Anthropic's founding team (Dario & Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, Chris Olah) just dropped their origin story.
This crew literally built GPT-3 at OpenAI before spinning out. Tom Brown was first author on the GPT-3 paper. Chris Olah pioneered interpretability research (the circuits work that's now foundational). Jared Kaplan co-authored the scaling laws paper that predicted how model performance scales with compute.
They left OpenAI in 2021 over safety disagreements and raised $124M to build Constitutional AI - basically training models to be helpful, harmless, and honest through self-critique loops instead of pure RLHF.
Claude's architecture uses their Constitutional AI approach: the model critiques its own outputs against a constitution of principles, then trains on the better responses. This is why Claude tends to refuse less and reason more compared to early ChatGPT.
The technical bet: interpretability + constitutional training >> pure scaling. They're trying to understand what's happening inside the model while making it safer, not just making it bigger.
Worth watching if you care about the actual technical lineage of modern LLMs and how the safety vs capabilities debate split the field.
This crew literally built GPT-3 at OpenAI before spinning out. Tom Brown was first author on the GPT-3 paper. Chris Olah pioneered interpretability research (the circuits work that's now foundational). Jared Kaplan co-authored the scaling laws paper that predicted how model performance scales with compute.
They left OpenAI in 2021 over safety disagreements and raised $124M to build Constitutional AI - basically training models to be helpful, harmless, and honest through self-critique loops instead of pure RLHF.
Claude's architecture uses their Constitutional AI approach: the model critiques its own outputs against a constitution of principles, then trains on the better responses. This is why Claude tends to refuse less and reason more compared to early ChatGPT.
The technical bet: interpretability + constitutional training >> pure scaling. They're trying to understand what's happening inside the model while making it safer, not just making it bigger.
Worth watching if you care about the actual technical lineage of modern LLMs and how the safety vs capabilities debate split the field.