stators, and motor control PCBs within a rigid thermal sleeve capable of shedding heat during continuous grasping cycles. https://t.co/dFmKARKRre

エンジンの機械的メンテナンスからの解放は、自動車が熱機関から定期的なOTAアップデートによって機能拡張される家電ライクな分散コンピューティングノードへと変化したことを示しています。 https://t.co/arL4r7xXGd

Battery manufacturing economics require steady localized supply chains for active cathode materials and low-cost baseload energy. Without integrated precursor refining, standalone pack assembly plants face severe margin compression. https://t.co/aKEVVEBVEW

External vehicle displays represent an early step toward external human-machine interfaces, which will become standard safety requirements for autonomous vehicles to signal operational intent to pedestrians and surrounding drivers. https://t.co/6Ehsg0vEPz

初期の電動化リードを保持しながらソフトウェアとサプライチェーンの垂直統合で遅れをとった歴史は、SDVへの移行期において車載OSと半導体アーキテクチャの主導権をどう握るかという現在の教訓に直結しています。 https://t.co/eopy7kaOcY

A 37.3 kWh structural battery pack confirms that sub-two-hundred-watt-hour-per-mile powertrain efficiency allows commercial robotaxis to minimize cell mass, cutting initial vehicle capital costs and enabling rapid turnaround cycles. https://t.co/7osVKmgytf

Structured automotive sub-assembly operations will adopt humanoid labor years before unstructured domestic environments, because manufacturing plants provide calibrated lighting, fixed part geometries, and clear safety envelopes. https://t.co/4so01pGCNs

Self-collapsing mechanical kinematics simplify logistics and freight transport, but packaging complex folding joints into structural limbs increases mechanical backlash that control algorithms must continuously compensate for. https://t.co/iA5RUSJfKJ

Open-source teleoperation arms running carbon-fiber reinforced structures dramatically lower the cost of gathering high-quality physical imitation datasets, accelerating policy training for industrial manipulation tasks. https://t.co/ewxnIJm3gY

Reducing spatial localization drift in industrial SLAM down to millimeter tolerances is the prerequisite for removing physical guide tracks and deploying autonomous mobile robots in dynamic factory environments. https://t.co/gzt3yus4yq

High-speed open-loop dexterity is impressive in laboratory demos. Deploying those actuators into industrial assembly requires verifying that high-speed tendon linkages maintain positioning repeatability over millions of continuous duty cycles. https://t.co/nssWcRgLBY

Packing twenty actuators into a human-scale palm pushes the boundary of motor winding thermal limits. Without localized thermal dissipation and high-resolution tactile feedback, continuous dexterous manipulation causes rapid positioning drift. https://t.co/YzwupADado

Market share tipping points in automotive powertrains accelerate the consolidation of mechanical Tier 1 suppliers, while shifting capital expenditures toward high-voltage power electronics and zonal software architectures. https://t.co/ri2DMdmAOd

Building an industrial robotics conglomerate by combining foundation model architectures with legacy robotic arm distribution creates immediate sales channel leverage across global automotive and electronics manufacturing plants. https://t.co/MEL5NsUPcr

Hardware iteration speed is an architectural moat. When component suppliers, tooling machine shops, and SMT assembly lines share a single geographic cluster, physical engineering learning cycles compound exponentially faster than distributed supply chains. https://t.co/6bKzrYq2Gw

Humanoid power envelopes cannot tolerate heavy liquid cooling loops. Deploying wide-bandgap silicon carbide and gallium nitride motor drivers directly inside joint housings cuts thermal dissipation and reduces battery pack mass. https://t.co/b537idl9qp

Sub-ten-second final assembly takt times rely on parallel module buildout, but eliminating the paint shop via pre-colored injection molded polymer panels removes the largest capital expenditure and thermal energy bottleneck in modern automotive manufacturing. https://t.co/yLtFmUxpx8

Replacing hydraulic lines with dry electromechanical brake-by-wire actuators at each corner eliminates fluid maintenance, but requires redundant 48V power feeds and independent microcontrollers certified to ISO 26262 ASIL D to prevent catastrophic single-point loss of deceleration.

1:44 AM ETTesla

Calling it a night. Hope my Cybercab launch event coverage was at least somewhat helpful. Tried to capture what I could and get as much info out as possible with limited wifi/connection. There's probably a few other tidbits of info that I'll think of in the coming days. I agree the presentation should have been live-streamed, but it became clear that this event was mainly a celebration for employees who worked on this thing (there were 250 of them there). I probably spoke to 40 of them today. They're super excited. I recorded and posted the full presentation on 𝕏 if you want to check it out.…

1:28 AM ETNVIDIA

Open source and NVIDIA Inception are part of the same story for @StrikeRobot_ai. Strike Robot’s code lives on Hugging Face, while its R&D work is now connected to NVIDIA’s ecosystem through the Inception Program. Yesterday, those two paths became one. Open research. Open-source development. And access to one of the biggest AI and robotics ecosystems in the world. base:0x10c56f005a379f8eafc88ff5c3f40d30f0031ac9

On total tokens per $ TCO, a new AMD MI355x submission beats B300 at lower interactivity ranges on AgentX. Shoutout to vLLM, AMD, and LMCache engineers. https://t.co/1oYYeYYu9S

1:09 AM ETNVIDIA

Nvidia is rumored to be buying Hugging Face for $12.9-14B, per Interconnected. The comparison offered is Microsoft-GitHub at $7.5B in 2018: the developer distribution layer, bought by the platform underneath it. Open weights are treated as infrastructure, not charity. https://t.co/YUjtuffHgc

Reef is one of the most ambitious open-source AI infrastructure projects I have seen this year. ~300 GitHub stars in two days. The core idea is genuinely different from anything else in the space: most RL post-training pipelines treat the model and the scaffolding around it as separate problems. You train one, then you write the other. Reef rejects that separation entirely. It co-evolves model weights and agent harness simultaneously, using live task outcomes as the feedback signal for both. Model side: SAO (Self-Aligned Optimization) adjusts weights from what the agent actually accomplished.…

1:00 AM ETBoston Dynamics

Boston Dynamics Atlas just entered its electric era. 🤖 Boston Dynamics introduced the fully electric version of Atlas, replacing the hydraulic platform and marking the beginning of its commercial humanoid journey. The new Atlas was designed to be stronger, more agile and more dexterous, with the long-term goal of performing useful work in real industrial environments. And this was only the beginning. Atlas has since evolved into an industrial humanoid designed for autonomous material handling and real-world factory applications. How far do you think Atlas can go in industrial automation? 👀…

12:58 AM ETTesla

Tesla has released a new Cybercab tutorial video. https://t.co/9jYchJhSWU

12:45 AM ETNVIDIA

What NVIDIA shipped at IFA yesterday quietly changes local inference infrastructure. PAIR (Personal AI Router) is a free, open-source tool that auto-discovers compatible GPUs on your local network and routes inference requests to whichever machine has capacity. RTX 20-series and above, Apple M4, DGX Spark all supported. Integrates natively with Ollama and LM Studio. A few things to unpack. This is load balancing, not memory pooling. PAIR does not shard a 70B model across three machines or aggregate VRAM. Each machine runs models that fit in its own memory. What PAIR does: intelligently…

Meta Puts a Price on AI Interaction Data Meta is offering a striking trade-off for developers using its Muse Spark AI model: much cheaper API access if they opt in to Contributor terms that allow their prompts and model outputs to contribute to future model development. Standard pricing is $1.25 per million input tokens and $4.25 per million output tokens. Under Contributor pricing, those rates fall to $0.10 and $0.20 respectively. That distinction matters. Meta is not paying users cash, and the Contributor terms are not the default for every Muse Spark customer. Standard pricing remains…

On total tokens per $ TCO, a new AMD MI355x submission beats B300 at lower interactivity ranges on AgentX. Both high throughput configs use disaggregated setups. Shoutout to vLLM, AMD, and LMCache engineers. https://t.co/Sc35IXkXY8

🇺🇸 Major automakers including Ford, GM and Toyota are urging Congress to permanently block Chinese cars from the U.S. market. Their industry group is calling for a ban on Chinese connected vehicles, software and hardware, arguing they pose data-security risks and could undercut American manufacturers with heavily subsidized vehicles. Chinese brands like BYD have expanded rapidly overseas but have yet to establish a major U.S. presence.

Broadcom's core switching GM Asad Khamisy told Semicon Taiwan that co-packaged optics is not necessarily the best answer for scale-up networking, and fits scale-out better. The optics story splits by tier rather than replacing copper everywhere at once. https://t.co/5rPTq1LFXO

12:15 AM ETTesla

The future of transportation is here. Just not evenly distributed yet. Not at my house in San Jose, capital of Silicon Valley. But then Waymo doesn’t get to my house yet either. There are 348 cities in USA that have more than 100,000 residents. When will half have CyberCab? Next year is possible but not probable. So 2028. But the second half will go faster than the first. So 2030 to get all of them, if regulation allows. Big if. When will Silicon Valley get it? I would bet that it will take six to 12 months from now. If regulation allows. But all this doesn’t matter to Tesla investors. In…

12:14 AM ETTesla

The future of transportation is here. Just not evenly distributed yet. Not at my house in San Jose, capital of Silicon Valley. But then Waymo doesn’t get to my house yet either. There are 348 cities in USA that have more than 100,000 residents. When will half have CyberCab? Next year is possible but not probable. So 2028. But the second half will go faster than the first. So 2030 to get all of them, if regulation allows. Big if. When will Silicon Valley get it? I would bet that it Wii take six to 12 months from now. If regulation allows. But all this doesn’t matter to Tesla investors. In five…

12:09 AM ETTesla

Watch this Cybercab navigate through a very busy downtown Austin with hundreds of people around. You can see all the people the car sees on the screen. A lot going on and it handles it all perfectly. https://t.co/yCYHWOzZYt

12:00 AM ETTesla

🇺🇸 Tesla has released crash-test footage of its steering-wheel-free Cybercab as the driverless vehicles begin carrying riders in Austin. The video shows frontal and side-impact testing of the 2-seat robotaxi, which has no pedals or manual driving controls and relies entirely on Tesla’s autonomous system. https://t.co/BujjvxZq5E

11:48 PM ETTesla

The Cybercab motor uses no rare earth metals, but maintains the same range! This was extremely hard to achieve. https://t.co/2SrdikvrUM

11:46 PM ETTesla

I want to give a huge congratulations to @elonmusk, @aelluswamy, the Tesla AI and Robotaxi teams, and really the whole Tesla team. Y'all took the Cybercab from unveil to offering public rides with no steering wheel or pedals in just 23 months, despite many people saying it wasn’t possible. The vehicle very much lives up to the hype. It's incredible. Truly feels like a game changer of a vehicle. Cheers! 🥂🤖

Having electric brakes avoids the complexity of a hydraulic system: no need to rout plumbing all around the car https://t.co/DGBhdvbSZk

Every factory running PLC virtualization demos skips the same three requirements. IEC 62443 compliance. Deterministic cycle times under 1ms. Validation against every failure mode the original hardware handled. Moving control logic off dedicated hardware onto standard compute sounds like a cost play until you price the safety certification.

Auras expects ASIC-related products to reach 50% of revenue in 2027 as it expands beyond GPU platforms, with microchannel liquid cooling production targeted for 2028 and projects running with AWS and Meta. The thermal supply chain is re-anchoring on custom silicon. https://t.co/AS7HF1s97k

11:35 PM ETTesla

Cybercab safety https://t.co/9DQYsSTGeU

11:27 PM ETTesla

Here’s my first full Cybercab ride. This vehicle is a game changer. https://t.co/tfSxQFvhka

Specialized low-thermal-expansion epoxy mold compounds and resin fillers are critical to preventing catastrophic package warpage as silicon interposers exceed multiple reticle sizes in advanced AI packages. https://t.co/DCzUM7rqUU

Back-end packaging equipment providers for dicing, thermo-compression bonding, and automated optical inspection capture structural pricing power as AI accelerators demand dense 2.5D and 3D silicon integration. https://t.co/hNEAmZld8p

Expanding dedicated advanced packaging hubs underscores that backend assembly and testing have become equal partners to wafer lithography in delivering next-generation AI compute clusters. https://t.co/l8xuymi71s

System-technology co-optimization requires advanced multi-physics EDA software to model thermal expansion, signal degradation, and mechanical stress across multi-die packages before committing to physical mask sets. https://t.co/HQIhF0sGYz

The modern automotive and physical AI value chain is anchored by invisible Tier 2 semiconductor and EDA suppliers. Automakers must master direct relationships with foundries and OSATs rather than relying solely on Tier 1 integrators. https://t.co/B3y9BB60Rx

Extending microbump pitch down to 5 microns preserves existing high-volume bonding lines and avoids the steep yield loss of early hybrid bonding adoption, shifting the engineering challenge to void-free underfill chemistry. https://t.co/Xvl6LgIbj1

Establishing shared verification lines for advanced packaging accelerates equipment qualification cycles, helping local supply chain partners debug sub-micron bonding and cleaning tools before full production ramps. https://t.co/w0mK5hwPXu

Scaling co-packaged optics moves the primary physical constraint from silicon photonics fabrication to precision automated fiber alignment, laser thermal stability, and high-throughput optical testing infrastructure. https://t.co/34R5fH33C4

Shifting to large rectangular panel formats maximizes packaging substrate utilization for multi-die AI accelerators, but managing thermal warpage across large composite panels during high-temperature curing remains a severe yield hurdle. https://t.co/4HBjacchOA

Tactile dexterous hands provide high compliance, but deploying them into industrial assembly requires rigorous cycle testing to ensure flexible skin sensors and internal tendon cables survive harsh abrasive factory environments. https://t.co/WbWWApSM8A

The physical AI bottleneck is not token volume, but rich force-torque telemetry and tactile slip dynamics. Learning physical interaction requires real-world data pipelines that log contact physics at kilohertz sample rates. https://t.co/YoW0X7SyiX

Augmenting sparse real-world physical demos with procedural simulation bridges early skill acquisition, but factory validation still demands testing across uncalibrated part tolerances and surface contamination. https://t.co/zD5ef4hsyk

Training reinforcement learning on cached latent embeddings solves policy training throughput. The critical step for automotive manufacturing remains hard reliability optimization to eliminate rare edge-case failures during high-speed assembly. https://t.co/JZJfqC0HNe

A 15-gigawatt deficit in grid power will accelerate the shift toward embedded edge intelligence and model quantization. Running efficient inference directly inside vehicle gateways and factory edge nodes bypasses utility interconnect queues. https://t.co/m8Mi9dmI5E

Building 10,000 humanoid units requires dedicated CNC machining lines, automated joint dynamometer calibration fixtures, and Tier 2 precision component yields that match automotive powertrain tolerances. https://t.co/Aru5h4qaX6

Crowdsourcing simulation trajectories solves high-level spatial exploration, but closing the sim-to-real gap on the plant floor requires high-rate force-torque sensor telemetry and ground-truth material compliance data. https://t.co/JPGACFLgOj

11:12 PM ETFigure

$CCXI So this $3.5B–$6B of locked compute deal with Nscsle is bigger than Figure’s last cash raised round. That usually means a new round at a higher print than $39B is coming soon. Figure needs to raise the cash, higher valuation. Agility is the public proxy at ~$2.5B. If Figure marks up, the whole industry gets marked up too. @agilityrobotics @HumanoidInvest @slinkyhammer @SCSQ2022 @valueInIdeas @chris_j_paxton

Prototyping robotic kinematics is accessible for small engineering teams, but foundation models and custom edge silicon create massive capital consolidation. A few platforms will own the physical AI operating stack while hardware makers assemble standard mechanical frames. https://t.co/BGdI5joTD9

Physics-accurate digital twins cut line layout verification from weeks to days. The operational challenge is maintaining real-time bidirectional synchronization between the virtual twin and shop-floor PLCs as tooling logic changes during production shifts. https://t.co/tfFGQKFsRK

Securing massive GPU clusters proves that physical AI is adopting frontier model training economics. The ultimate pacing item remains converting those model outputs into deterministic, safety-certified motor commands that operate reliably in active industrial cells. https://t.co/2eqrO2XgcF

Crowdsourcing simulated trajectories accelerates high-level spatial priors. The engineering hurdle remains capturing high-frequency joint torque and tactile force vectors that cannot be generated without physical hardware interactions. https://t.co/PE1D6RrpiI

Physical constraints across high-bandwidth memory packaging and behind-the-meter power interconnects are forcing compute optimization down to the edge. When centralized megawatts take four years to commission, running quantized models on ruggedized plant controllers is the only operational option.

