Researchers at MIT CSAIL have demonstrated "weight-space programming," using the GPT-6 Astra model to directly interpret and alter the raw weights of neural networks without training loops or gradients. In one test, Astra authored network weights from game rules alone, outperforming models trained with PPO for 100 million steps. In another, Astra altered just 0.016% of the weights in the small model Qwen3.5-0.8B, which raised its MATH-500 accuracy from 7.9% to 48.1% and reduced instances of the model hitting token…
Key points
Standard linear program solvers can match Astra's eight-hour weight-space programming pro…: Standard linear program solvers can match Astra's eight-hour weight-space programming process within approximately two one-minute runs.
Targeted linear programming modifications require changing eight times fewer weights than…: Targeted linear programming modifications require changing eight times fewer weights than the trial-and-error approach utilized by Astra.
Direct weight manipulation resolves recurring reasoning loops but fails to inject novel f…: Direct weight manipulation resolves recurring reasoning loops but fails to inject novel functional capabilities into models that do not already loop.
Sources checked by our analysis pipeline: 1
Last week, MIT CSAIL published Sparks of Weight-Space Programming. Given the raw weights of a small network, GPT-6 Astra could tell which game it had been trained to play. Given only a game’s rules, it could write the weights of a small network from scratch and beat networks trained with PPO for 100 million steps. No gradients, no training loop, just reasoning about what the weights compute. They coined this weight-space programming.
In another experiment, they asked Astra to make Qwen3.5-0.8B better at math by editing its weights. Astra changed only 0.016% of the weights, and MATH-500 accuracy jumped from 7.9% to 48.1%.1
Looking at Astra’s edit, I realized most of its search was solving a linear program by trial and error. It spent about 8 hours and 28 forward passes choosing which weights to change and by how much. A couple of LP solves, about a minute each, do the same job with 8× fewer changed weights. In this post I’ll walk through the edit, why it’s an LP, and what that says about the fix itself.
Specialized cloud infrastructure providers are seeing renewed market momentum that directly reinforces the broader semiconductor supply chain. Market observers interpret these recent gains as strong evidence that enterprise artificial intelligence spending is expanding rapidly rather than cooling off. This trend provides critical reassurance regarding long-term revenue visibility for key hardware suppliers.
Global financial markets are bracing for a historic milestone as one of the world's leading artificial intelligence developers heads toward a landmark public listing. The move will serve as a crucial barometer for global investor appetite regarding the long-term commercial viability of generative technology. Extensive backing from prominent industry giants has positioned the transaction to potentially rank among the most valuable market debuts ever recorded.
While pitching investors ahead of a public offering, artificial intelligence leadership is outlining sweeping transformations for global healthcare. The enterprise anticipates its technology will drive rapid solutions for some of humanity's most complex and lethal medical afflictions. These public projections underscore the aggressive timeline and massive societal impact tech executives are attaching to modern model capabilities.