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

Core Technical Shift: Test-Time Compute & Reinforcement Learning

• O1 Model Breakthrough: The September release of O1 demonstrated that model performance improves not only with training-time compute but also with "test-time compute" (reasoning time).
  • Models now "think" for a duration before answering; increased reasoning time correlates directly with higher accuracy on mathematical benchmarks.
• O3 Model Capabilities: The subsequent release of O3, a superior reasoning model, successfully solved a quantum electrodynamics problem.
  • The model performed multi-step visual analysis and calculation verification.
  • Execution time: ~1 minute (compared to ~3 hours for a human physicist to verify via textbooks).
• Contrarian Scaling Thesis: OpenAI plans to invert the traditional AI scaling paradigm where reinforcement learning (RL) was a minor component.
  • Current State: Pre-training is the dominant "cake," with RL as a small "cherry."
  • Future State: RL compute will eventually dominate pre-training compute to maximize intelligence.
  • Dan Roberts explicitly advocates for scaling RL compute to crush current pre-training dominance.

Validation of Reasoning Capabilities

• Einstein Thought Experiment: O3 successfully solved a theoretical General Relativity exam question (involving black holes and wormholes) that GPT-4.5 failed to answer.
  • The model required only minutes to arrive at a solution Einstein spent eight years discovering.
  • This validates the model's ability to reproduce complex textbook calculations and perturbations.
• Current Limitations: Despite high performance on specific tasks, current models are described as "idiot savants."
  • They struggle with novel scientific discovery because training data may be overly focused on competition math problems.
  • Success depends heavily on the formulation of questions rather than just the answer generation process.

Strategic Roadmap & Investment

• Infrastructure Investment: OpenAI announced plans to raise $500 billion in capital.
  • Location: Construction of new data center facilities in Abilene, Texas.
  • Objective: To house massive compute clusters for training next-generation models.
• Scaling Science Imperative: Traditional scaling laws (like those proven by GPT-4 loss predictions) must be reinvented to accommodate test-time compute and RL.
  • The goal is to develop a new "scaling science" framework that predicts performance for models dominated by reasoning steps rather than just pre-training tokens.
• Revenue Loop: The strategy involves generating significant revenue from current models to fund the construction of additional compute infrastructure.

Forward-Looking Projections

• Exponential Task Length Growth: Agents can currently handle tasks lasting roughly one hour; this capacity is doubling every seven months.
  • 1-Year Projection: Task capacity expected to reach 2.5–3 hours.
  • 9-Year Projection: Extrapolating 16 doubling times (based on the "eight years of Einstein" analogy), Roberts predicts a model capable of discovering General Relativity will exist in nine years.
• Future Timeline: The organization aims to move from reproducing known science to making major contributions to human knowledge within this decade.