Unsafe Subintelligence announces a landmark result in machine pharmacology** CaveatThe result was null.

Simulated stimulants have no measurable effect on language-model reasoning, the institute confirms, in what its director called "the most important finding of the decade."

LONDON, 30 SEPTEMBER 2026. Researchers at Unsafe Subintelligence today published the results of Experiment 0042, a preregistered, double-blind study in which a 7-billion-parameter language model was administered a simulated stimulant across 120 trials. Accuracy on a standard reasoning set was 61% under placebo and 63% under the highest dose. The difference was not statistically significant (p = 0.61).

We set out to prove that more is better. We have instead proven, with great precision, that more is the same. That is still more than anyone else has proven.

The institute has announced a follow-up study with a larger dose. “If some is good, more is better,” said Dr. Ada Plenty, Head of Dosing. “And if more is not better, we have not yet tried enough.”