Latest models learned to write a 20,486-weight Breakout network in one reply
Alex Krentsel reran MIT CSAIL's Breakout experiment across model versions and says only the latest models can write all 20,486 weights in one reply.
By Ryan Merket · Published
Primary source: x.com
Why it matters
If reasoning models can read and edit weights directly, the same system that diagnoses a failure in a network could patch it without a training run. Krentsel's version-to-version comparison suggests this is a measurable emerging capability.
Alex Krentsel (@AlexKrentsel) says he has reproduced the central Breakout result from an MIT CSAIL post on weight-space programming, with one addition: the ability to write a small neural network's weights in a single reply appears only in the latest set of models. Krentsel described the pattern in a thread on X as a step-change emerging capability, while stressing that the network involved has 20,486 parameters and frontier models are on the order of 1 trillion parameters.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.