Meta researcher Rylan Schaeffer leaves to found startup, regrets not betting on inference scaling

Meta researcher Rylan Schaeffer published a long post on October 20 announcing his departure to start a company, systematically reviewing two years of pretraining research and career reflections. He candidly shared his two biggest regrets: not betting on inference-time compute scaling in January 2024, and not betting on automated AI research in April 2024—believing he saw the directions early but never went all in.

Confirmed

Schaeffer disclosed that as early as January 2024 he proposed to Ethan Perez studying the exchange rate between pretraining scaling and inference scaling, based on Andy Jones's famous paper—nine months before OpenAI released o1; he was also doing related research at Berkeley at the time.
Pretraining paper results with collaborators include: a generative evaluation scaling law with Joshua Robinson and others—using the model's log-likelihood of gold reference answers predicts GSM8K pass rates more tightly than compute scaling laws; scale-dependent data repetition effects; and non-trivial interactions between data repetition and model scale.
He closed the post thanking collaborators and sharing an insight: many laws of machine learning transfer to life and to designing larger systems.

Why it matters

Schaeffer's experience reflects the dilemma facing big-lab researchers amid paradigm shifts: even identifying the inference-scaling direction months early, failing to bet decisively can mean losing the first-mover edge. He frames his departure to found a startup as itself a bet on the next wave—automated AI research.

2026-10-20 ~ 2026-10-20 · 5 related posts

Primary sources

Ex-Meta Researcher on His Biggest Regrets: Skipping Inference Scaling and AI Automated Research — RylanSchaeffer · 2026-10-20 02:22 (UTC+08:00)
He pitched pretraining-inference compute exchange rate research 9 months before o1 — RylanSchaeffer · 2026-10-20 02:22 (UTC+08:00)
Gold-likelihood predicts GSM8K pass rates better than compute scaling laws — RylanSchaeffer · 2026-10-20 02:22 (UTC+08:00)

[source] He pitched pretraining-inference compute exchange rate research 9 months before o1 — RylanSchaeffer · 2026-10-20
[source] Gold-likelihood predicts GSM8K pass rates better than compute scaling laws — RylanSchaeffer · 2026-10-20
Schaeffer wraps pretraining work: generative eval scaling laws and data repetition — RylanSchaeffer · 2026-10-20
[source] Ex-Meta Researcher on His Biggest Regrets: Skipping Inference Scaling and AI Automated Research — RylanSchaeffer · 2026-10-20

1 near-duplicate retellings: RylanSchaeffer