Posts dispute how directly Astra can interpret and improve neural-network weights

One post reports that GPT-6 Astra can inspect, edit and generate neural-network weights, including a reported improvement to Qwen3.5-0.8B. A response argues that the headline combines different methods: the Qwen experiment used Python and model runs, unlike the smaller tests that examined raw weights without running a network.

• The first post says Astra edited 0.016% of Qwen3.5-0.8B’s weights and raised its MATH-500 score from 7.9% to 48.1%, with IFEval also increasing threefold.· 1 source
• The critique says the Qwen test used Python and up to 50 batches of practice problems, while raw-weight inspection without tools or network runs was limited to smaller experiments.· 1 source
• The posts also describe different smaller-network results: the first reports successful weight reading and game-network generation, while the critique says the game-generation test beat 500-million-step PPO only in Breakout, not Seaquest or Asterix.· 2 sources