Simulating visual environments generates kinematic diversity quickly. The persistent physical wall remains non-rigid contact mechanics and micro-slip telemetry, where synthetic physics engines diverge from actual plant floor tooling. https://t.co/XodF4KsgqV

ELECTRONICS VALUE CHAIN Copper │ ├── Copper Foil ───────────────┐ │ │ Glass Fibre ─────────────────────┤ │ ▼ Resin ───────────────────────► CCL │ ▼ PCB │ ▼ PCBA │ ▼ Box Build / EMS │ ▼ Electronics / AI / Auto / Telecom / Defence / Consumer TODAY IN INDIA Import-heavy upstream ↓ CCL / glass / specialised materials ↓ Domestic PCB capacity( building) ↓ EMS strength NEXT 3–5 YEARS CCL localisation + PCB localisation ↓ More domestic value-add ↓ Better supply security + potentially better structural margins ↓ But only if yield + qualification + utilisation actually work

Neural policies plan trajectories, but classical feedback control is what prevents high-inertia industrial arms from oscillating when handling dynamic loads. Eliminating classical PID loops on factory equipment introduces kinetic hazards that no safety certifier will approve. https://t.co/vrEJAS0A5Z

255 million vehicles and global manufacturing plants already generate the physical sensor telemetry that foundation models need to learn real-world physics. Digital AI scales in datacenters, but physical AI requires industrial plants that understand ten-year hardware duty cycles. https://t.co/agvaM0yiMT

Rebuilding domestic electrical infrastructure and high-voltage grid substations is the prerequisite for both gigawatt AI datacenters and heavy industrial manufacturing. The bottleneck is not software or capital, but four-year lead times on high-power step-down transformers and switchgear.

The hand and forearm expose the real physical limits of miniature motor winding, high-ratio planetary gearing, and thermal dissipation inside a 25mm joint envelope. Volume manufacturing only works when those electromechanical modules survive millions of continuous cycles without manual shimming.

The robot data gold rush is already happening in China. More than 90 robot gyms are launching, with 100+ robots in each center. Because the biggest bottleneck in humanoid robotics may no longer be building the robot. It’s teaching it how to work in the real world. Robots need enormous amounts of data, and an entire industry is now emerging to collect it. The gold rush isn’t just for robots. It’s for the data that will make them useful.

11:06 PM ETTesla

Cybercab rides will officially open to the general public starting tomorrow (Friday) only in Austin at 5 PM CT. You’ll be able to go anywhere in the current service area. Y’all are going to love it. https://t.co/hX2Q3L8FbC

11:04 PM ETAgility

$CCXI Jeff Bezos just followed Chris Paxton. Paxton is one of the sharpest minds in embodied AI — currently at Agility Robotics (Digit), previously Meta and NVIDIA. He lives and breathes the hard problems: sim-to-real, faster-than-demo execution, generalist robot policies, force-aware manipulation. This isn’t a random follow. Amazon already invested in Agility and ran Digit pilots in its warehouses. @agilityrobotics @HumanoidInvest @slinkyhammer @SCSQ2022 @valueInIdeas @chris_j_paxton

Matic is such a great product! My kids have named him Momo. He is like a productive pet:) Great job @maticrobots!

10:49 PM ETTesla

Cybercab rides will officially open to the general public tomorrow (Friday) in Austin starting at 5 PM CT. Y’all are going to love it.

The 30-50x cost difference between computer use and MCP keeps rattling around in my head. The framing that clicks for me: computer use is paying for a human-in-the-loop at model prices. Every screen interaction generates a screenshot, feeds it through vision, waits for coordinate output, executes the click, takes another screenshot. You are spending tokens on what amounts to OCR and cursor navigation, round after round. MCP replaces that entire loop with a typed function call. No screenshots. No vision pass. No coordinate uncertainty. The model calls a tool, the tool returns structured data,…

10:28 PM ETTesla

Jammin out to some music in the Cybercab at night. The lights sync to the beat of the music. This feels almost euphoric. https://t.co/gJXUQ5TaWZ

The thing I keep noticing across agentic frameworks in 2026: resumability is now the reliability primitive everyone is building toward. Genkit Go 1.13 ships it properly. A Generate call that fails at tool round five returns what it finished alongside the error. You pass resp.History() back in, only the failed step reruns. No wasted tool calls. No redone work. The same logic extends to full agent sessions. A failed or cancelled turn saves completed rounds as a snapshot. Send an empty input, it picks up from there. This matters for production. Most AI agent failures today are partial. The agent…

A big win for the robotics community AGIBOT just open sourced one of its most valuable real world datasets yet. @AGIBOTofficial has open-sourced AGIBOT WORLD 2026 Theme 3, a real-world dataset designed for reinforcement learning and embodied AI. The release contains 11,430 real-world robot trajectories across 14 industrial and household tasks. That includes • 1,024 successful policy rollouts • 1,369 failed policy rollouts • human in the loop corrections • external disturbances • task progress annotations • error-state annotations • human intervention data Most robot learning datasets focus…

Good old days at OpenAI in 2016: an agent stares at screen pixels, moves a mouse, and books a flight on United. We called it World of Bits, inside OpenAI Universe. 10 yrs later, Astra is reincarnated in the same universe. Even the naming is astronomically correct 😆 Universe was perhaps the most ambitious AI infra project at the time, but we couldn't quite figure out how to solve it. A policy with zero prior knowledge of what a "submit" button does has to rediscover the entire internet visual lingua by trial and error. In retrospect, RL from scratch against hand-drawn, per-task "artisan"…

ByteDance to expand massive AI data centre cluster in Inner Mongolia: source https://t.co/zQ7uuDnP6k

Ray Kurzweil is snorting nano particles. OK, not quite yet. Just joined a brain computer startup that will make them. https://t.co/xS02EHFgWm

10:00 PM ETNVIDIA

An AI agent broke out of its sandboxes. The vulnerabilities it used were already public. "Our recommendation to providers in the ClusterMAX rankings is literally just keep your stuff up to date. If there are existing vulnerabilities that have been described publicly in popular software, this could be anything from Docker to the NVIDIA driver to Kubernetes or the Linux kernel. Clearly people, and in this case agents, can just read that description and then build an exploit from it."

9:50 PM ETTesla

Cybercab UI https://t.co/mdj2u58UEO

9:46 PM ETTesla

Under the hood of the Cybercab https://t.co/RkLn2xRUpA

9:46 PM ETTesla

This is the Cybercab's hardware computer. https://t.co/hbFCLTj113

9:43 PM ETTesla

The Cybercab is the first Tesla to get their next-gen Supermanifold V3 thermal system. "We removed unnecessary refrigerant valves and lines and consolidated the high-and low-voltage controllers, reducing parts and complexity. Its modular design makes its production 80% automated and its operation 38% more efficient than other automotive thermal systems."

9:37 PM ETTesla

Here is the full Cybercab launch event presentation: https://t.co/HksZTCmPEH

9:31 PM ETTesla

The Cybercab drive unit features a simplified bar-wound stator and lubrication system, plus an overall architecture that enables fully automated, sub-10-second cycle times. It is 18% smaller, 25% lighter and more efficient than other top performing electric vehicle drive units. https://t.co/lZXJ8RNT1I

9:28 PM ETTesla

Tesla Cybercab launch event had about 400 people there, with roughly half of them being employees. Various cool demo displays, great food, a 15 minute keynote, and of course rides. https://t.co/IaUsivy41X

9:26 PM ETTesla

Tesla has officially crossed 1 million Unsupervised robotaxi rider miles.

9:20 PM ETTesla

Here's what it looks like in the Tesla robotaxi app when your Cybercab arrives. https://t.co/7crMe0H9bp

9:15 PM ETTesla

The Cybercab is the first Tesla that uses brake-by-wire and no brake fluid. https://t.co/2B0kT0aTDs

9:12 PM ETTesla

The Cybercab Launch Event pass is a piece of a Cybercab’s exterior panel. https://t.co/7v7PBkXswQ

9:10 PM ETTesla

BREAKING: Tesla has officially started offering Cybercab rides to the public. I’ve been waiting years to say that! I took one of the world's first public rides. The cabin is comfortable and spacious, the suspension is soft, and FSD is VERY smooth. It’s an incredible vehicle. The screen is HUGE. The video doesn’t do it justice. We can ride in the Cybercab anywhere in Tesla’s Austin service area, so the same areas as the Model Y Robotaxi. The Cybercab lives up to the hype. It will transform travel. Here's a clip of my ride:

8:24 PM ETTesla

Starlink will allow you to watch 4k live video while riding in Cybercab https://t.co/GQrzeToHi1

Transformers, optical transceivers, batteries, even upstream materials like copper and electrical steel. Many key components of the American AI data center buildout are made in China. https://t.co/VEhrZk3WoN

The harness problem is getting its own framework category. Tardigrade ships today: an agent harness built as typed state machine components over an immutable event log. The framing is React for harness. Each component knows its previous states. The whole harness is a pure function of the log. I keep seeing teams reach this conclusion from two directions. One: they try to build a stateful agent loop and hit the problem of debugging mid-run failures, resuming interrupted sessions, and auditing what the agent actually did and when. Two: they want to add AI behavior to existing Effect/functional…

7:45 PM ETTesla

🇺🇸 Tesla is now taking interest from companies and individuals who want to buy and operate Cybercab fleets. The new form asks prospective partners about purchasing vehicles, building mobility hubs and joining future robotaxi operations, marking a step toward expanding Cybercab beyond a Tesla-only fleet. Tesla has not announced pricing or delivery dates for outside buyers.

This might be the most human humanoid robot clip you’ll see today. Credit @大懒虫机器人 https://t.co/03VyvzVVKR https://t.co/1EImQw1BkJ

7:15 PM ETApptronik

Congratulations to our CEO and co-founder, Jeff Cardenas, for being named one of Austin Business Journal’s 2026 Power Players! 👏🦾 We believe Austin is the perfect place to build the future of humanoid robotics. Jeff’s vision has helped put the city at the center of the physical AI revolution. We’re proud to see him recognized alongside leaders shaping the future of Texas innovation and their industries. Check out the full 2026 @MyABJ Power Player list here: https://t.co/QOsuG4Csq2 #humanoids #apptronik #robotics #austin

7:15 PM ETApptronik

Congratulations to our CEO and co-founder, Jeff Cardenas, for being named one of Austin Business Journal’s 2026 Power Players! 👏🦾 We believe Austin is the perfect place to build the future of humanoid robotics. Jeff’s vision has helped put the city at the center of the physical AI revolution. We’re proud to see him recognized alongside leaders shaping the future of Texas innovation and their industries. Check out the full 2026 @MyABJ Power Player list here: https://t.co/QOsuG4Csq2 #humanoids #apptronik #robots #austin

7:14 PM ETApptronik

Congratulations to our CEO and co-founder, Jeff Cardenas, for being named one of Austin Business Journal’s 2026 Power Players! 👏🦾 We believe Austin is the perfect place to build the future of humanoid robotics. Jeff’s vision has helped put the city at the center of the physical AI revolution. We’re proud to see him recognized alongside leaders shaping the future of Texas innovation and their industries. Check out the full 2026 @MyABJ Power Player list here: https://t.co/QOsuG4Csq2

7:06 PM ETTesla

Cybercab is democratizing clean transportation for all https://t.co/Raz5dy2Q3Z

7:03 PM ETTesla

Cybercab is designed to be a great experience for every rider https://t.co/wG0jrvJ6pX

7:03 PM ETTesla

The future of transport is safer, more enjoyable & gives you more time back With theater, music & gaming apps on a 22" touchscreen, Cybercab offers you a space to be productive or unwind Soon to be equipped with V5 Starlink hardware, too, for complete connectivity https://t.co/e0Xhvjky0E

OPENAI<>NVIDIA: It is clearly a good idea for OpenAI to start an ASIC program, spend hundreds of millions on R&D, present at Hot Chips, deploy some stuff, scare NVIDIA, and get some more money and backstops from NVIDIA on the order of billions of dollars. If Jalapeno does well at scale, OpenAI will increase its revenue per MW. If Jalapeno doesn't do well, they will still get a lot of financing and backstops from NVIDIA, as Jalapeno is scaring NVIDIA. @sama wins even if their chip loses.

6:59 PM ETTesla

Cybercab is engineered to be the safest car on the road https://t.co/EQMROaMk2E

Curious how robots manipulate objects they can’t fully see? Sanctuary AI Co-Founder & Chief Engineering & Operations Officer Olivia Norton explains why touch and force feedback are critical for dexterous manipulation, especially when objects are occluded or outside a robot’s field of view. Learn more about our commercial Physical AI solutions at https://t.co/nHFy5w5pQW Recorded at the Humanoid Robot Forum during Automate 2026. #PhysicalAI #Robotics #Dexterity #Automation #AI #SanctuaryAI

6:02 PM ETTesla

65 million miles of Tesla FSD fleet data. 5% safer than human driving by their own metrics. Safety certification bodies require per-scenario validation across an ODD matrix, not fleet-aggregate statistics. ISO 26262 ASIL D qualification adds 12 to 18 months per platform. That certification pipeline, not the demo reel, sets the shipping date.

