My infant year as an AI researcher – Moving from physics to AI (alfredyao.github.io)

A former physics postdoc describes a year-long transition into AI research, driven by the desire for a field with more experimental feedback and faster career opportunities for early-career researchers. They argue that large-scale AI research resembles 17th-century thermodynamics: we lack a complete theoretical understanding of neural networks, but systematic experiments (e.g., scaling-law studies) reveal regularities that guide rapid capability gains. This empirical, engineering-driven progress attracted them away from theoretical physics and toward AI work that yields immediate, measurable impact. They joined Anthropic (starting Oct. 1) and contributed during the Claude 3.7 → 4.5 period, but resigned Sept. 19 and moved to Google DeepMind on Sept. 29. Their departure was partly political—strong disagreement with Anthropic’s public anti‑China stance (~40%)—and partly due to internal, unspecified reasons (~60%). They praise Anthropic as a strong entry point for physicists but criticize the lab-specific knowledge silos and a trend where core research increasingly happens behind closed doors rather than through papers. The account underscores two broader signals for the AI/ML community: empirical, systematic experimentation remains central to progress, and institutional culture and geopolitics are now material factors shaping researcher movement and where core innovations happen.