The video covers MIT research in which GPT-6 Astra was used to work directly on the weights of other neural networks. The presenter, Dr Károly Zsolnai-Fehér, describes three results. First, Astra examined a network's weights and identified which games it had been trained on, including simplified versions of Breakout, Pong and Pac-Man. Second, it edited the weights of a failing driving agent so that a car which had been stuck on a ramp then got through.
The third result is the one the video dwells on. According to the presenter, Qwen, an open-weights model that can be run at home, solved about 8% of tasks on a math benchmark before the edit and about 48% afterwards. The explanation given is that Qwen was not weak at maths but was falling into repetitive thinking loops. Astra tested the model repeatedly, located the parts of the network responsible for the loops and edited them to suppress that behaviour.
The second half is a list of corrections to coverage elsewhere. The presenter says Astra did not teach the model mathematics, it helped unblock it. He says the three abilities were shown in separate experiments, mostly on small systems with custom setups, not all at once. He adds that independent experiments have already read which game a network plays, using smaller open models, so the capability is not exclusive to OpenAI. Astra also failed to find meaningful improvements for Gemma 3, so the approach does not always work.
He closes with a further result: Astra writing a game-playing network from scratch, without reinforcement learning. He notes that a benchmark for this kind of weight space programming is being built. The conclusion is that the work is impressive but narrow, and the video ends with a sponsor segment for Lambda's GPU cloud.
Key takeaways
- The video describes three separate experiments: reading a network's training games from its weights, repairing a driving agent, and improving Qwen's maths score.
- The presenter reports Qwen rising from about 8% to about 48% on a math benchmark after the weight edit.
- He attributes the gain to removing repetitive thinking loops, not to teaching new mathematics.
- The abilities were shown in separate setups, mostly on small models, not combined in one system.
- Astra reportedly found no meaningful improvement for Gemma 3, and similar reading experiments exist with open models.
- The video ends with a Lambda sponsor read.
Our take
The useful part of this video is its scepticism. The presenter is the one flagging that the headline results were separate, small-scale experiments, and that the Qwen figures come from suppressing a failure mode, not from added capability. Anyone running open models at home should read it as a hint about where weight-level tooling may go, not as something they can use today.
As a caveat, this is one creator's summary of the research, and he says himself he may be wrong. The 8% and 48% figures are his account of the paper, and we have not independently checked them. None of our current coverage bears directly on this work.