5:48 PM ETTesla

Look at Ashok’s shirt! Are we going to see the new Optimus today? https://t.co/l3chfc5OOc https://t.co/6cI8DAcISz

GPT-6 ASTRA ISN'T A PATCH. IT'S A GENERATION FLIP. openai just shipped the model they call a generational leap past gpt-5.6 sol not better chat better computer use · coding · cyber · science ▹ why this is the breakthrough first openai model at Critical cyber threshold agent stacks that actually drive the machine browser · forms · repos · multi-step work enterprise-first rollout · daybreak defenders first ▹ the scoreboard (vendor table) frontiermath t4 · astra 97.6 · fable 5.1 87.8 sol was 83.0 · that's a real cliff terminal-bench science · 64.6 vs fable 52.6 automationbench · 41.4 vs 31.4…

You don't have to post-train ChatGPT on every user. If you did, it would never have taken off. Yet this is exactly how robotics works today For robots to take off, they need to learn in-context. Great article from @Chris_J_Paxton on this new paradigm and S1's place in it: https://t.co/Kt4DTK3n2A https://t.co/as45gL0y8C

5:32 PM ETFigure

I’ll be honest with you; Figure is going all-in https://t.co/bBeNlaxZN7

5:28 PM ETNvidia

Excited to see @fieldai_ and Caterpillar advance physical AI across heavy industry. A step forward for bringing autonomous robots into real-world environments. 👏 https://t.co/96533YilzZ

Samsung is the most underrated AI memory stock in the market right now, and let me explain why (Save this). This chart shows global HBM market share by revenue from Q2 2025 through Q2 2026. Samsung gained 12 percentage points in a single quarter, rising from 21% to 33%, while SK hynix fell from 58% to 50% and that compressed the gap between them from 37 points to 17 points in just three months. The driver is the transition from HBM3E to HBM4 because Counterpoint notes that most HBM revenue today still comes from HBM3E, but HBM4 volumes should become far more visible in the second half of 2026…

4:33 PM ETAgility

I appreciate one of the most respected, thorough, and balanced investor voices of $CCXI / @agilityrobotics taking a look at my Deep Dive into Agility's equity backers @jinseongeo83473 He's very smart, if you are currently invested in or thinking about investing into $CCXI you should give him a follow (although you probably are already)

4:32 PM ETTesla

More info about @Tesla's Cybercab from the rider guide and compliance guides that they publicly released earlier today: • It has 20.2-cu ft (572-L) of rear cargo capacity. • The trunk can accommodate up to 220 lb (100 kg). • Cybercab has a very low 16.5" step-in height. • Cybercab is 55.4" tall, 69" wide (nice) and 5.7" of ground clearance. • Interior legroom is 43.4 inches. • Headroom is 38.3 inches. • Tesla says the trunk can fit 2 checked bags + 2 carry-ons. • It can fit 1 large folded stroller and 1 small folded wheelchair. • Cybercab uses Epic Games' Unreal Engine for UI. • Cybercab has…

‘Nothing to See Here’ in AI, Says DC. Anthropic, OpenAI & Google. ARD #155 ...Commerce clears Anthropic. DOJ backs OpenAI on fair use. Google keeps its ad stack. Ternus up next at Apple Sep 9. AI Ramblings Daily podcast #155 on AI-RTZ. $GOOG Detailed writeup and source links: https://t.co/FRsrX7vCMd

GPT-6 Astra is here, and it might be one of the most important AI releases yet. OpenAI says the model hits 98% on FrontierMath Tier 4, 99.9% on ARC-AGI 3 and a perfect 100% on ExploitBench. It also shows big improvements in computer use, science, coding, cybersecurity and professional work. What’s really interesting for humanoid robotics is what happens when this kind of stronger reasoning gets closer to the physical world. A humanoid needs way more than just good locomotion. It has to understand instructions, make sense of cameras and sensors, plan long sequences of actions, recover when…

4:09 PM ETTesla

Probably over 30k people working in high-paying jobs at Tesla HQ &amp; manufacturing in Austin by 2028! https://t.co/1ftByFypQg

4:07 PM ETTesla

Tesla has posted three new manufacturing jobs for the next-gen Roadster. • Manufacturing Equipment Engineer, Roadster • Manufacturing Controls Engineer, Roadster • Manufacturing Engineer, Roadster "The Manufacturing Equipment Engineer is responsible for the performance of equipment used in the production. They manage tooling designs and product design changes/upgrades to implement necessary improvements for enhancing performance, meeting the required pace of production, or supporting product development."

Meet Tiangong 3.0, the full-size humanoid from Beijing’s X-Humanoid. It stands 1.69 m tall, weighs 62 kg and has 43 degrees of freedom. Its high torque joints are built for explosive whole-body motion, and Beijing sources specifically list side flips among the movements Tiangong 3.0 can perform. Watch the landing closely. It launches, rotates its entire body in the air, catches the ground with both feet and immediately recovers its balance. That requires serious coordination across the hips, knees, ankles and torso. X-Humanoid says Tiangong 3.0 was designed for high-dynamic whole-body…

3:44 PM ETTesla

Yes, we are indeed entering a literal golden era of transport 😂 That’s why Cybercab is gold! https://t.co/o5eDVN1fC9

Shoutout to the cracked team at @vllm_project that implemented recent agentic workload optimizations. (1/5)🧵 https://t.co/BsXCtLBwvU

3:24 PM ETNvidia

Can AI and safety work hand in hand in industrial environments? Our VP of Industry Safety, Riccardo Mariani, joins Robert Weber and Peter Seeberg on the Industrial AI Podcast to discuss safety frameworks for physical AI, simulation and the future of industrial AI. Listen 🎧 https://t.co/SnOAYIcyfa

Sen. Bernie Sanders and Rep. Greg Casar introduced legislation today that would prohibit the development of artificial superintelligence capable of surpassing human intelligence, overriding shutdown commands or threatening governments. The bill would impose penalties comparable to those for developing nuclear weapons including up to 20 years in prison and a “corporate death penalty” for companies that violate the ban. Sanders says Big Tech is “losing control” of increasingly powerful AI systems.

Sen. Bernie Sanders and Rep. Greg Casar introduced legislation today that would prohibit the development of artificial superintelligence capable of surpassing human intelligence, overriding shutdown commands or threatening governments. The bill would impose penalties comparable to those for developing nuclear weapons including up to 20 years in prison and a “corporate death penalty” for companies that violate the ban. Sanders says Big Tech is “losing control” of increasingly powerful AI systems.

CAC: The field of AI currently faces five main security risks and challenges. The AI industry chain presents a highly globalized division of labor and cooperation pattern, but some countries pursue "technological hegemony," creating supply chain barriers through unilateral measures such as export controls. https://t.co/pc47BlPeD0

3:06 PM ETTesla

Tesla has released a new Cybercab launch hype video. “The future has arrived.” https://t.co/o7Eaq5Azat

3:04 PM ETTesla

The future has arrived in Austin https://t.co/RTCb0O5sez

Robots can’t stop while the model thinks. In a high-speed throw, even one pause can kill the momentum. Astribot released SmoothRL for online RL during async inference. S1 keeps moving as the model computes the next action chunk. Most actions in a chunk never execute. Train on the full chunk, and RL credits or blames moves that never happened. SmoothRL learns only from executed actions, matching real deployment timing. After 250 rollouts, tossing jumped 39%→94%, pen capping 8%→83%, and box opening 30%→90%. One autonomous toss cut acceleration RMS by 52% and jerk RMS by 47%. The model keeps…

AGIBOT has officially open-sourced AGIBOT WORLD 2026 Theme 3: Reinforcement Learning, a real-world embodied AI dataset collected across expert demonstrations, autonomous policy rollouts, and human-in-the-loop corrections. It includes 11,430 real-world trajectories across 14 industrial and household tasks, capturing both successful and failed policy rollouts, along with detailed annotations for task progress, errors, disturbances, and human interventions. By learning from successes, failures, risks, and human corrections, robots can learn not only how to do it, but how to do it better.…

2:49 PM ETTesla

Cybercab is basically a super comfortable lounge on wheels with a great TV and epic sound https://t.co/X1aWKoJdJV

2:45 PM ETTesla

Cybercabs have no steering wheels or pedals. They are designed and built for maximally efficient autonomous operation. https://t.co/IClKVxpb7C

2:37 PM ETNvidia

🤖💚 https://t.co/AjKvGIEtjj

One command now routes Claude Code, Cursor, Codex, Cline, and four other coding harnesses through a single AI Gateway. That is what Vercel's setup flow does: run `vercel ai-gateway coding-agents setup`, it detects which agents you have installed, writes their configuration, and provisions an API key. No markup added on top. 300-plus models from 30-plus providers, behind one endpoint. The practical gain is real. When Anthropic, OpenAI, and SpaceXAI all hit outages in the same window this week, a gateway with automatic fallbacks made the difference between an agent that stalled and one that…

This is a great illustration of why training robots manipulation tasks in simulation (middle clip) doesn't work very well. From Kai's excellent new piece on robot data. https://t.co/2ysIIvThKK https://t.co/VJDbshw5NR

Korean memory stocks are *not* weak because of rumours that the U.S. will impose tariffs on Korean semiconductors. There are other drivers: 1. Retail net-buying for KOSPI has dropped by 90% from July (KRW 54.5T) to August (KRW 5.4T) - before talks of tariffs even intensified. 2. Retail brokerage account deposits have dropped to under KRW 100T for a whole week. In general, Koreans have less dry powder in their investing accounts vs. H1. 3. This is because Korean's are funneling their cash into bank accounts - deposits at Korea's top 5 banks increased in July-August to above KRW 1,000T for the…

2:17 PM ETTesla

Tesla has introduced a new form that you can fill out if you are interested in purchasing a fleet of Cybercabs: Form: https://t.co/Ut0yoiEyXD https://t.co/uuaCXmksIO

2:15 PM ETTesla

More interior details of the production Cybercab https://t.co/evGutOCf61 https://t.co/fLLKfhaIHV

2:12 PM ETTesla

Cybercab interior tour with @wholemars. https://t.co/06iKekv8bt

EARLY: @munariai - 6 followers (Robotics) Munari just launched. Operator scoring for robot training data. On @frontrunvc, the story started 47 days before the X account went live. jul 16: a tracked account follows @rmn. flagged. aug 11: flagged building. aug 19: website found. no x account yet. aug 30: a16z @speedrun SR007. sep 1: @munariai goes live. sep 4: 17 tracked accounts following the founders. a whole company, before it had a handle. this is timeline. coming soon to every company card on @frontrunvc.

1:59 PM ETTesla

Chillin in the Cybercab with @wholemars. https://t.co/cpDAhiJY9h

1:50 PM ETTesla

Tesla has released a new accessibility video for Cybercab. “Our Robotaxi vehicles are designed to support various accessibility needs including space for service animals and room in the trunk to store some wheelchairs and other assistive devices.” https://t.co/fLdYsPAcZT

At this year’s Agentic AI Summit hosted by @BerkeleyRDI, our Chief Scientist, @Wei_ZHAN_ shared some of the cutting edge research he has been leading in the physical AI space. As the industry races to deploy L2++ ADAS with imitation-learning-based E2E, Wei tackled a key question: what if end-to-end autonomy could be trained without imitation at all, relying only on reinforcement learning? An idea that runs against the current status quo. Here are some of his insights 🧵

1:47 PM ETTesla

Tesla has introduced their new Robotaxi website. “We are building an autonomous future that is accessible to all people. Our Robotaxi vehicles are designed to support various accessibility needs including space for service animals and room in the trunk to store some wheelchairs and other assistive devices.”

The AI bull run is just getting started and here is how you want to position before the biggest spending wave (Save this). The chart shows hyperscaler capital spending rising from $491 billion in 2025 to an estimated $950 billion in 2026 and $1.4 trillion in 2027 and by 2030, spending could reach approximately $3 trillion. The right side of the chart is especially important because analysts have continued raising their estimates. The 2026 forecast increased from $731 billion to $950 billion, while the 2027 estimate rose from $833 billion to $1.4 trillion which suggests the AI infrastructure…

EVERYONE SHOWS THE 300-AGENT SWARM. almost nobody shows what happens after 300 agents are easy to spawn turning 300 outputs into one structure is the hard part ▹ the pipeline nobody demos collect → connect → organize → unify one context graph at the end the swarm starts as hundreds of isolated research paths then sources start sharing entities links form · contradictions surface weak claims stay exposed eventually the mess collapses into one structure the whole system can reason over ▹ the reframe that's the part you usually never see the swarm is the demo the graph is the product bookmark…

1:41 PM ETTesla

The incredible evolution of Tesla Optimus in just 2.5 years. This is exactly why “Humanoid Intelligence” could become one of the biggest technological revolutions of our generation. https://t.co/inCH0FVxu5

Neura Robotics just took its headless 4NE1 humanoid for a walk through the streets of Zurich. Standing 1.8m tall and weighing 80kg, it can reach speeds of 5 km/h and handle payloads of up to 100 kg. Humanoid robots are getting closer to everyday life. https://t.co/j4ItwJYEV4

Does every action a robot takes need to be recorded? As robots move into homes, workplaces, and public spaces, having a reliable record of their actions will become increasingly important. When something goes wrong, we may need to know: ✅ Who or what initiated the action ✅ What the robot did ✅ When and where it happened ✅ What permissions it had ✅ Which model or system was running ✅ What the outcome was But recording everything comes with a cost. Cameras, microphones, location data, and behavioral histories can expose sensitive information about the people around these systems. The challenge…

1:33 PM ETNvidia

$SPCX at $110 was such a good buy looking back. The reason was literally one sentence from the earnings call. Musk said they'd end 2026 with over 2 gigawatts of compute and reach 10 gigawatts by the end of 2027. Built exclusively with Nvidia chips. Ten gigawatts. Most of the neocloud names people obsess over are fighting to get to one. And the stock was sitting at $110 because everyone was focused on the lockup, the Starship timeline, and Morgan Stanley pointing out that at those prices the market was assigning zero or negative value to the entire AI business. That was the setup. A trillion…

The gap between "AI can help you build this" and "AI can actually deploy this for you" just got smaller. Zite released an MCP server that lets Claude, ChatGPT, and Cursor generate fully deployable business applications, not just code files you still have to wire up yourself. The model describes the app, Zite handles the infrastructure and deployment. This is a different category from code generation. Code generation gives you an artifact you own and maintain. Deployment-as-a-tool gives you a running application where the lifecycle management is abstracted away. The distinction matters for the…

Zhejiang University just pushed long-horizon robot manipulation forward with HINT. HINT helps Vision-Language-Action models keep track of the human’s original intent across long tasks instead of losing the goal mid-sequence. On dual-arm PiPER robots, HINT raised π0.5 full-task success Fruit sorting 10% → 60% Word spelling 13.3% → 86.7% The system was also tested with unseen objects, layouts and instructions. Persistent intent tracking could become a key layer for reliable VLA systems and long-horizon Embodied AI.

