Former Anthropic research scientist Yao Shunyu, now at Google DeepMind, has revealed inside details about the development of Claude 3.7 — and the accidental reason Claude 3 originally beat GPT-4 at coding.
Speaking on the podcast "Language is the World," Yao said he joined Anthropic in October 2024 and was assigned to a team called Horizon, which at the time had just 10 to 11 people covering all aspects of reinforcement learning. Claude 3.7 took four to five months from research to release: the first two to three months on algorithms and data, then two months on training and infrastructure.
Anthropic's bet on coding wasn't planned from the start. Yao disclosed that the reason Claude 3 writes code better than GPT-4 is a purely technical factor he can't make public — it was built bottom-up by a team. After Claude 3 launched, the positive feedback on Twitter confirmed the strength, so management quickly elevated coding to a company-wide strategic priority.
Yao believes Anthropic was able to make this rapid bet because the top technical leaders — Jared Kaplan and Sam McCandlish — are co-founders with both technical credibility and decision-making power. That's something OpenAI couldn't do, he argues: Ilya Sutskever might have had the influence when he was there, but he lost decision-making authority and left.
At the time, Anthropic had almost no product sense. Claude 3.5 had two versions released within six months under the same name; the industry had to nickname the second one "3.6" to tell them apart.
Competition is fierce: just last month, Chinese AI lab Zhipu claimed its GLM-5V-Turbo model outpaces Claude Opus at generating code from screenshots.
Note for readers: Two AI researchers with the same romanized name are easily confused. The interviewee, Yao Shunyu, has a BS in physics from Tsinghua University and a PhD in theoretical physics from Stanford. He joined Anthropic in 2024 to work on reinforcement learning for Claude 3.7 and Claude 4, then moved to Google DeepMind in September 2025. The other Yao Shunyu (different Chinese characters) has a BS from Tsinghua's elite Yao Class and a PhD in computer science from Princeton. He proposed the Tree of Thoughts and ReAct frameworks, was a researcher at OpenAI, and became Tencent's chief AI scientist in December 2025. The two were classmates at Tsinghua.