Summary
According to a post on X by @neilkale, Neil Kale and Jerry of Trajectory Labs post-trained Qwen 27B and 35B-A3B models to write Atari-style networks by hand, using WeightProgramBench with 322 simple games and five MinAtar games held out. On the held-out Freeway game, Qwen-27B's score rose from 21.0 to 47.1 points, compared with 63.9 for Astra. The post also says Playground reported that GPT-6 Astra hand-wrote a 19.8M-parameter manipulation policy, but the snippet cuts off before describing its results.
AI-assisted summary based on the listed source.
What happened
New since the earlier weight-space coverage: Neil Kale and Jerry (Trajectory Labs) post-trained Qwen 27B and 35B-A3B to write Atari-style networks by hand using WeightProgramBench (322 simple games, five MinAtar games held out); on held-out Freeway, Qwen-27B rose from 21.0 to 47.1 points vs Astra's 63.9....
Why it matters
Researchers and builders tracking weight-space programming can treat the Freeway gain as an early signal of progress on this approach. The figures come from a single social post, so they should be checked against the full work before being compared with other systems.
Signal Intelligence
Signal Strength 78%
Technical label WATCH
Public Interest 31
Category DEVELOPING
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 0
Practical Impact Score 0
Novelty Interest Score 94
Consequence Score 26
Curiosity Score 36
Shareability Score 45
Why this is here
VQV surfaced this signal because it is recent, relevant to AI Search, connected to Scobleizer News Pulse (X).