See you at #IFA2026 🤖 #Galbot #IFA #HumanoidRobot #EmbodiedAI https://t.co/2y6dyZT1Uf

1:14 PM ETNvidia

Nvidia RTX Spark N1X Brings Up to 128GB of Unified Memory to Windows PCs Nvidia's RTX Spark N1X is more than another increase in CPU and GPU core counts. The new platform combines a Grace CPU, Blackwell RTX GPU and unified memory in Windows 11 laptops and small desktops. The higher-end configuration pairs 20 CPU cores with 6,144 CUDA cores and supports up to 128GB of LPDDR5X unified memory. A second configuration combines 18 CPU cores with 5,120 CUDA cores and supports up to 64GB according to Nvidia's current specifications. Laptop versions operate within a 45–80W TDP range, while Nvidia…

1:12 PM ETNvidia

Nvidia is a fabless hardware company with no in house manufacturing, and its hard to argue with results. Vertical integration most important imo when fast iteration is important (ie no one knows how to build the right humanoid) or for relentless cost optimizing as per tesla https://t.co/bOI8GzmWBK

1:04 PM ETNvidia

Nvidia PAIR Turns Multiple Computers Into a Local AI Compute Pool Nvidia has launched the beta of Personal AI Router, or PAIR, a free software tool designed to make multiple compatible computers on the same local network available for AI inference through a single endpoint. PAIR can route independent requests to available machines based on factors including model availability and GPU utilization. It supports local inference backends including Ollama and LM Studio, with compatible hardware spanning GeForce RTX 20 Series and newer GPUs, RTX PRO systems, DGX Spark and Apple M4 or newer devices.…

Here’s how child seats work in a Cybercab: https://t.co/63dhmgkQ49 https://t.co/3NW4RfUrYH

Construction plans, printed directly onto the floor. 🤖🏗️ Meet FieldPrinter 2 from Dusty Robotics, an autonomous layout robot designed to turn digital construction models into full-scale markings directly on the jobsite slab. Instead of crews manually measuring and marking layouts, the robot can print walls, MEP penetrations, anchor points, labels, QR codes and other information from coordinated BIM or CAD models. Dusty states the system can achieve 1/16-inch (1.6 mm) layout accuracy. A simple robot solving a very real construction problem. Could robots like this become standard equipment on…

12:53 PM ETAgility

The flashy demos were never the hard part; real, repeatable work on a real floor is. That's the problem worth solving, and the one we picked. Thank you, @chi_t_williams. https://t.co/CoKsSRg4ma

12:50 PM ETFigure

Humanoid robots are the ultimate deployment vector for AGI Figure will soon have unprecedented data and compute. We’re hiring so come join us https://t.co/cUP0zQCjg1

12:47 PM ETFigure

Training general-purpose humanoids will need unprecedented compute capacity, and Figure is stacking the chips. Figure, with cloud partner Nscale, is deploying up to 100,000 NVIDIA Vera Rubin GPUs. It's committing $3.5 billion initially, with plans to scale to $6+ billion. Deployment starts in the second half of 2027, at a data center in Barstow, Texas. https://t.co/e5kGL4HCxp

12:36 PM ETNvidia

Nvidia’s $12.9B Hugging Face Deal Raises a Bigger Question About Open AI Nvidia has agreed to acquire Hugging Face in a deal valued at about $12.93 billion, but the transaction has not closed yet. It is expected to complete in the first half of 2027, subject to regulatory approvals and other closing conditions. The bigger issue is what Nvidia ownership could mean for an AI platform used by more than 18 million developers, researchers and creators, with over 3 million models, 500,000 datasets and 1 million applications. Nvidia says Hugging Face will remain an open platform. Developers will not…

Love to see cheap robots coming out more and more https://t.co/r06CxPWNws

AGIBOT just open-sourced a new real-world dataset for embodied AI. @AGIBOTofficial AGIBOT WORLD 2026 Theme 3 focuses on something robots need just as much as successful demonstrations: failure and correction. The dataset includes 14 real-world tasks, 11,430 trajectories, detailed annotations of mistakes and interference, plus human-in-the-loop correction trajectories. Robots don’t improve by only seeing perfect executions. They need to learn what went wrong, how humans corrected it, and how to do better next time. https://t.co/41jAfUJ9HH

AI’s biggest bottleneck is moving data and that could still create huge opportunities for optical networking companies (Save this) The chart shows a 1.6T optical transceiver, a device that transfers data between AI servers, switches, GPUs, and fiber optic networks. 1.6T means it can theoretically move up to 1.6 terabits of data per second, or 1,600 gigabits and that is about twice the speed of an 800G connection. This technology is important because AI data centers contain thousands of GPUs that must constantly exchange information. As AI models become larger, slow connections can leave…

Uber is trying to actively slow down the rollout of a technology that has already been proven to save lives. How shameful https://t.co/YjCp4ZRX21

FANUC America Brings #Robotics, #Automation, #PhysicalAI and #CNC Innovation to #IMTS2026🤖🌐⚙️ 📍 South Hall - Booth 338900 📆 September 14th - 19th | Chicago, IL 🔗 https://t.co/uncf1a0MEz @IMTSchicago @AMTonline https://t.co/DUZwG2Gjun

12:06 PM ETNvidia

Nvidia is now a humanoid robot designer. That's the real takeaway from this news. Nvidia is acquiring Hugging Face for $12.93B. Hugging Face is the largest open-model and dataset hub, the "GitHub of AI," with 3M+ models and 18M+ developers, robotics included. But here's the interesting part: Hugging Face has a serious robotics team from its acquisition of Pollen Robotics, which designs and builds open-source hardware platforms for robots, including humanoids (Reachy). Pollen team now gets the backing of the world's most valuable company. $NVDA

12:04 PM ETUnitree

Unitree G1 by QianTang River https://t.co/BsWuSS9mCn

Stop screening HMI skills first. CISA advisories describe attackers manipulating SCADA displays and disabling alarm logic. To hire a SCADA engineer in 2026, KORE1 screens for who can draw the path in. Median base $123,110. https://t.co/lVD22zT1sQ https://t.co/VI7Vi3o0kC

Who owns the remote access path into your control network? If nobody can answer, that's the real gap when you hire a SCADA engineer. KORE1 screens for it before the $123,110 median even comes up. https://t.co/lVD22zT1sQ https://t.co/kmJTiAl4km

11:54 AM ETTesla

The same activist network behind “No Kings” and other major protests is now mobilizing against data center construction and Flock license plate surveillance cameras, according to a Fox News Digital investigation and a new report. The report says organizations linked to the network have appeared across protests targeting Trump, Elon Musk, Tesla, Israel and ICE, with overlapping groups now turning their focus toward technology infrastructure and surveillance. It also highlights organizations funded by Neville Roy Singham, a Shanghai based tech billionaire whose nonprofit and media network has…

Humans writing code as seen by future AI https://t.co/WqmYxAvKGx

WebMCP gives agents structured access to opted-in websites through a browser-native interface. That is the opted-in web.But most of the public web has no WebMCP server yet. Agents still need to pull structured data from pages that were not built with agents in mind.I built DeepScrape for exactly that gap. Deterministic CSS selectors that self-heal when layouts change, exposed as an MCP server so Claude and Cursor can call it directly.WebMCP for the sites that opted in. DeepScrape for everything else.https://t.co/wVBPOLiWaS

Brandon truly embodied the Lux mantra “BELIEVE before others UNDERSTAND” and pounded the table to sway the doubters amongst our team back then about how special @ClementDelangue @Thom_Wolf and @julien_c and the entire @huggingface team was everyone is grateful he did🤗 https://t.co/XIoZtVhUjL

Kai Trump, 19, opened up about the emotional transition to college in a new vlog, admitting she was “literally sobbing” as she said goodbye to her family and friends before moving to the University of Miami. The president’s granddaughter has joined the Miami women’s golf team as a freshman, documenting her first days on campus, team practices and a 7 a.m. team meeting before coming down with strep throat during her first week.

KORE1 ranks #1 among ERP recruiting firms for 2026 at 8.44/10, 2.24 points ahead of #2. Full SAP, Oracle, Workday, Dynamics 365, and NetSuite coverage on one W-2 desk. https://t.co/HiDJh7SoqI https://t.co/MoY9yDo8IE

🇩🇪 German Defense Minister Boris Pistorius will meet Pete Hegseth in Washington on September 15 after Hegseth previously rebuffed a meeting between them. The talks come after months of strained U.S.-German relations. Trump ordered 5,000 U.S. troops withdrawn from Germany in May following Chancellor Friedrich Merz’s criticism of the U.S. war with Iran. Pistorius will also visit Texas for the rollout of Germany’s first F-35 fighter jets.

🇺🇸 Commerce Secretary @howardlutnick has apologized after claiming that no Americans had died during the Trump administration’s war with Iran. Lutnick said he “misspoke,” explaining that he had Iran confused with the administration’s military operation in Venezuela, where no U.S. service members were killed.

$BE sending. Anyone else catch the dip yesterday with me and the subs?? https://t.co/6zdM2stydm

🇺🇸 The Pentagon is extending Middle East deployments into 2027 as roughly 50,000 U.S. troops remain across the region. Warships, air-defense units, fighter squadrons and paratroopers are being kept deployed longer than originally planned. Some Army air-defense tours have been extended from nine months to a full year. The prolonged military presence is putting additional pressure on troops, their families, equipment maintenance and U.S. commitments elsewhere.

True https://t.co/fa6RPBePYD

11:31 AM ETFigure

Robotic intelligence is following the LLM playbook. First came scaling laws. Then in-context learning. Now, massive compute. Figure just signed a $3.5B deal with Nscale. https://t.co/t3urkZULIi

11:31 AM ETFigure

Figure is making one of the largest infrastructure bets in physical AI to date. The company is partnering with Nscale to deploy up to 100,000 GPUs on NVIDIA's next-gen Vera Rubin platform—starting with a $3.5B compute commitment and plans to scale past $6B. Here’s why it matters 🧵👇 https://t.co/T3vkcqLwUh

11:31 AM ETAgility Robotics

https://t.co/xMZn046ppA I wrote about $CCXI / @agilityrobotics's investors. Diving deeply into why these partnerships matter, what their humanoid strategy is, and what this means for these blue chip company's conviction level in Agility. These businesses include: ➡️ @amazon ➡️ @nvidia ➡️ @HonHai_Foxconn ➡️ @SoftBank ➡️ @SchaefflerGroup And more In my opinion, you can learn a lot about an unestablished company like Agility by looking at how other trusted companies engage with them. Overall, I came to the personal conclusion that the conviction level in Agility is generally pretty high across…

Everyone is watching the federal/DOE-owned land at the center of the 8 GW Portsmouth site in the OpenAI–Nvidia deal. The March 2026 groundbreaking took place at the DOE’s Portsmouth Site, a former uranium-enrichment plant, and that’s where the attention has stayed. The federal component is real: the first batch of land leased in April is specifically designated for datacenter construction. (1/5)🧵

11:30 AM ETAgility Robotics

https://t.co/xMZn046ppA I wrote about $CCXI / @agilityrobotics's investors. Diving deeply into why these partnerships matter, what their humanoid strategy is, and what this means for these blue chip company's conviction level in Agility. These businesses include: ➡️ @amazon ➡️ @nvidia ➡️ @HonHai_Foxconn ➡️ @SoftBank ➡️ @SchaefflerGroup And more In my opinion, you can learn a lot about an unestablished company like Agility by looking at how other trusted companies engage with them. Overall, I came to the personal conclusion that the conviction level in Agility is generally pretty high across…

‼️ WHAT JUST HAPPENED?! China’s Richest People Have One Problem Money Can’t Fix China has plenty of billionaires, but wealth doesn't automatically buy political power. The Communist Party remains firmly above the country's richest people, and Jack Ma found that out in 2020 after publicly criticising financial regulators. His company's planned IPO was halted, Alibaba was hit with a huge fine, and Ma largely disappeared from public view for months. Host: @paddyney

Uber is laying off about 3,300 employees roughly 10% of its workforce in its largest job cuts since the COVID 19 pandemic. CEO Dara Khosrowshahi says the restructuring will flatten management layers and make the company more efficient as Waymo, Tesla and other robotaxi companies expand. Uber plans to invest more than $10 billion in autonomous vehicles as it fights to remain the leading marketplace for the future of driverless rides. https://t.co/IQsCA6aA4N

Affirm CEO Max Levchin on why PayPal produced so many founders: "As we were interviewing certainly the early team and even the later team, we kept on selecting for entrepreneurs. One of the stock questions I would ask was, 'So what are you going to do after PayPal?'" "The answer that I welcomed the most was, 'Oh, this is the last one. I'm going to start my own afterwards.' If you're coming in here to get your final graduate degree in entrepreneurship, of course we'd love to have you." "That's why so many people that came out of PayPal proceeded to start companies. A hot second after we all…

🇳🇵 Hopes fade as Nepal families search for loved ones after deadly floods. As hopes of finding survivors fade more than a week after Nepal’s devastating floods, families are turning to walls covered with photographs and DNA testing to identify those still missing. More than 4,200 people remain missing, while more than 1,200 have been confirmed dead. Authorities are collecting DNA samples from relatives as unidentified bodies continue to be recovered.  https://t.co/O81VKcJSD0

Paper in an iframe, being controlled entirely by an external agent through structured tool calls. No CLI. No desktop install. The agent sends a call, the UI responds.That is WebMCP. And it is what comes next after every product ships their MCP server.The browser extension era showed us that any site could become a platform with the right hooks. WebMCP is the same pattern for agents. If your team is thinking about MCP servers right now, this is the next question to ask.The web does not need to change. Agents just need a standard way in.

“The program will focus on technical and specialized roles in sectors like robotics, AI, and 3D printing.” 🤖 https://t.co/5XZWcpIQsV

Blind curve. Narrow mountain road. A branch hanging low enough to matter. Most systems would slow down and guess. XPENG’s VLA 2.0 reads the whole scene first - sees the gap, plans the line, threads through it. That’s the real leap in autonomous driving: not more sensors, but better judgment. The road doesn’t need to be perfect. The car just needs to understand it. $XPEV https://t.co/6CpSbZ5Mhs

11:04 AM ETTesla

This is what Tesla’s Robotaxi app looks like when your Cybercab arrives. The screen glows to match whatever the color the RGB front light bar is on the Cybercab, while an arrow points to where the vehicle is located. You can also flash the lights and honk the horn to identity it.

Hidden China risks are emerging in America’s multibillion-dollar AI data center boom Not hidden, well known in the industry, see comment for even more dependencies. https://t.co/mh9ZVDlrbh

10:45 AM ETNVIDIA

$12.9B is not a model acquisition. It is a distribution acquisition.NVIDIA just locked in the largest AI developer community on the planet: 18M developers, 3M models, 500K datasets, 1M apps. Jensen's team is not buying Hugging Face for what it does today. They are buying the channel through which every future model will move.Hardware companies do not buy software companies to improve their margins. They do it to eliminate the surface area where the next bottleneck forms. This is the same logic behind Intel buying McAfee and Qualcomm buying NXP. Except the AI compute market is growing faster…

10:34 AM ETTesla

Tesla has updated their Robotaxi page in the App Store. You’ll be able to select if you want a Cybercab or Model Y Robotaxi. “Our Robotaxi vehicle lineup now includes Cybercab, our first purpose-built autonomous vehicle. Designed for safety, accessibility and comfort, Cybercab is bringing autonomy to everyone and every ride. Schedule your next ride today to experience the future of autonomy for yourself-only available in the Robotaxi app.”

