A new technical critique is challenging how experiments involving the Astra model have been presented, arguing that separate tests using substantially different methods were bundled into a broader claim that the system could understand and improve neural networks by directly reading and writing their weights.

The analysis, published by software developer and natural-language-processing researcher Matthew Honnibal, distinguishes small-network experiments from the headline result involving Qwen3.5-0.8B. According to Honnibal's review, experiments that placed raw floating-point weights in the model's context without tools involved networks ranging from 17,000 to 172,000 weights. The language-model work, by contrast, used Python programs and opportunities to run the target model.

That difference matters because the Qwen result has been described as Astra identifying a cause of reasoning loops and improving performance by changing a small share of model weights. Honnibal says the procedure was not based only on inspecting those weights. His account says Astra could run Qwen on as many as 50 batches of practice problems while devising and applying an intervention through code.

The critique does not dismiss the outcome as trivial. It characterizes the work as evidence that a capable AI system can carry out tasks resembling those of a human machine-learning engineer: writing analysis code, examining model behavior and modifying parameters productively. The objection is to treating that tool-assisted workflow as the same capability tested when raw weights are supplied directly in context.

Honnibal also separates claims about game-playing networks. In his summary, a network generated from zero-initialized weights had no existing learned structure for Astra to interpret. Reported comparisons were mixed across games: the generated system exceeded one PPO baseline on Breakout but trailed it on Seaquest and Asterix. Those details make a single, general claim about reading and authoring network intelligence harder to sustain.

The disagreement illustrates why method descriptions are especially important in demonstrations of unusual AI capabilities. Reading a compact network's numeric parameters in context, writing code that probes a larger model, and improving performance after repeated test runs are all technically interesting, but they answer different questions. Combining them under one headline can encourage readers to infer a common mechanism that the experiments may not establish.

The critique is one expert interpretation of the published experiments, not an independent replication. Its core contribution is methodological: it asks that results be labeled according to the information, tools and feedback available in each test. Further evaluation would need to reproduce the procedures and compare them under consistent constraints.