TypeSafe

The bottleneck in most agent stacks right now isn't intelligence. It's the cost of the thousand small decisions an agent makes before it writes the sentence a user reads: classify this ticket, check this tool call, score this signup as…

The decision layer inside AI agents just became its own product category, separate from the model doing the talking. TypeSafe AI shipped Jev, a model that never generates text. Hand it a piece of state, a ticket, a log line, a tool list,…

The shift I keep coming back to this week: not every AI call needs to write a sentence. Diogo Almeida's Jev proved that in nine days. Jev is what TypeSafe calls a System One model. It skips token by token generation entirely. Feed it a…

The bottleneck in agent costs just moved, and I don't think most people are measuring the part that actually got expensive. For a year, agent economics meant output tokens: what a model charges to generate text. TypeSafe just showed the…

Jev doesn't generate text. It decides. TypeSafe AI came out of stealth Sep 15. System One is a classifier. Returns choice, score, or null. Not an LLM. Something fundamentally different.

System One AI models just entered production. Here's what that means for builders: 1/5 - A new model category Jev from TypeSafe does one thing: it classifies and decides. It takes structured input and returns typed output with…

What changes when a model returns typed decisions instead of text? Everything about cost, speed, and reliability. TypeSafe just shipped Jev -- a System One Model for decision-making.

The category is crystallising: System One models are purpose-built for high-volume, low-latency decisions—routing, scoring, classification. Not text generation. Jev by TypeSafe AI: typed probabilistic decisions, 40-200x faster than…

We're not just getting faster frontier models. A completely different model category shipped this week. TypeSafe AI released Jev on Sep 15. Built by Diogo Almeida, one of the people behind ChatGPT's reinforcement learning from human…

The Berkeley HarnessTax study deserves more attention than it got. 21 model-harness pairs tested across SWE-bench Lite and Terminal-Bench 2.0. The result: the same model running in different harnesses produced outcomes ranging from 38% to…

The bottleneck moved from generating text to making decisions. Jev from TypeSafe AI launched this week out of stealth. It is not a chat model. It is a System One model: it accepts any unstructured context and returns a typed, probabilistic…

The routing problem in a multi-harness agentic stack has been an LLM call problem. You route to Claude Code, Codex, Opencode, or another harness by putting logic in a system prompt and burning tokens on every decision. Jev changes that…

My read: When software needs to route between agents, classify intent, or make a binary API decision, it doesn't need a paragraph. It needs a number. That's what TypeSafe AI built Jev for. System One models, trained with RLCD…

There is a new category of AI model forming, and Jev just named it.The models we have are excellent at conversation and terrible at being components inside software. Priced per million tokens, optimized for long generation, calibrated for…

FUNDRAISE 💰: @typesafeai just came out of stealth with a $40M seed led by @DCVC. On @frontrunvc, the story started 71 days ago. jul 6: @typesafeai flagged. in stealth. jul 21: @frontrunvc posted in EARLY at 40 followers. sep 9: 9 tracked…