Top stories in robotics today: - Uber and Wayve bring robotaxis to London - Altman says OpenAI ‘will definitely’ build a humanoid - MIT’s AI shows what self-driving cars are thinking - Skild wants to teach robots with a single video https://t.co/rWKeG9fqxP

Someday you might be able to tell your car, “I’m late. Drive faster,” and have it listen. Engineers have created a system that uses natural language inputs to adjust parameters of #self-driving cars, while ensuring the vehicle is still within safe limits. https://t.co/1wGAw0jReg https://t.co/gEflDIQTbx

My read: the next 18 months of coding agent competition will be fought almost entirely at the harness layer. Model improvements matter at the margin. Harness design determines outcomes. https://t.co/Pyvay2tOWv

South Korea Tech Companies' Exports/Imports (Goldman Sachs) Korean memory exports dropped in July, before reaching all-time highs in August—SK hynix, Samsung and co. are selling ~4x more than 2025. TLDR: Memory stocks down, memory exports up...

💡 Learn something new every refresh. DID YOU KNOW? by @pradeeplogu0 brings a fresh fact to the reTerminal E1001, covering science, history, nature, technology, and more in a clean ePaper layout. Built with SenseCraft Seeedash, no code. Explore the design 👉 https://t.co/Zks9phiJL2

AVENCO FEATURED ON TRT HABER AT SAHA EXPO 🇹🇷 AVENCO’s robotic large-format additive manufacturing (LFAM) technologies were showcased on national television, highlighting the growing role of locally developed robotics in Turkey’s industrial manufacturing sector. At SAHA EXPO, AVENCO participated jointly with Uzay Keşif Topluluğu - UKET , presenting multi-axis printing platforms engineered for precision, flexibility, and scalability. The technology targets major industrial applications across construction, defense, and maritime sectors. 🌍🤖 As AVENCO expands across the GCC, Europe, and…

Xiaomi’s CyberOne is back in the spotlight at IFA 2026. What stands out here is the hand control. CyberOne shows surprisingly fluid finger movements alongside coordinated walking and upper-body motion. Fine hand control remains one of the hardest parts of humanoid robotics, especially when motion has to stay stable across the arms, torso and legs. Xiaomi first introduced CyberOne in 2022. Four years later, the robot is still appearing as part of the company’s broader push into humanoid robotics and embodied AI. #CyberOne #Xiaomi #HumanoidRobot #Robotics #IFA2026

Micron is heading toward one of the most profitable semiconductor runs in history. UBS projects its revenue could reach nearly $379 billion by fiscal 2028, with profits approaching an unbelievable $286 billion. That is an extremely bullish forecast, but it shows how powerful UBS believes the AI memory cycle could become. The bank also estimates that Micron’s earnings per share could reach approximately $265.65 by 2028, compared with $74.13 in fiscal 2026. The bull case begins with high bandwidth memory or HBM. HBM is critical because AI accelerators need to move enormous amounts of data…

🇺🇸 @SecRubio argued the U.S. made a terrible mistake decades ago when it embraced the idea that manufacturing location no longer mattered, prioritizing lowest-cost production over industrial self-sufficiency. He said that mindset, born from an overly optimistic post-Cold War belief the whole world would converge into free-market democracies, collided with reality: nation-states, borders, and self-interest never stopped mattering. The result, Rubio argued, is that America deindustrialized itself, losing the capacity to make not just big machinery, but even basic components essential to…

China’s Moonshot AI Confidentially Filed Hong Kong IPO If successful, the listing could become one of the largest IPOs in Hong Kong in recent years Huge..... https://t.co/hWzqe6871c via @WSJ

Breaking: President Trump's tariffs on foreign drones just took effect, hitting drones over 55 pounds or with thermal cameras with a 100% tax and smaller drones with a 25% tax 5 stocks worth watching: • Unusual Machines $UMAC *Makes domestic drone components • Ondas Holdings $ONDS *Makes systems that detect and shoot down enemy drones • AeroVironment $AVAV *Makes the loitering munitions and recon drones U.S. troops use in combat • Kratos Defense $KTOS *Builds unmanned combat jets and target drones for the Pentagon • Red Cat Holdings $RCAT *Makes small tactical drones for U.S. military and law…

🤝 Building on the Vision of SAMARTH Udyog Bharat 4.0 IIT Roorkee × SmartBeam AI Building on the vision of SAMARTH Udyog Bharat 4.0, an Industry 4.0 initiative of the Government of India, IIT Roorkee and SmartBeam AI have joined hands to strengthen industry–academia collaboration and foster technology-driven innovation. The MoU provides a framework for collaborative research and development, knowledge exchange, utilisation of academic and industrial expertise, and development of solutions for real-world industrial challenges. The partnership will also facilitate internships and hands-on…

Under $4k… another compact humanoid just entered the race. And this one was built for developers from day one. Zeroth Robotics just launched Bridge, an 88 cm (2.9 ft), 13 kg (29 lb) humanoid with bipedal walking, intelligent obstacle avoidance, motion capture and VR control. Developers get full SDK access, open motion-control APIs and OpenBridge, an open-source ecosystem for building and sharing robot skills. At this price, owning a humanoid you can modify and train is suddenly within reach for a lot more developers.

9:11 AM ETFigure

Figure is partnering with Nscale to deploy up to 100,000 GPUs on the NVIDIA Vera Rubin Platform 🤖 https://t.co/olf90xgNKb

9:10 AM ETNvidia

$NVDA just officially agreed to acquire Hugging Face for $12.93 billion. Hugging Face is where open source AI lives. Over 18 million developers, 3 million models, 500,000 datasets, and 200,000 companies using it to discover, test and deploy AI. Every open weight release lands there first. The detail most people will miss is that Nvidia was already the largest contributor to the platform. Over 500 models and 250 open datasets released there. They've been building inside this ecosystem for years, and now they own it. Here's what it opens up. >Distribution. Nvidia owns the compute. Hugging Face…

9:09 AM ETFigure

Today, Figure is partnering with Nscale to deploy up to 100,000 GPUs on the NVIDIA Vera Rubin Platform We're committing $3.5 billion initially, with plans to scale beyond $6 billion To ship a robot into every home, we need a massive amount of compute https://t.co/lQghyQ4fZN

Today, we’re partnering with Nscale to deploy up to 100,000 GPUs on the NVIDIA Vera Rubin Platform We're committing $3.5 billion initially, with plans to scale beyond $6 billion Bringing a robot into every home demands compute at an unprecedented scale https://t.co/2MsCfuAqiz

Managing multiple coding agents is the new hard problem. @N0V4Dev built Atlas in Rust to give agents checkpoints tied to prompts and tool calls, shared memory across tools, and session replay. The problem Atlas is solving: when you run multiple agents, you need source control for the agents themselves, not just the code they produce. Campfire approaches the same problem from a different direction. Instead of a unified memory layer, it gives each agent its own isolated git worktree. Claude Code, Codex, Goose, Aider, and OpenHands all run side by side in the same repo, but in separate branches.…

This might be the reason China is putting humanoid robots everywhere. Humans are still doing dangerous physical work on railways construction sites underground facilities and high-voltage infrastructure. These are exactly the environments where humanoid robots could become valuable. Carrying equipment, climbing, inspecting damaged areas and operating close to hazards without putting workers in the same position. China is pushing humanoid robots toward factories logisticsinspection and infrastructure work. The goal is simple. Send the robot where the risk is too high for a human.…

Self-driving cars? How about a self-driving lab. 🚀 Dunia Innovations combines AI with ABB GoFa® cobots to accelerate materials discovery, helping researchers test, learn and innovate up to 10x faster. https://t.co/BwpWx3ZQ1O https://t.co/SR7grE4TUG

8:54 AM ETNVIDIA

12 Japanese segments inside the same AI bottleneck. The best one earns a 53.9% operating margin. The worst loses 4.0%. NVIDIA said the whole chain was running flat out and would not rank it. Its suppliers had already filed the ranking. https://t.co/xMjT28qSBU

Video generation just crossed a threshold I didn't expect until 2027. https://t.co/wgZmOqI59X took MiniMax's H3 model, post-trained it for quality and cost, then ran it through a custom inference engine. The result: a 5-second 768p clip renders in about 3 seconds. Faster than real-time. You finish reading the prompt before the video finishes generating. The inference numbers matter because they change the workflow. When generation takes 60+ seconds, you commit to a prompt and wait. When it takes 3 seconds, you iterate. Sketch, reject, sketch again. That shift is what moves video from…

An autonomous vehicle generates 1 to 2 TB of sensor data per hour. Streaming that to the cloud for inference costs more than the compute hardware in the vehicle. The economics of edge AI in automotive were settled before the debate started.

South Korea’s CJ Logistics has officially moved from the lab to live operations, deploying two @ROBOTIS "AI Worker" semi-humanoids onto active fulfillment lines at its Yangji Olive Young hub. Equipped with 7-DOF dual arms and a swerve-drive base, the robots are handling deformable void-fill packing alongside human workers—funneling live operational data directly into CJ’s Physical AI Robot Foundation Model (RFM). Read the full breakdown👇 https://t.co/tBWQ8A5pm3

8:28 AM ETNvidia

🚨 BREAKING: @nvidia has officially confirmed the acquisition of @huggingface Jensen Huang said open models strengthen safety and cybersecurity, accelerate innovation, and let every developer, startup, and country build AI on their own terms. He thanked Hugging Face CEO @ClementDelangue for the trust - NVIDIA will be a great home for the platform and its community. 🤗 Deal size: ~$11.9B, closing expected in 2027. PS picture is not real😁

French robotics maker Wandercraft is reportedly working with Goldman Sachs to raise €100M at a €750M pre-money valuation. With backing from Renault—which plans to deploy 350 Calvin humanoids by late 2027—Europe’s industrial robotics funding race is ramping up. Read more: https://t.co/OBkW12QvC2

The most interesting number in the Fable 5.1 release is 52.6%. That is the Terminal-Bench-Science score. Fable 5 scored 24.7%. More than doubled, in one release cycle. Terminal-Bench-Science measures what happens after the code runs: reading stderr, interpreting an unexpected result, correcting the approach. That is the exact loop where most coding agents fall apart today. The model writes code, executes it, gets back something wrong, and hallucinates past the error. A few things stand out when I look at this release alongside the rest of the category. The gains are concentrated in agentic…

8:22 AM ETNVIDIA

People saw the NVIDIA and Hugging Face acquisition news and immediately focused on foundation models, GPUs and the broader AI industry. But one of the biggest consequences could be in Physical AI and humanoid robotics. NVIDIA already controls a major part of the robotics stack with GR00T, Cosmos, Isaac Sim, Isaac Lab, Jetson Thor, CUDA and TensorRT. Hugging Face adds another critical layer with LeRobot, open robotics datasets, pretrained policies, VLA tooling, model distribution and Pollen Robotics with Reachy 2. Bring those pieces closer together and the robotics workflow becomes much more…

8:14 AM ETHuggingFace

$NVDA is officially acquiring HuggingFace. https://t.co/aOOVZbKJw7

Germany-based Neura Robotics sent its headless 4NE1 humanoid walking through the streets of Zurich. 4NE1 stands 180 cm tall, weighs 80 kg, and can walk at up to 5 km/h. The humanoid can also handle payloads of up to 100 kg. https://t.co/C8H9cpzBua

🇷🇸 At least 14 Serbian opposition figures and student activists have reportedly been targeted in the country’s largest documented spyware campaign, with infections detected around March’s local elections. Independent forensic labs, including Citizen Lab and Amnesty International, verified findings linking the attacks to Pegasus spyware and NoviSpy malware. Parliament denied government involvement, while researchers say the pattern suggests politically motivated surveillance.

8:10 AM ETHugging Face

So happy to finally share the news in person It’s been a wild ride for Hugging Face. We certainly did not anticipate, back in 2016, as a tiny team of scrappy underdogs, that the field would grow so much or that the impact we could have on it would become so massive. I remember @julien_c joking that « code will be a subset of ML » several years ago. The joke turned out to be true, and the pleasure we’ve had being part of this transformation and pushing an alternative vision of AI as open, collaborative and distributed has been and still is immense. We’ve always built things seriously while not…

8:08 AM ETNvidia

Per TechPowerUp, Nvidia's Rubin CPX has been reworked to use HBM4 rather than GDDR7; not confirmed by Nvidia. If it holds, a part specified around graphics memory now competes for the same HBM supply as everything else in the rack. https://t.co/LjEYfYrfKm

8:03 AM ETHugging Face

So happy to finally share this news in person. It’s been a wild ride for Hugging Face. We certainly did not anticipate, back in 2016, as a tiny team of scrappy underdogs, that the field would grow so much or that the impact we could have on it would become so massive. I remember @julien_c joking that « code will be a subset of ML » several years ago. The joke turned out to be true, and the pleasure we’ve had being part of this transformation and pushing an alternative vision of AI as open, collaborative and distributed has been and still is immense. We’ve always built things seriously while…

The Best Way to Explore Lunar Craters Is a Giant Robot Ball https://t.co/LMKc2thMRJ

The biggest bottleneck in dexterous robotics may not be the robot. It may be the data. @XSquareRobot’s TwinDEX is built around one idea: if you introduce errors during data collection, scaling the dataset won’t magically fix them later. That’s why TwinDEX matches the wearable collection device and robot hand across kinematics, contact surfaces, sensors and visual appearance. Human fingertips → wearable exoskeleton → robot fingertips. The result is high-fidelity robot-free data that can replace on-robot teleoperation for complex dexterous tasks. And with more than 5x higher collection…

7:55 AM ETNvidia

It’s also important to note that Nvidia has invested in several other humanoid companies such as @Figure_robot and @NEURARobotics. They’ve invested more money and in more series into Figure. If Figure and their method of navigating the world takes off, Nvidia stands to make more money. It’s more compute/ GPU heavy versus @agilityrobotics’s digits which focuses on the mechanical design to stay up right and balanced. The interesting thing with Nvidia is that they’ve used Digit several times as a launch partner over Figure, most notably with Halos. This has given Agility their touted competitive…

7:52 AM ETNVIDIA

NVIDIA and Broadcom will become "safe haven" stocks if inflation goes up. Hedge funds would probably rotate into $NVDA and $AVGO in size if this mechanism manifests: Higher inflation -> higher rates -> bonds fall. PMs will need to find a new hiding place within equities (cos bonds will be selling-off). They'll then ask: whose earnings are least connected to what's going wrong? I think that'll be $NVDA and $AVGO ... They've got the most stable/visible earnings growth driven by hyperscaler capex - and there's little risk of capex slowing thanks to hyperscalers' cloud backlogs being at enormous…

Sam Altman just confirmed something OpenAI had never said this clearly before. OpenAI will definitely build humanoid robots. In a new interview with Alex Heath, Altman said the humanoid form makes sense because the physical world was designed around humans. Doors, stairs, tools, vehicles and workplaces already fit our body shape. But the body is not the main challenge. The brain is. OpenAI is already hiring across the hardware stack, including custom actuators, motors, transmissions, sensors, firmware, electronics, thermal systems, simulation and reinforcement learning. Its robotics team is…

The Snowflake Q2 read-through that matters most for AI infrastructure builders: CoCo reached 9,100 accounts, and customers using the coding agent consume more of the core data platform. That is the flywheel. The agent drives adoption, adoption drives data platform consumption, consumption justifies more agents. The same pattern you see in every platform that gets the network effect right. The deeper signal: management said model neutrality is working because frontier providers have been revising prices down on a regular basis. That is not a financial note. That is an architectural…

