Nicholas Andresen

More chain of thought, not neuralese.

So the frontier remains text.

OpenAI's O-Series, Anthropic's Extended Thinking, Google's Gemini, DeepSeek, all of them reason in words.

No major lab has shipped a Neuralese model that can compete.

But Neuralese hasn't won yet is not Neuralese won't win.

Eventually someone might crack it.

And if they do, there's nothing to monitor.

Neuralese or not, the pressures towards opacity are real.

Reasoning probably won't stay monitorable forever.

The question is whether it stays that way long enough.

Keeping reasoning monitorable has a cost.

It means foregoing whatever optimizations would make chains of thought faster or more capable but harder to read.

For a given budget, that's performance left on the table.

Researchers call this the monitorability tax.

Right now, the tax is small.

Frontier labs can pay it and remain competitive in the market.

Nobody's going bankrupt from keeping their model's reasoning in English instead of glean, disclaim, disclaim, vantage, illusions, overshadow.

But suppose new release cracks the scaling problem.

If thinking in vectors proves substantially faster, cheaper, more capable than thinking in words, the tax explodes.

Any lab that pays while competitors don't will ship worse models on every axis the market cares about.