@0xLogicrw: Shunyu Yao, former Anthropic research scientist and current Google DeepMind research scientist, first revealed the internal R&D process of Claude 3.7 on @zhang_benita's podcast "Language is World". He joined Anthro…

Shunyu Yao, former Anthropic research scientist and current Google DeepMind research scientist, first revealed the internal R&D process of Claude 3.7 on @zhang_benita's podcast "Language is World". After joining Anthropic in October 2024, he was assigned to a team called Horizon. At the time, the team had only 10 to 11 members but handled all of Anthropic's reinforcement learning efforts, covering data, infrastructure, and algorithm research. From initial research to final release, Claude 3.7 took roughly four to five months: the first two to three months focused on algorithm and data research, followed by two months on training and infrastructure setup. Anthropic's bet on coding capabilities wasn't planned from the start. Yao revealed that Claude 3's superior coding performance over GPT-4 stems from a purely technical reason he cannot publicly disclose, which was developed bottom-up by a specific team. After Claude 3's launch, overwhelming positive feedback on Twitter validated this strength, prompting Anthropic's management to elevate coding capabilities to a company-wide strategic priority and go all-in. He attributes Anthropic's ability to pivot and bet heavily so quickly to the fact that its top technical leaders, Jared Kaplan and Sam McCandlish, are also co-founders. They command technical respect and hold final decision-making authority. OpenAI lacks this dynamic; it might have worked when Ilya was there, but he left after losing that authority. Back then, Anthropic had almost no product sense. They shipped two versions of Claude 3.5 within six months under the exact same name, only distinguishing them thanks to the community-coined nickname "3.6".