9 Years to AGI? OpenAI’s Dan Roberts Reasons About Emulating Einstein

OpenAI’s Dan Roberts argues reinforcement learning will dwarf pre-training and projects AGI-level scientific discovery in 9 years.

• o3 solved a quantum electrodynamics problem in ~1 minute; Dan Roberts, a physicist, needed 3 hours to verify the same calculation.
• Test-time compute is a new scaling dimension: the longer o3 thinks, the better it performs — independent of training compute.
• Roberts’s contrarian thesis: RL compute will eventually “crush” pre-training, inverting Yann LeCun’s cake-vs-cherry analogy from 2019.
• GPT-4.5 failed a general relativity exam question that o3 answered correctly, used as a proxy for Einstein-level reasoning.
• Agent task length is doubling every 7 months; current models handle ~1-hour tasks, implying ~2–3 hour tasks by next year.
• 9 years to a model that could independently discover general relativity — extrapolated from 16 remaining doubling periods.
• OpenAI is raising $500B and building compute infrastructure in Abilene, Texas under the Stargate initiative.
• Old pre-training scaling laws are obsolete; OpenAI must reinvent predictive scaling science for RL and test-time compute regimes.

2025-05-08 · Watch on YouTube