Extropic trains a model to replicate Hinton's experiments and triples its score

With about fifty exercises taken from Boltzmann's machines and the wake-sleep algorithm, Extropic increases Qwen3.6-35B-A3B from 0.127 to 0.361 on new search tasks.

Extropic, the start-up that designs thermodynamic computing chips, released on October 1st a first experiment called "Baby Thermo RSI". The objective: to teach a language model to conduct research in its field, probabilistic machine learning, by making it reproduce classical experiences of the connectionist tradition of Geoffrey Hinton, such as Boltzmann's machines or the wake-sleep algorithm.

A two-part reward

The team is part of the open model Qwen3.6-35B-A3B (35 billion parameters, including 3 billion active) and has post-strengthened it (GRPO) for 100 steps on Prime Intellect infrastructure, from about 50 adapted code tasks of open source reproductions. The reward weighs the execution of the code and fidelity to the reference metrics to 70%, and to 30% an evaluation grid rated by Nemotron 3 Super 120B, responsible for verifying the architecture and experimental procedure.

On tasks set aside, the score rose from 0.127 to 0.361, almost three times as much, however far from the 0.747 achieved by GPT-5.5. Extropic presents this result as a first spark and hopes, eventually, to have these models test new algorithms on its thermodynamic chips expected in 2027.