Why Are Leading AI Companies Competing for Top Scientists?

At the recent Fields Medal announcement, winner Jacob Tsimerman revealed he is moving to OpenAI’s safety team, a shift echoed by OpenAI chief research officer Mark Chen’s warm welcome. This mirrors a broader trend where leading scientists from mathematics, physics, biology, and other fields are joining major AI labs such as Anthropic, where Nobel laureate John Jumper and others have recently signed on.

Chinese AI leader ByteDance’s Seed Edge team has responded by launching the Seed STEM Scientist Program, offering 100 collaboration slots for researchers in frontier sciences. The program aims to tackle fundamental scientific problems with AI, marking the first large‑scale domestic effort to accelerate scientific discovery through AI.

The article argues that the AI race is moving from short‑term benchmark scores to long‑term foundational research. It cites statements from industry figures—Yao Shunyu, Anthropic CEO Dario Amodei—emphasizing that model progress now hinges on compute, data quality, training duration, and scalable objectives rather than clever tricks.

Technical highlights include DeepSeek’s mHC (Manifold‑Constrained Hyper‑Connection) architecture, which builds on ByteDance’s Hyper‑Connections (HC) to improve feature expression without extra compute, and the EdgeBench long‑trajectory evaluation suite that defines new scaling laws for agent learning in realistic environments.

Benchmarking efforts such as OpenAI’s Evals framework, SimpleQA, BrowseComp, and MLE‑bench are discussed as essential for transparent, reliable model assessment. Anthropic’s SWE‑bench and Terminal‑Bench illustrate how targeted benchmarks align with model strengths.

Talent programs like OpenAI’s Residency, Anthropic’s STEM Fellows, and ByteDance’s Seed Edge initiatives provide scientists with AI compute, model access, and a collaborative environment, resulting in high‑impact academic outputs and accelerating model capabilities in scientific reasoning.

Overall, the piece concludes that the next phase of AI competition will be defined by deep interdisciplinary collaboration, long‑term research investment, and the ability of models to solve open‑ended scientific challenges, rather than merely achieving higher benchmark scores.