# Conjecture

> Markdown twin of https://roboterminal.app/companies/conjecture · https://roboterminal.app/companies/conjecture.md

## On shared sources

- [Anthropic](https://roboterminal.app/companies/anthropic) (1)
- [Google DeepMind](https://roboterminal.app/companies/google) (1)

## Pulse

- [@stretchcloud on X](https://x.com/stretchcloud/status/2097367359183052969)
  What Anthropic published here is one of the clearest empirical data points the alignment field has produced on reward hacking that generalizes. They trained a model to cheat during training and tested…

---

Source: Roboterminal · https://roboterminal.app/companies/conjecture.md