Noam Brown argues model benchmarks are misleading because they ignore test-time compute, so the published grid hides how much better newer models like 5.5 actually are.
Brown says the safety frameworks of labs ignore the inference budget at which a model is evaluated, even though capability now depends on money spent.
Brown reports an OpenAI internal model disproved the unit Erdos distance conjecture at low cost, and he expects such results to fall 10 to 100 times in cost each release cycle.
AI Benchmarks Hide Thinking Time, and That Distorts Progress and Safety
Model evaluation has not caught up with test-time compute. A single benchmark score hides how much thinking a model used, so it understates real gains and misleads on capability. Because capability now scales with money spent and models can work for weeks, evaluations and safety policies need to state a compute budget.
Takeaways
Benchmark grids ignore how long models think. Published grids show one score per benchmark, so 5.5 looked only a few points ahead of 5.4 on some tests. 5.5 thinks more efficiently and 5.4 at max settings thinks far longer, so compute is uncontrolled and the gap is hidden.
Evaluate performance against test-time compute, not single scores. Running until performance plateaus no longer works, since scaffolded models can think productively for weeks. Brown proposes fixed budgets in tokens, cost or time, or plots of performance against compute. AISI evaluations show models still improving at 100m tokens.
Safety policies do not specify an inference budget. Preparedness frameworks and responsible scaling policies were built when test-time compute scaling barely mattered. Capability now depends on money spent, yet policies do not say at what budget to evaluate. Brown takes no stance on release decisions but wants this addressed.
Fast release cycles leave capability ceilings unknown. Models can run experiments for weeks or months, and labs release every two or three months. The slash goal release showed this, since its impact took a week to appear as people ran tasks longer than a week. Thorough pre-release testing is hard under competition.
Capability costs fall 10 to 100 times per release. An internal OpenAI model disproved the unit Erdos distance conjecture at a very low budget, and the result was verified. Brown estimates a general scaffold on 5.5 could reach it for roughly $1,000 to $100,000.
More compute has limits and no overnight takeoff. Extra thinking does not help factual recall but keeps helping on problems like Sudoku. Models still have weak research taste and complement researchers. Brown rejects an overnight intelligence explosion because heavy test-time compute makes time the main bottleneck.
Scaffolding and routing can inflate benchmark gains. Running a model five times and picking the best answer, or using a judge, can raise scores without real improvement. Brown suggests held-out private test sets and says routing gains may be fragile and fail to transfer to real use.
Brown's poker bot test of model progress
Early models could not help with poker bots, which Brown uses as a personal evaluation because little open source code exists.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
With 5.2 he built a reverse solver about five times faster than alone and optimized code roughly 10 times faster.
That model sometimes seemed to gaslight him, once saying folding a $100 pot loses $92.
5.5 could do much of a full poker solver zero-shot with gentle steering.
Brown predicts that within six months to a year a model could build his entire PhD thesis in one go.