OpenAI lets Apollo, METR and two more groups test models from inside
Apollo Research gets training code, datasets and mid-training checkpoints for a four-month pilot, but OpenAI can still redact evidence from its reports.
Every ClaudeGPT Si story about Alignment, newest first: news, analysis, reviews and guides.
2 stories · Updated
Apollo Research gets training code, datasets and mid-training checkpoints for a four-month pilot, but OpenAI can still redact evidence from its reports.
Reinforcement learning from human feedback turned raw language models into assistants. Here is how it works, where it fails, and how DPO simplifies it.