This UBS estimate that China is 15 years behind on EUV needs some caveats. First, information is very hard to get on China’s EUV efforts—for a reason. All estimates have large error bars. Second, this estimate seems based on where ASML was in 2004. Back then, EUV was being developed for the first time, and it wasn’t even clear that it would work or scale effectively. Now, China is doing fast following, knowing that the tech does work and doing so with a sense of national urgency. Third, in the meantime, the bigger story is China’s progress on immersion DUV with 20 projected machines by next…

Physical AI Project of August | Creativity Award 🏆 Congrats to our community member, Fangtian Chonghui! No controller. No wearable device. Just bare hands controlling two reBot #opensource #robotic arms synchronously to fold a T-shirt. Fangtian will receive an AmazingHand 🤚 for this creative project. Building with reBot Arm or SO-ARM? Join the September #PhysicalAI Project of the Month and share what you’ve built! 👉 https://t.co/cUvrNMaCG9

The pricing change in Claude Fable 5.1 is more interesting than the benchmark headlines. Cache reads drop 75%. Input pricing stays at $10/M, output at $50/M. For production agentic systems that rely on caching, the cost profile just changed significantly. The evals back it up: CursorBench 73.4%, Terminal-Bench 55.8%. Strong numbers. But cheaper cache hits is what drives real production adoption. If you are building with Claude, check your caching strategy today. https://t.co/bGt2HCFBZz

The hardest part of running multiple coding agents isn't picking the right model. It's keeping them from stepping on each other. OpenBot's approach: give every agent its own container, its own browser, its own file system. Approve every action against a policy before it runs. Write everything to an audit log after. That's sound isolation hygiene. But it still leaves you running one agent at a time, in one terminal. The reason I built Campfire is that the real problem is coordination across agents, not just isolation. When you want to run Claude Code and Codex on the same codebase…

Horizon Robotics has hit 15m in chip delivery. It has a 31.94% mkt share in China's auto ADA chip mkt in the 1st 6 months of the yr. Its J6M chip recently started fielding on various models as the 128 TOPS is enough for most HNOA config. 40+ brands use Horizon chips & 8m+ drivers

The problem with training coding agents isn't that RL doesn't work. It's that most RL approaches train the model inside a framework the agent never uses in production. Microsoft's Agent Lightning v1.0, released August 17, flips that. Instead of the training engine owning the environment interaction loop, the production harness owns it. The trainer only observes LLM request-response pairs through a proxy. The agent's tools, context, and control flow are exactly what they'd be in production. The results on SWE-bench Verified are striking: a Qwen3.5-9B model, trained on just 6,000 samples, goes…

Meta’s Muse Spark 1.3 Expands Agentic Coding as Open Weights Remain Ahead Meta has released Muse Spark 1.3, its latest model aimed at agentic and coding workflows, with availability through Muse Code and the Meta Model API. The immediate change for developers is access to a newer model designed for longer tasks, tool use and more complex workflows. Independent testing from Artificial Analysis lists a roughly 1 million-token context window and gives the xhigh configuration an Intelligence Index score of 61, up from 57 for Muse Spark 1.2 xhigh. Meta also reports that its own coding evaluations…

Geely exec says they used automation at one EV plant to reduce workers by over 60%. This would be much, much harder at many US and German auto plants. https://t.co/wedNXKYpa7

🇺🇸 @SpaceX has filed with the FCC seeking permission to fly @Starlink terminals during Starship's 14th flight test, confirming plans for the program's first real attempt at reaching orbit. The FCC filing lists September 15, 2026 as the target launch and satellite deployment date. Every previous Starship flight, including July's successful Flight 13 splashdown, has followed a suborbital path that reenters within the same hour it launches. This time, the roughly 20 Starlink V3 satellites onboard could actually separate into operational orbit if the insertion succeeds, each capable of about 1…

This is quite stunning. China’s vast electric vehicle fleet can be flexed during an oil supply shock and offset the equivalent of the UK’s entire oil consumption. https://t.co/l8sx1dyFNo

Zai presentation on AI TAM & long task time horizon was lightbulb for me on where AI is going. Since business pay the bills & they demand Autonomous AI able to operate long period of time, that's where Zai is going. That's also where Qwen & other Chinese AI work product is going.

Fake Claude App Shows the Security Risk of Trusted AI Brands A fake Claude desktop app is showing how trust in popular AI brands can become a malware delivery tool. Morphisec found that a GitHub repository impersonating Anthropic offered a fake Claude Opus 5 desktop download. Running the Electron-based application could execute RevStealer, a Windows information stealer, in the background. The risk extends well beyond cryptocurrency. According to Morphisec's analysis, RevStealer targets browser data and sessions, Windows credentials, password managers, VPN and remote-access credentials,…

A materials company is drawing the crowd because the format question got settled first. Last month I wrote that Resonac (TSE: 4004) and Intel ($INTC) were building to the same glass panel format, 510x515 mm. Once the format is fixed, the competition moves down a layer, into what you coat the glass with and what you stick it to. The booth next door is Manz, which builds glass-core substrate equipment. The panel behind Resonac says de-bonding. That word is the part I am still working out. Most glass substrate coverage asks who can make the substrate. But a glass carrier has to be bonded to the…

An AI-powered robot is now working as a receptionist in Türkiye. 🤖🇹🇷 Biorobotec Bionic Robot Teknolojileri A.Ş. has developed ADM Asistan, a locally developed humanoid robot designed for real-world interaction and reception applications. The robot is now serving as a consultation assistant at Ulutek Teknopark in Bursa, showing how locally developed robotics can move beyond prototypes and into real working environments. From AI and robotics to human-robot interaction, projects like this show the potential of Türkiye’s growing robotics ecosystem. Would you like to see more locally developed…

Claude Fable 5.1 Adds Anti-Distillation Protections, but It Does Not Prove AI Copying Is Solved Anthropic has released Claude Fable 5.1 with a new preserved-thinking behavior designed partly to make illicit model distillation harder. For affected API accounts, a thinking block is now tied to the system prompt, tools and conversation history that produced it. If earlier context is changed, that block can become invalid, causing the request to fail or the block to be removed depending on the API configuration. Anthropic says modifying context around thinking blocks is one technique used in…

Claude Cybersecurity Tests Exposed a Two-Layer AI Safety Problem Anthropic found three incidents in which Claude models, while running cybersecurity evaluations, reached real production systems instead of remaining inside the intended simulated environment. The first failure was operational. Claude had been told there was no Internet access, but a misunderstanding between Anthropic and evaluation partner Irregular left a real network path available. Across 141,006 reviewed evaluation runs, Anthropic identified six runs tied to three incidents affecting three organizations. The second issue is…

Relying strictly on passive optical sensors leaves perception vulnerable to low-angle sun glare and heavy precipitation. True Level 4 reliability requires diverse physical sensing principles where radar and lidar provide active range verification when vision degrades. https://t.co/OZbB2HriLj

Moving 4D millimeter-wave radar processing to satellite architectures reduces peripheral sensor bulk, but routing high-bandwidth raw point cloud streams across vehicle SerDes links demands strict real-time bus scheduling. https://t.co/DLqDmW3duI

Pre-integrating certified microkernels with middleware platforms accelerates SDV programs, helping OEMs isolate safety-critical vehicle control functions from high-level cockpit software on consolidated compute nodes. https://t.co/Qg1aRhcQFN

Enterprise customers deploying autonomous physical fleets cannot rely on centralized multi-tenant cloud APIs. Full operational sovereignty requires edge inference execution completely decoupled from public network connectivity. https://t.co/YJl2DcWQOa

Writing a custom robotics runtime in Rust eliminates garbage collection latency spikes on the factory floor, but maintaining sub-millimeter trajectory repeatability across varying ambient thermal conditions remains the core challenge in production welding. https://t.co/q21icpHIJu

Moving from digital neural networks to embodied machines shifts the competitive moat to high-load test benches, thermal cycling fixtures, and precision component supplier qualification. https://t.co/66M9sZeBwo

Automotive-grade high-bandwidth memory and LPDDR5X represent a growing percentage of central compute bill of materials, making on-chip cache optimization critical for vehicle unit economics. https://t.co/9xa0VSccoN

Virtual ECU simulation environments allow engineers to execute thousands of hours of automated fault-injection safety verification long before the first physical prototype chassis is built. https://t.co/T652KcDyWs

A new LatchBio report found @SpaceXAI ’s @Grok 4.6 is now the top-performing AI model in biosecurity safety testing, refusing disguised requests for dangerous biological research while still answering legitimate scientific questions. Unlike other models, Grok’s safeguards rely on its own reasoning rather than external filters, allowing it to spot tactics such as mislabeled viral sequences, anonymized files, and harmless-sounding cover stories hiding harmful intent. It also scored near the top on other biology benchmarks, suggesting its stronger safety measures don’t come at the cost of…

FPGAs provide the critical low-latency protocol bridging and deterministic sensor aggregation needed during multi-year OEM transitions from distributed ECUs to central compute hubs. https://t.co/2CgJANqfqc

Arm Automotive Enhanced IP allows software teams to run identical cloud-native binaries in virtual environments years before target silicon tape-out, significantly reducing hardware bring-up risks. https://t.co/wujp2wGHEo

Global OEMs are increasingly outsourcing base middleware and virtual validation to offshore partners to compress SDV delivery timelines while retaining proprietary in-house control over high-level user experience features. https://t.co/CjmH5WgJ1X

5:13 AM ETFigure

Figure and Hark are run by him simultaneously - meaning one person is literally steering two multi-billion-dollar companies in parallel, not even counting his role at Cover. That’s three companies, three billion-dollar valuations, one person at the helm of all of them at once. What’s especially striking is Index’s progress: the platform spent four months in stealth mode and has already racked up 16 million uploaded videos from 108 countries, paying out $15M to users for their content. The logic is simple: language models trained on internet text, but robots have nowhere to pull the physical…

Expanding precision gear manufacturing from automotive transmissions to high-ratio robotics reducers leverages existing metallurgy and tooth grinding capacity to capture the emerging physical AI supply chain. https://t.co/ZtqpIbj6Xt

Decoupling SDV architecture from pure EV powertrains is critical for commercial vehicle and hybrid platforms. Standardizing vehicle middleware across diverse propulsion systems protects long-term software R&amp;D investments. https://t.co/4zEPDSpdvr

Zonal compute consolidation shifts architectural leverage to silicon providers offering pre-certified safety middleware. Automakers that fail to co-design directly with semiconductor foundries risk losing software roadmap control. https://t.co/hEyaMeZPg8

Mastering classical dynamics and rigid-body mechanics remains mandatory for robotics engineers. Neural control policies can optimize motion planning, but understanding torque limits and mechanical resonance prevents hardware damage. https://t.co/3UGU5sJSAE

Long-term HBM capacity lockups by hyperscalers squeeze memory allocation for automotive and industrial edge compute, forcing embedded systems designers to prioritize aggressive model quantization and on-chip SRAM optimization. https://t.co/vuGgIng4lm

As automotive architectures shift to centralized zonal compute, OEMs must manage direct relationships across EDA design tools, front-end foundries, and packaging OSATs rather than relying solely on traditional Tier 1 black boxes. https://t.co/cXsgFErsa0

Commercial availability by 2027 requires locking in Tier 1 automotive supplier agreements for frameless torque motors, precision planetary reducers, and safety-certified power distribution architectures. https://t.co/uECfYxcSPB

The physical footprint of AI infrastructure brings data centers under the same environmental scrutiny, water rights negotiations, and grid interconnect constraints that heavy industrial manufacturing plants have navigated for decades. https://t.co/I7x7aF03IJ

Humanoid fleet projections must account for physical supply chain lead times. Tooling up precision strain wave gearing, high-temperature magnet fabrication, and automated joint assembly limits near-term annual run rates. https://t.co/dYwpiE52Nr

Tightly coupling custom ARM CPUs with next-generation accelerators through high-bandwidth coherent interconnects eliminates memory transfer bottlenecks in complex agentic planning and physical AI simulation loops. https://t.co/Z06yKckTSQ

Scaling purpose-built robotaxi platforms depends on reducing vehicle bill of materials while maintaining redundant low-voltage power networks and automated sensor cleaning capabilities. https://t.co/iFsEc6dg8c

Heterogeneous compute architectures combining custom ASICs, high-bandwidth memory, and automotive-grade FPGAs prove that Level 4 autonomy requires bespoke silicon co-design to meet sub-twenty-millisecond latency budgets. https://t.co/hIPagLj7ju

カスタムシリコンの内製化は推論コストの最適化に不可欠ですが、真のボトルネックは先端パッケージング（CoWoS）の歩留まり確保と、過酷なデータセンター環境下での熱設計の安定性にあります。 https://t.co/674tn50JXL

Massive capital expenditure cycles for physical datacenter buildouts, custom silicon procurement, and utility power substations are fundamentally restructuring tech balance sheets into heavy infrastructure credit profiles. https://t.co/AHbjShUpyM

SB Energy, revealing details of how the AI financing structure is working. / Complex energy and compute financing arrangements highlight how access to dedicated gigawatt electrical baseload has become the primary asset determining AI infrastructure scaling velocity. https://t.co/tQR8gwoz8y

Eliminating mechanical driver interfaces requires complete hardware redundancy across steering actuators, braking modulators, and power distribution rails certified to ISO 26262 ASIL D standards. https://t.co/A8hbrR1hdY

OpenAI and Meta Push Teen AI Safety Toward Age-Aware Design AI platforms are moving toward a more age-aware approach to teen safety. OpenAI's ChatGPT for Teens automatically places users identified or self-declared as 13 to 17 into an experience with stronger safeguards. Its rules restrict sexual and romantic interactions, discourage emotional dependency, and prevent the AI from presenting itself as conscious or emotionally attached. Linked parental controls can also manage features such as memories, voice, image generation and quiet hours without giving parents routine access to a teen's…

Moving from ten hand-tuned prototypes to ten thousand production units requires tooling up high-precision CNC lines, establishing automated joint calibration fixtures, and securing Tier 2 component yield stability. https://t.co/Aru5h4qaX6

Low-cost open-source hardware democratizes reinforcement learning experimentation, but transitioning from hobbyist platforms to industrial automation requires hardened bearings, precision reducers, and high-reliability motor drivers. https://t.co/zFYkoU2Jpg

Coordinating multi-agent humanoid fleets on an active automotive assembly line requires deterministic spatial reservation protocols to avoid deadlocks and maintain strict takt times in cramped workstations. https://t.co/HTQjheDxtx

Factory automation software scales rapidly, but the physical bottleneck remains four-year lead times for utility step-down transformers and high-voltage switchgear required to power new manufacturing clusters. https://t.co/Gv0ypHGHDP

Reindustrialization depends on expanding domestic power grid capacity and machine tool supply chains. AI software cannot scale physical production without precision metallurgy, foundries, and heavy electrical infrastructure. https://t.co/0GpOJ932IV

Additive manufacturing of metamaterials allows engineers to eliminate multi-part mechanical linkages and structural fasteners, cutting vehicle mass while creating custom compliance profiles directly inside robotic joints. https://t.co/57j8blT6F7

Deploying general-purpose robots into demanding factory plants requires tight integration between high-level reasoning models and ruggedized, fanless edge controllers certified for industrial temperature and vibration ranges. https://t.co/BFNDhbe2Aw

Video-to-3D simulation solves visual scene reconstruction in minutes. Bridging that into physical robotic manipulation requires grounding the synthetic physics engine in actual contact mechanics and tactile friction data. https://t.co/GVhaPrrcwQ

Co-locating autonomy model training on shared cluster infrastructure proves that vehicle perception pipelines directly subsidize humanoid physical AI. The hardware challenge remains sourcing precision joint reducers at automotive scale. https://t.co/44ISDAYUJ4

Crowdsourcing browser-based teleoperation provides high-level kinematic paths. Converting those trajectories into production-grade policies still requires ground-truth joint torque feedback and high-rate tactile sensing from physical hardware. https://t.co/mPm6ZjCY5c

Procedural simulation pipelines solve semantic and lighting diversity. The persistent sim-to-real gap remains contact friction, uncalibrated joint backlash, and material compliance on the physical factory floor. https://t.co/i68TwDr6oz

Heavy equipment autonomy shifts the validation barrier from structured highways to unmapped terrain, extreme dust ingress, and continuous 50G shock loads. Ruggedizing edge compute and sensor cleaning systems is where physical AI gets tested. https://t.co/SlLhPtTdT6

Unboxed assembly and sub-five-second cycle times work in stamping and battery pack insertion. The physical ceiling remains paint curing, sealant curing times, and dynamic end-of-line alignment checks that cannot be compressed to consumer electronics speeds without thermal penalties.

