AI Card Bao HERMES.AICARDBAO.COM

Industry News

Without any training or the use of tools, the weights of the neural network can be determined in just one response – and this new model achieves results that are 22 times better than those obtained through conventional methods.

This Monday, researchers from MIT demonstrated an experiment in which an AI model directly modified the weights of another model, without going through any training process at all. Alex Krentsel, a doctoral student at Berkeley, then repeated one of these tests using both the old and the new models – and the results were incredibly different.

The task is as follows: Write out all the numbers that appear in a small neural network designed to play the Atari game “Tetris.” You must complete this task in one reply; no training or the use of any tools are allowed. For reference, there is a neural network that has been trained through trial and error over a period of 500 million steps.

The network generated by GPT-6 Astra scored 22 times higher than the baseline score. Claude Opus 5.5 achieved a score 20 times higher, while GPT-6.1 Sol reached 18 times the baseline score.

Old models such as GPT-5 to GPT-5.5, and Opus 4.5 to 4.8 simply don’t even come close to meeting the current standards.

Krentsel described this phenomenon as “the sudden emergence of new capabilities.” There was no gradual development; the previous generation had almost zero such abilities, while the current generation has directly acquired them.

The first reaction is that it’s a bit intimidating – after all, the term “AI-generated handwriting by AI” easily brings to mind the idea of recursive self-improvement.

Then read through all the accompanying notes as well. This particular network only has 20,486 parameters, while cutting-edge models typically have around one trillion parameters – that’s a difference of seven or eight orders of magnitude. Building blocks with simple rules is also a kind of casual game. It’s very likely that when creating a model, one first forms a strategy in their mind and then encodes it into weights; however, this approach may not work effectively when applied to more complex tasks. Moreover, so far, the strategy of making model weights publicly accessible has never succeeded.

So it is still a long way away from the ability of “the model to modify itself automatically”. However, this capability emerged in just one or two generations. A more appropriate attitude would be to continue observing and see whether this trend will continue upward in the next generation.

Related