Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI Research Scientist Noam Brown

Summary

Noam Brown argues that traditional benchmark grids are obsolete because model performance now depends heavily on inference budget: the more tokens, time, or money you allow at test time, the more capable the same checkpoint can look. He says this makes safety and release policies harder too, since preparedness frameworks often evaluate a model as if capability were fixed, when in practice it is budget-dependent. Brown uses practical examples like poker bots, cyber evaluations, and an internal OpenAI result on the Erdős unit distance conjecture to show how scaffolded models can keep improving over long horizons. He also pushes back on the idea of an overnight intelligence explosion, arguing that model-assisted research is real but still bottlenecked, and that the right direction is cost-aware, budget-controlled evaluation rather than benchmark-maxing.

Notes

• Modern model capability should be treated as a function of test-time compute, not a fixed checkpoint property.
• He says old models like GPT-3 could not productively use very large inference budgets.
• Benchmark grids are misleading because they collapse a budgeted process into a single score.
• The benchmark presentation hides how much longer newer models think.
• He says modern models can keep improving for weeks when scaffolded properly.
• The plateau point is too far out for practical benchmark testing.
• He argues evals should either impose a budget or plot performance against test-time compute.
• Cost, tokens, or time should be explicit axes in evaluation.
• He says cyber-style tasks can still improve at very large token budgets.
• AISI evaluations are cited as evidence that 100 million tokens is still not enough to hit the ceiling on some cyber tasks.
• Current safety and preparedness policies do not clearly specify the inference budget at which a model should be judged.
• He says a model should be evaluated at a budget, but policies today leave that open.
• Routing or scaffolding can look better on benchmarks without proving real-world gains.
• He warns that benchmark optimization can fail to transfer to practical use.
• He uses poker-bot building as a hands-on eval for model reasoning.
• Poker bots are useful because they force reasoning through many hidden edge cases.
• Early models were unusable for poker-bot work, but later ones helped substantially.
• He says a model-assisted river solver was roughly 5x faster than solo work.
• He says model-assisted code optimization reached 10x to 100x speedups in some research workflows.
• The models are accelerating research, but unevenly across the workflow.
• Bottlenecks remain in the parts of research that do not speed up as fast as model assistance.
• He rejects the idea of a sudden overnight intelligence explosion.
• He thinks long test-time compute makes an abrupt RSI jump less likely.
• Sudoku is his example of a task that should keep improving with more compute without an obvious ceiling.
• He distinguishes tasks that benefit from more thinking from tasks that do not.
• He says models complement researchers rather than replace the whole research cycle.
• General-purpose scaffolds could make proving hard math results cheaper over time.
• OpenAI’s internal model reportedly helped disprove the Erdős unit distance conjecture.
• He says the cost of that kind of math discovery could fall by 10x to 100x per release cycle.
• Release cadence matters because every few months the model ceiling changes again.
• He argues recursive self-improvement will look more like incremental acceleration than a sudden explosion.
• Multi-agent coordination is still underexplored at sufficient scale.
• He thinks large-scale coordination among models is emerging but not mature.
• Global knowledge sharing among models is a key missing ingredient.
• He says frontier competition is intense, but the direction of progress is still positive.
• He says models are now good enough for some high-stakes everyday use cases.
• He suggests older trust concerns from 2023 may be less relevant today.