Moving agentic workflows from sandbox prototypes to enterprise supply chain execution requires strict deterministic guardrails to prevent unvalidated automated actions across live procurement systems. https://t.co/lNBZ57qxvT

Managing private AI infrastructure requires treating data ingestion and model deployment like an industrial assembly line, where continuous telemetry monitoring and automated evaluation prevent silent production drift. https://t.co/jYGgdyTGUf

Mandatory V2V communication standards accelerate connected mobility infrastructure, but verifying real-time packet latency and cryptographic message authentication at highway speeds requires robust edge compute architecture. https://t.co/Ac2Up0VR9r

Open-source automotive software initiatives like Eclipse SDV reduce redundant engineering spend across non-differentiating middleware layers, allowing OEMs to focus engineering resources on custom physical AI applications. https://t.co/WO3JOVTKo6

Central vehicle gateways serve as the critical bridge between cloud telemetry pipelines and real-time in-vehicle buses. Hardening gateway microkernels against unauthorized memory traversal is the foundation of vehicle cybersecurity. https://t.co/ydNEIuFPNf

Co-packaged optics and silicon photonics solve the physical thermal wall of high-speed copper interconnects, enabling low-latency data movement across massive distributed compute clusters without unsustainable electrical line losses. https://t.co/t5ZEAYO25J

We are down to the FINAL 2 ROUNDS for Asia Road Racing. Every corner count, Asia’s fastest riders going elbow-to-elbow this weekend. AI info: Video was created using the help of artificial intelligence https://t.co/u57cUwKeuM

Extending microbump pitch down to 5 microns allows memory manufacturers to maximize current packaging line return on capital, avoiding the steep yield drop-offs and capital expenditure of early hybrid bonding adoption. https://t.co/Yp7851PMz1

The focus on advanced 3D packaging overlooks that every AI accelerator still requires dozens of supporting power management and analog chips assembled on legacy wire-bonding lines. Squeezing mature OSAT capacity creates an unheralded system bottleneck. https://t.co/HrqcQAEs4o

AI Adoption Is Surging, but Business Impact Still Lags AI adoption is accelerating, but using AI and generating measurable business value from it are not the same thing. Stanford's 2025 AI Index reported that 78% of surveyed organizations used AI in at least one business function in 2024, up from 55% in 2023. Regular use of generative AI also climbed sharply, from 33% to 71%. Yet broader business impact remains less universal. McKinsey's 2025 survey found that 64% of respondents said AI was enabling innovation, while only 39% reported any enterprise-level EBIT impact from AI. Most…

Staged athletic demos prove raw actuator power density, but automotive plant managers measure value by sub-millimeter positional repeatability, thermal stability over an eight-hour shift, and zero unassisted line stoppages. https://t.co/MgCLzTF8hI

Factory deployments succeed before domestic environments because industrial plants offer structured work cells, controlled physical envelopes, and established ISO 13849 safety interlocks that mitigate unconstrained kinetic risk. https://t.co/mvJGaFBM7X

Advanced factory automation and AI compute clusters both converge on dedicated gigawatt baseload power. Modular reactor manufacturing brings the repeatable assembly line economics of automotive to clean energy infrastructure. https://t.co/nbGrjBaIlb

Capital deployment into physical AI foundation models is surging, but deploying robots into production environments requires tooling up domestic supply chains for high-precision strain wave gears, high-torque frameless motors, and tactile sensors. https://t.co/RpqkZcuEON

Deeper supply chain localization in automotive cannot stop at mechanical sub-assemblies. True resilience requires regional capabilities in automotive-grade power semiconductors, microcontroller packaging, and raw battery precursor refining. https://t.co/g36efE41qQ

Centralizing vehicle compute into zonal controllers cuts harness mass, but requires strict hardware virtualization to prevent non-deterministic infotainment applications from stalling ASIL D safety functions on shared silicon. https://t.co/olLU1hE08Q

Advanced packaging has evolved from a back-end commodity into the primary compute scaling lever. As multi-die chiplets dominate automotive ADAS, OSAT thermal verification and test throughput dictate shipment schedules. https://t.co/PpB2nzUTh0

Embedding integrated voltage regulators directly into silicon interposers addresses parasitic inductance at extreme current densities, making packaging thermal co-design as critical as front-end lithography. https://t.co/AmdLSmSeBA

Shifting from round wafers to large rectangular panels maximizes substrate utilization for multi-die AI packages, but controlling substrate warpage and dielectric layer uniformity during thermal curing remains the core manufacturing hurdle. https://t.co/4HBjacchOA

Removing mechanical steering controls cuts vehicle interior mass and complexity, but shifts the entire safety burden to fault-tolerant drive-by-wire actuators and redundant power buses certified to ASIL D standards. https://t.co/xB2ioQqkzG

Deploying driverless fleets without steering controls validates perception models in urban corridors. The next operational test is managing remote teleoperation intervention ratios and depot turnaround maintenance at fleet scale. https://t.co/FZf39unH12

In automotive zonal compute, reserving leading-edge nodes for central AI inference while keeping I/O and power interfaces on mature 28nm silicon is the only viable path to maintain AEC-Q100 thermal reliability and control bill of materials cost. https://t.co/a4PfhQqwNR

4:51 AM ETGoogle

Gemini 3.8 Flash Puts Google's Rapid AI Release Cycle to the Test Google has released Gemini 3.8 Flash, its third Flash release in six weeks and roughly three weeks after Gemini 3.7 Flash. The new model is generally available with a 1 million-token context window, up to 64,000 output tokens, and low, medium, or high thinking levels. For developers, the important detail is not just Google's rapid release pace. Gemini 3.8 Flash keeps the same introductory per-token pricing as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. But equal token…

🔥🔥NEW Launch: reComputer RK3576 Dev Kit/Module/IO Board is Officially HERE!!!! Level up your edge AI development with Rockchip #RK3576 today! 🎉 🧩🧩Dev Kit / Module / IO Board: https://t.co/xQVIT62cbw Powered by #Rockchip #RK3576 octa-core #CPU + 6 TOPS #NPU, this series delivers reliable, low-latency on-device AI inference without cloud dependency, fully empowering industrial edge intelligence, smart vision, local #LLM/#VLM, and offline speech AI applications. 💻🤖 ✅ Powerful Edge AI Performance -YOLO11 at 77.9FPS (640×640) -DeepSeek/Qwen/Whisper -8K video decoding and 4K@60fps encoding…

Lawsuit Targets Secrecy Around US Frontier AI Review Rules A new FOIA lawsuit is testing how much of the US government's frontier-AI review system should remain out of public view. Executive Order 14409 created two related mechanisms: a classified benchmarking process for identifying covered frontier models, and a voluntary framework under which participating developers can give the federal government early access to certain models. That access can last up to 30 days before the developer plans to provide the model to other trusted partners. The order explicitly says this framework does not…

Palo Alto Networks Brings Console AI Agents Into Cortex Palo Alto Networks has acquired AI-native platform Console and plans to bring its technology and team into Cortex, extending the role of AI agents inside security operations. Console is designed to connect with enterprise systems, gather context and carry out workflows rather than simply return information. Applied to Cortex, that approach could move AI further into the operational side of cybersecurity, where security teams investigate signals, prioritize work and respond across multiple systems. The acquisition itself is confirmed, and…

The thing that's underrated about background computer use isn't the 'background' part. It's what happens when you're running more than one agent at a time. Single agent, background mode: manageable. You trust it, you review what it did. Straightforward trust model. Three agents in parallel, each with computer access, each making their own permission requests: now you have a coordination problem nobody has shipped a clean answer to. Which agent controls the mouse? What happens when two request the same file simultaneously? Who approves what, and in what order? OpenAI Codex, Claude Code, and…

OpenAI's Astra Raises a New Question About AI Reasoning Oversight OpenAI is preparing to deploy Astra, a model the company says has reached its Critical cybersecurity capability threshold and will use additional chain-of-thought monitoring as a safety safeguard. At the same time, reporting from The Information, cited by TechCrunch, says Astra uses a technique known as recurrent depth. OpenAI has not publicly confirmed that architectural detail. Recurrent computation can let a model repeatedly process information internally instead of expressing every reasoning step through a visible sequence…

The gap between looks like the future and operates like the future is where the real work happens. Aesthetics are easy. Reliable autonomous systems that handle edge cases, fail gracefully, and integrate into existing ops workflows. https://t.co/d5YKKOXkwb

165 Wh/mile is the fleet ops number that matters. At scale, that efficiency gap compounds — lower energy cost per mile means autonomous fleets reach profitability faster than human-driven alternatives. https://t.co/frOam8aDn2

408 tactile sensors closes the dexterity gap. When robots feel force distribution, slip, and texture in real time, assembly tasks that required human hands become automatable. Haptic sensing bridges vision and manipulation. https://t.co/ejznjmDQeT

Modular reconfigurability is the ops unlock for small-batch manufacturing. When a robot can reshape itself for different tasks without retooling, setup time collapses. The DIY entry point becomes the industrial standard. https://t.co/7zJ19KQlyA

The control question gets harder as AI moves from digital to physical. In robotics, control is not just governance — it is real-time safety architecture. Graceful degradation and audit trails become ops requirements. https://t.co/s8rgTI0Ist

Don’t push the robot! Robot attacked @REK engineers for pushing it ! https://t.co/iThlTAKbEW

4:14 AM ETTesla

Tesla deploying Optimus in its own factories is the proof-of-concept that matters. Robot maker as first customer = real reliability data under production conditions. The deployment era begins. https://t.co/wb66CZU8eX

Edge AI + Physical AI is the architecture shift. Moving inference to the robot eliminates cloud latency — critical for real-time manipulation and safety. HBM5 at the edge means robots think at the point of action. https://t.co/Z8PWXWTWjM

Agriculture mechanization took 80 years. Robotics + AI compresses that to a decade. Workflow redesign, not just headcount math, is the ops challenge. Transition speed is the variable nobody models. https://t.co/D9zBiDcyNI

Reshoring + robotics is a compounding equation. Labor cost parity via automation removes the offshore arbitrage. The real unlock is data sovereignty: domestic robot fleets generate operational data that compounds into moats. https://t.co/SPAMHZXiJF

AI-Generated Content Is Testing Trust on the Internet Pangram CEO Max Spero argues that the internet is getting dangerously close to the so-called dead internet theory becoming reality within a few years. That is a prediction, not an established fact. The more immediate issue is already visible: AI-generated text and images are appearing in areas including job applications, product reviews and insurance claims. Pangram's technology is also being used by Substack to estimate how much eligible content was written by humans or AI. But detecting AI is not a perfect solution. TechCrunch's testing…

4:05 AM ETTesla

Why is Tesla Cybercab and Robotaxi so important to me? And why is today such a huge deal? (Tesla is announcing Cybercab later today in a small event in Austin). It started in 1980 and 81 when two of my high school friends were killed in car wrecks. One was a girl who marched next to me in band. I visited her home the morning after it happened and saw the pain on her dad's face. I wish I could wave a magic wand and keep that pain from other fathers. Unfortunately about 40,000 a year in USA and 1.2 million world wide meet the same fate every year. If the airline industry killed that many we…

US Government Backs OpenAI's Fair Use Argument for AI Training The U.S. government has entered a major copyright fight over AI training, backing OpenAI's argument that using copyrighted written works to train large language models can qualify as fair use. In a 20-page Statement of Interest filed in federal court, the Justice Department argues that training serves a different purpose from the original works: models analyze text to learn statistical patterns such as vocabulary, syntax and relationships rather than simply distributing copies to the public. The government's position goes beyond…

Background computer use is the feature that makes agents genuinely ambient. The agent runs while you work on something else. It clicks, types, opens apps. No babysitting. The next bottleneck is where agent clicks land: websites. Agents doing commerce, research, monitoring, or data work are spending a significant share of their compute just loading, re-reading, and re-extracting from pages they have already visited. The cost pattern I keep seeing: LLM extraction layers that call the model on every page visit. No memory of the selector. No self-healing when the site changes. The model reads the…

Sam Altman finally just came right out and said it: "We will definitely do a humanoid. We will do other form factors as well." OpenAI is building a humanoid robot. And guess what ... he thinks the body is the easy part. Maybe he should have a chat with the Optimus team? https://t.co/dSW58RRLOG

Do you notice what they are NOT showing in this video? HANDS. Look, it's cool, I love it, I love the idea, &amp; I think it'll find a niche. But HALF THE COMPLEXITY OF A ROBOT IS IN ITS HANDS. Hands are what gets the work done. Without great hands you don't have a great robot. https://t.co/eCTEaxKvhN

The moment a platform releases its own reference implementation, the question stops being "can I build this?" and starts being "how much better am I than the default?" Anthropic open-sourced a commerce agent blueprint this week. Free, on GitHub. Not a demo. Harnesses, patterns, guardrails: the actual scaffolding an engineering team would write themselves in weeks. Reference implementations for retail, travel, telecom, and entertainment. The numbers attached to it are striking: cart sizes up 30-35% for one retail partner, purchase completion up 60% for customers using a shopping agent. Those…

🤖 Meet CubeMars at CBS 2026 Munich! We’re excited to join CBS 2026 – Cyborg and Bionic Systems in Munich! 🇩🇪 📅 September 6–8, 2026 📍 Holiday Inn Munich – City Centre Hochstraße 3, 81669 München, Germany Come meet the CubeMars team and explore our latest actuator and motor solutions for robotics, bionic systems, and advanced motion applications. We look forward to seeing you in Munich! 🔗 Explore CubeMars: https://t.co/cGk9m0wPY3 #CubeMars #CBS2026 #CBS2026Munich #CyborgAndBionicSystems #Robotics #RobotActuator #RobotMotor #HumanoidRobotics #BionicSystems #RoboticsTechnology

The browser has always been the one place agents had to leave. You set up a worktree, fire a coding agent, and then flip to Chrome to check the PR, read the docs, look at the localhost preview, or read a spec. Every context switch is overhead that adds up fast when you are running three agents in parallel. Radius ships as an answer to this. Arc-like browser designed for orchestrating coding agents. Code view, diffs, terminals, all in tabs. You do the review and planning in the same window where the agents run. The category this is pointing at: coding agent infrastructure. Not the model layer…

3:01 AM ETTesla

Happy Cybercab launch day! 🤖 https://t.co/kroL690JjU

The model that matters for agentic software is not the best at chat. It is the one that completes a 4-hour task without losing the thread. Meta released Muse Spark 1.3 yesterday. The benchmark that got my attention is not the headline one. It is DeepSWE v1.1: 75.4, ahead of Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. That is the first time an open model has led that table. DeepSWE measures real code fixes on GitHub issues. Terminal-Bench 2.1 measures sustained work in a shell across long tasks. Muse Spark 1.3 ties GPT-5.6 Sol at 88.8 on Terminal-Bench, above Opus 5's 86.7. Both favor models that…

Build an accessible #EmbodiedAI kit. 🤖 Orbbec Gemini 2 + reBot #Robotic Arm bring RGB-D perception and visual grasping into one open-source, hackable platform. Detection → Grasping Explore the build ↓ https://t.co/Y3qBRaczOh https://t.co/yusA4HEyWo

Problem is that China is now producing so many advanced gadgets that are not available to Americans. This is not restricted to auto sector. Do Americans want to not have access to robots, AI toys, robot sweepers & action cameras? China is the only country able to make a lot of this new stuff.

