Interpretability researchers have spent years trying to make sense of neural network weights. A new MIT preprint claims that frontier AI can now read and write them unaided. According to this preprint, GPT-6 Astra handcrafted a Breakout-playing network without any training, and it outscored a conventionally trained one. It also took a small language model that kept getting stuck on math problems, edited just 0.016% of its weights, and lifted its score from 8% to 48%.

Questions remain on whether these results generalize beyond a handful of small-scale experiments. An AI that writes weights is also one step away from rewriting its own, possibly opening a new path to the risks of recursive self-improvement.

In this month’s benkyoukai, we will look at the MIT preprint alongside METR’s independent assessment of AI R&D acceleration at Anthropic. Join us to separate the evidence from the hype!

We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, using a variety of recent tools and algorithms or even via direct weight editing. These results show sparks of Weight-Space Programming, which may extend beyond what we’d expect from humans or conventional algorithms. This might be a big deal for both optimization and interpretability.

Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do. And without any environment feedback, it has no trouble writing, from scratch, small game-playing networks that outplay networks trained with PPO for 100 million steps.

Likewise, asked to make Qwen3.5-0.8B better at math without gradients, Astra identifies an issue in its weights causing reasoning loops that Qwen’s own model card warns about, and fixes this by natively editing 0.016% of the weights. With the same sampling settings, this weight correction raises MATH-500 from 7.9% to 48.1% and generalizes out of the box to instruction following, raising IFEval by 3x.

Rikiya Takehi et al., LLMs can natively read & write neural network weights (2026)