This post is a frame-by-frame fact-check of a Hacker News thread that reached nearly 900 points, covering an MIT CSAIL blog post (rikiiyat.github.io, Nov 2, authors Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab) claiming that a model the authors call "GPT-6 Astra" can natively read and write neural network weights, and used this to improve Qwen3.5-0.8B's math performance without touching gradients.
Key points
- The claimed results: Editing 131,072 weights (0.016% of 0.8B parameters; verified as 64 columns × 2 MLP layers × 1024) lifted MATH-500 from 7.9 to 48.1, GSM8K from 13.8 to 67.1, IFEval from 21.0 to 62.3, and cut token-limit failures from 93.0% to 33.1% (at a 32,768-token limit with thinking enabled). Astra located a "thinking loop" direction in layer 15 and 19 MLPs and subtracted it from 64 output columns per MLP, using Python analysis and up to 50 forward passes.
- The core critique (Matthew Honnibal, Nov 3): The headline bundles two different techniques into one hat, like a magician. The raw-weight experiments are real: Astra identified Breakout, Pong, Pac-Man, and Freeway from a 17,253-parameter PPO policy alone (4/4), and identified languages from mid-layer-only character transformers — but only works for networks small enough that all weights fit in context (17K–172K parameters). The flagship Qwen results required Python tooling and up to 50 forward passes — a different skill. Honnibal's hardest line: "it's not a single technique." His critique targeted the abstract/TL;DR wording, not the blog title itself. He conceded that models from a year or two earlier likely couldn't do any of these tasks.
- Reported failures (credit to the original authors for publishing them): Fine-tuned-for-sycophancy Gemma 3 could not be recognized from raw weights (only with attention patterns); Gemma 3 270M could not be math-repaired the same way, moving at most one point — obstacles can be removed, but missing capabilities cannot be installed.
- Independent replications: Simon Willison ran the raw-weight game-identification locally with Qwen 3.8 27B (17GB Q4 quantized), correctly identifying 3 of 4 games. Kimi K3.1 Pro via API identified Breakout in 9 minutes 41 seconds. Sean Goedecke noted the Qwen repair was essentially standard ML engineering performed to a script.
- Verdict: Not fraud, but headline metrics — welding "can read weights" and "can fix models" into one sentence makes readers pay an extra tax of imagination. The fix is simple: score raw-reading and tool-assisted modes separately under the same harness.
- Pong/Pac-Man conflation: Willison's notes show Pong took 1h 4m 54s (75% confidence, token count unreported); the 43,506 inference tokens belonged to Pac-Man (46m 12s, 72%). Breakout was 52m 25s at 88%; Freeway failed. The original post's "1h5m / 40k+ tokens" fused two runs.
- The 48.1 MATH-500 score only holds at 50 forward passes: 10 and 25 passes yield 12.3 and 14.1 — a threshold-style jump, not a smooth ramp.
- "No third-party verification" was too absolute: Qwen scores lack independent full re-computation, but a Hugging Face rebuild from public logs reportedly scored 51.0%; Willison replicated only the raw-weight game identification.
- HN metadata: The main HN post was created Nov 2, 18:21 UTC; Nov 3 brought the rebuttals and replications. The 898 figure cannot be traced to a primary snapshot; the current reading is 2,949 points and 892 comments — the proximity of the two numbers invites confusing comments for upvotes.
- A diagram error in this post itself: The 7.2M (actually 7.16M: 4 layers × 6 heads × width 384) character transformer was drawn under the "raw weights only" branch, but Experiment 3's setup explicitly allowed Python analysis (Honnibal's table: "Reads raw weights? No / Python? Yes"), and 7.16M weights do not fit in context — only Experiments 1, 2, and 4 (17K–172K) were true raw-reading. The diagram is left in place as a record of the error; this text is the correction.