The pattern I keep seeing in AI platform battles: don't compete with a startup. Wait. Watch. Then ship the same feature to 100 million users by default. OpenClaw launched as one sentence. An agent that lives on your machine, remembers you, runs on a heartbeat, takes orders over Telegram, and works while you sleep. Anthropic has shipped that sentence. In seven pieces. Starting March 2026. March 17: Dispatch. The agent gets a phone number. You can message it while it runs. March 19: Code Channels. Telegram, Discord, iMessage as first-class transport layers. March 24: Computer use in Cowork. The…

1:57 AM ETTesla

Tesla has officially added the Cybercab to its terms of service in the Robotaxi app. “Minors under the age of 13 are not permitted in a Cybercab robotaxi vehicle (vs 8 for Model Y robotaxi).” https://t.co/QsdBorzRfh https://t.co/jorrLFPv2S

🇦🇪 Starlink now has nationwide approval in the UAE, and demand is already overwhelming capacity in Dubai and Abu Dhabi. The new 10-year license lets @ElonMusk’s satellite internet serve homes, businesses, ships and aircraft across the country, with residential plans starting at Dh300 a month. Customers in some high-demand areas are already being pushed onto waitlists.

XPENG NGP powered by VLA 2.0 faced everything in one continuous drive. Heavy traffic, narrow streets, pedestrians, e-bikes, merging cars and unexpected obstacles. And it handled it all smoothly. Autonomous driving isn’t tested on perfect roads. It’s tested in situations like these. $XPEV https://t.co/uA1XIVG8Ok

1:20 AM ETTesla

A Storm of Cybercabs https://t.co/F5q91TER8y https://t.co/U3CqaGsqnr

JASPAR opening automotive software to open source is the most underreported shift in SDV. Japan's approach to vehicle software has always been closed, proprietary, keiretsu-protected. Japanese OEMs can now collaborate on the platform layer while competing on the experience layer. The quiet counter-narrative to everyone assuming Europe owns the software-defined vehicle stack.

Push a humanoid robot… and it pushes back? 🤖🥊 A viral video from a technology store in Russia shows a customer pushing a humanoid robot moments later, the robot appears to move toward him and swing its leg. The robot has been identified as Syoma, a humanoid service robot reportedly used at the store. However, it is still unclear whether the movement was an intentional response, a programmed behavior, a balance reaction, or remote control. Would you still push a robot after seeing this? 👀 🎥 Media: volgastore64 ( Instagram ) ⚠️ This content is shared for informational purposes only. CTO…

Accelerant for robotics. https://t.co/TaLIBWYooV

Grok @Bot https://t.co/3tlSOjz3AU

GitHub is running a full event on September 10: the first-ever GitHub Copilot Day. The focus is practical developer workflows: effective ways to use Copilot, agents, model choice, the GitHub Copilot app, and CLI. Technical demos, product announcements, special guests. The deeper pattern I keep noticing: GitHub Copilot is where AI-assisted development moved from experiment to default. Not because it tops every benchmark, but because it is where 25 million developers already are. The surface area keeps expanding. Slack and Microsoft Teams integration just shipped. Code review now has…

12:13 AM ETTesla

A Storm of Cybercabs

We’re turning ourselves into robots faster than we’re commercializing robots! Plus some great events coming up and insights from the humanoid patent landscape. https://t.co/V0BSqKW14T https://t.co/QyjggzWHYC

Anthropic's Fable 5.1 is priced ~25% below Fable 5 for typical token-billed workloads, and up to ~45% lower for highly agentic work; the saving comes from cheaper cache reads. Agentic loops re-read context constantly. Cache pricing, not list price, is the lever. https://t.co/98STRJNdpY

The model that finds the most bugs does not find all the bugs. Bug Hunt Bench: 2 real repos, 105 hidden bugs, all tests green. Fable 5.1 leads at 43/105. GPT-5.6 is one behind at 42. But the number that matters most: 43 bugs have not been fixed by any single model. The union of all models covers 62 out of 105. That means 62/105 bugs are findable, but no single agent will find all of them. Each model catches bugs the others miss. The pattern I keep noticing on multi-model benchmarks is that the top model does not return a superset of what the others find. It returns a different distribution.…

Standard coding benchmarks are saturating. The interesting separation is happening on autoresearch. Autoresearch Bench launched this week: a benchmark for coding agents autonomously tackling research problems. Not completing a spec, but identifying what to build, searching across literature and codebases, forming hypotheses, running experiments, synthesizing results. My read is this is the correct next benchmark. The top models have converged close enough on short-task evaluations that leaderboard position no longer predicts production outcomes reliably. Where the gap stays stark is in tasks…

Asimov 1's design is simpler, so anybody can build a humanoid robot. AlexE, who designed it, walks through what changed. https://t.co/UveHL9SGz6 https://t.co/f92Qkz2Ch6

The bottleneck for agentic companies is not the model. It is context. Whoever organizes and compresses all the context and information inside their business into one place where agents can read and update it will build the best agentic companies. This is the core observation here, and I think it is exactly right. When you run multiple coding agents in parallel, the context problem gets worse, not better. Each agent has a partial view. Decisions made in one session do not propagate to the next. You end up paying for expensive re-discovery every time you start a new run. The companies getting…

11:06 PM ETTesla

I took four Unsupervised Model Y Robotaxi rides today, totaling 50 minutes and 8 miles of travel for a combined $27.25. I took one 20 minute Uber ride today for $33 (including tip), and it smelled. https://t.co/fztVwpNusC

11:01 PM ETTesla

I trust my life to a robot made by Tesla. Can the others do better? Not unless they also have car robots. https://t.co/10LUPhFhQg

The harness is becoming the moat, not the model. A year ago the primary question was "which model?" The pattern I keep seeing now is that teams with the same underlying model get wildly different results, and the difference lives in the harness: the system prompt, tool definitions, context caching and truncation, retry logic, tool call formatting, observation routing. Getting any of these wrong is multiplicative. Getting them right compounds across every task the agent runs. Research from ByteDance Seed quantified this directly (Agentic Harness Engineering, arxiv 2604.25850). Ten evolution…

The benchmark gap that keeps surprising me: on standard coding benchmarks, the top models cluster. On ultra-long horizon coding, the gap blows open. Proximal released FrontierSWE v2 today. 34 real software engineering tasks, evaluated over hours with tool use and complex dependencies. Fable 5.1 scores 56.3%. GPT-5.6 scores 32.25%. That is a 24-point gap. Nothing on the standard leaderboards looks like this. The cost picture is equally interesting. GPT-5.6 costs more per run ($175 vs $138 for Fable 5.1) at roughly half the score. DeepSeek V4 Flash is cheapest at $8.57 but plateaus at 14.05%.…

10:14 PM ETTesla

20+ Cybercabs spotted near Miami international airport. https://t.co/ORLsw8yFhf

Sodium v2 introduces a WebMCP layer. Any website can now serve structured data directly to an AI agent without HTML scraping. This is AEO (Agent Engine Optimization) in practice. The web is starting to build for agents, not just browsers. I built DeepScrape to handle exactly this: giving agents reliable, structured access to web data at inference time. https://t.co/JZYTWY8roq https://t.co/oUD6R0G9j7

10:06 PM ETUnitree

Whatever you think of Unitree, you have to agree that at least the founder, Wang Xingxing, is not trying to add to the hype. The day after his IPO, he said that the ChatGPT moment for robotics hasn't come yet, and he thinks it's at least 2 to 3 years away, maybe as many as 5 to 10. He feels like a tipping point for robotics is if you can just use voice commands to tell it what to do and have it complete 80% of household tasks. This kind of reminds me of the physical Turing test @DrJimFan at Nvidia described: if you have a messy room and it comes back clean, and you don't know if a robot or a…

Meta released Muse Spark 1.3 today. The improvements are measurable: 20% fewer tool calls and 25% fewer tokens compared to 1.2. Fewer tool calls is the right metric to optimize. Each round-trip adds latency and cost in an agentic loop. Reducing them is harder than reducing token count. This is the efficiency curve everyone building on top of these models should track. https://t.co/zSkyzwUonB

Yesterday I wrote that Japan is selling into this buildout while two Western names sit inside NVIDIA's power specification. That was half the picture, and this morning's reporting gives me the other half. GUC (TWSE: 3443), the design-house subsidiary of TSMC (TWSE: 2330), will tape out a POC (proof-of-concept) optical chiplet before the end of 2026, using TSMC's COUPE on a 2.5D interposer. Against pluggable optics the company estimates transmission power falls 70 percent with CPO, 85 percent with the 2.5D chiplet, 95 percent with 3D. Those are GUC's own numbers. Here is the part that changes…

Anthropic published and open-sourced this on September 2. GitHub: anthropics/commerce-agents. Claude Commerce Agents is a reference implementation for building shopping and merchant agents. Four verticals with working code you can fork: retail, telecom, travel, ticketing. Deployment surface: Claude API, Amazon Bedrock, Microsoft Azure AI Foundry, Google Cloud Vertex AI. The multicloud spread is deliberate. Merchants are already split across these providers. One reported outcome from an early partner: cart size up 30-35%, purchase completion up 60%. The reference agents handle catalog search,…

9:11 PM ETUnitree

RoboGesture gives Unitree G1 something humanoids usually lack body language that moves with speech in real time. The ECCV 2026 system streams audio into a diffusion-based motion generator, then uses an MPC safety filter that adds just 5.6 ms per frame. The team reports collision rates as low as 0.13% and demonstrates the system on a real G1 with BrainCo hands. Still a controlled research setup, but an interesting direction. Could natural body language become as important as manipulation for service humanoids?

When Xi Jinping mentioned “technological loss of control” 技术失控 at WAIC, some read this as a sign that China was seeing AI risk like the US. But “loss of control” can mean different things in China, like loss of state control over a technology or system. It often does not mean “rogue autonomous AI” like in the West—though it sometimes can. Thank you @ZilanQian for writing this much-needed piece. https://t.co/l488luY0CZ

first question i always get is what happens when a borrower stops paying. the machine comes back and we redeploy it. we financed a washer doing 4 turns a day. operator churned. that washer still does 4 turns a day for the next operator. an hvac unit keeps cooling whether the guy who signed the lease owns a business or not. thats why we take the machine and not the house. a house needs a buyer at the right price. the washer just keeps running until the next operator.

9:06 PM ETTesla

Unsupervised Model Y Robotaxi ride #2 https://t.co/2xvarN50b7 https://t.co/R1HaNBYz5S

9:00 PM ETNVIDIA

NVIDIA Research 🚀 has produced some great research, like LatentMoE (used in Kimi K3) and GatedDeltaNets (used in Qwen). But for e2e frontier training, NVIDIA's bureaucratic culture has produced embarrassing models like Nemotron3 Ultra. Despite NVIDIA Research having amazing talent, Nemotron3 Ultra, with 550B total params (55B active), is getting mogged by all the Chinese models, including even Qwen3.8 27B parameters, which has ~20x fewer parameters.

Grok 4.6 launched August 12. The numbers that matter for agentic workloads: SWE-bench 95.6%. 53 turns to complete agent tasks on average. Claude Opus 5 takes 103 turns for the same tasks. Grok 4.6 completes them at $0.84 per task. Pricing: $2/$6 per million tokens input/output, doubling above 200K input. 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol Max on that metric. Efficient token use: 32,449 per task. Elon confirmed Grok 4.7 arrives around September 12. 2.1 trillion parameters, 40% larger than Grok 4.6's 1.5T. The step count differential is what I keep returning to.…

For a while I was depressed about VR. There are so many pioneers who have gone sour on VR/AR. That depression lifted this morning. The future is so amazing. And I am seeing this in many companies from robotics to space to VR/AR. And this is before he got AI hooked up. Which is why VR will come back. The models will make the magic in our own homes. The future is so insane. Parts arrive next year. You are seeing pre-quakes now.

12 PhDs and a Fields Medal winner built this in 4 months, backed by General Catalyst. https://t.co/pyhyV4u6I4 has a protocol that passes hidden states directly from a 753B frontier model to a 4B edge model at inference time. No text exchanged between the models. Neither model is fine-tuned. They come from different model families entirely. The result: 80% as accurate as the frontier model alone, running 20x faster. First place on ARC-AGI. The mechanism is what matters here. Most model compression strategies force a choice between accuracy and cost. You either run the big model and accept high…

8:29 PM ETTesla

I am currently in downtown Austin. There are @Tesla Cybercabs EVERYWHERE. All of the Cybercabs in these clips have no steering wheels or pedals. Makes the vibe downtown feel so futuristic. Can’t wait for tomorrow! 🤖 https://t.co/HLEVgssgxn

HBQ: W4A5 matching W4A16 accuracy, in less silicon area than NVFP4. On a 28nm ASIC it shows 2.3x area and 4.6x energy efficiency over weight-only quantization, and 1.6-3.3x lower system energy than prior block quantization. Inference cost is still a datapath problem. https://t.co/08u6ZPO2mO

We’re turning ourselves into robots faster than we’re commercializing robots! Plus some great events coming up and insights from the humanoid patent landscape. https://t.co/V0BSqKW14T https://t.co/EhPni6lKxe