@Phoenixyin13: Finished reading a long post today by OpenAI researcher Noam Brown — a reality severely underestimated by the industry. The true ceiling of LLM capabilities is far higher than what any current benchmark shows. The reason: too little test-time compute. And as models...

I read through OpenAI researcher Noam Brown’s long post today — a reality that is seriously underestimated by the industry.

The true capability ceiling of LLMs is far higher than what any current benchmark shows.

The reason: they are given too little test-time compute. And as models evolve, this problem will only become more pronounced.

When GPT-5.5 was first released, its benchmarks only showed a slight improvement over GPT-5.4, leading many to feel it wasn’t that impressive.

But the actual experience was a clear step-change.

Why? Because test-time compute was not controlled for during evaluation.

Once you plot tokens on the x-axis, 5.5 is clearly stronger at the same compute budget. A single benchmark number masks this gap.

A more critical observation: stronger models benefit more from additional test-time compute.

The performance improvement curve is steeper, the plateau keeps getting pushed further away — it may not even exist.

Karpathy’s autoresearch experiments showed performance still rising after hundreds of trials.

The AI Security Institute’s cyber attack/defense evaluation showed top models still improving rapidly after 100 million tokens.

This suggests the capability ceiling may be much further away than we think.

The implications for AI safety and preparedness are huge.

Most current safety evaluations are only conducted under low inference budgets.

But truly dangerous capabilities — especially long-horizon, multi-step planning, complex agent behaviors — are likely only revealed under high compute budgets.

The controversy around Gemini 3 Deep Think exposed this problem. Scaffolding plus high inference can let ordinary users approach high capability, but the system card did not reflect this curve.

Noam points out a more stubborn reality: evaluating long-horizon tasks is becoming increasingly difficult.

Some capabilities can only be verified through actually running them over long periods, and the model iteration cycle may already be shorter than this verification time.

In the future, labs may face a dilemma: either delay release significantly, or release with incomplete safety evaluations.

More than two years have passed since o1, yet the industry still habitually reports single numbers, and safety organizations are still surprised when spending 100x more inference leads to significant capability improvements.

This is actually the most underestimated paradigm shift of the post-o1 era.

In the future, the real gap between models will be reflected in who better understands and leverages the intelligence-compute curve.

For those working on agents, long-horizon reasoning, and high-risk deployments, re-understanding this curve has become a top priority — an absolute must.