Summary
According to a post on X by @xingyudang, a new Writing Weights Index finds open-source models struggle to write weights directly. The post says a retest found GPT-6 Astra, GPT-6.1 Sol and Opus 5.5 reproduce the effect, while older models mostly do not. @csinva says Claude models can write functional neural-net weights from scratch, while Neel Nanda reads the result as models being surprisingly good at old-school interpretability. Log readers are also cited on Qwen's large gain, but the snippet is cut off before their full point.
AI-assisted summary based on the listed source.
What happened
A new index finds open-source models struggle to write weights directly, and a retest says GPT-6 Astra, GPT-6.1 Sol and Opus 5.5 reproduce the effect while older models mostly don't. @csinva says Claude models can write functional neural-net weights from scratch, while Neel Nanda reads it as models being...
Why it matters
Researchers and builders working on model interpretability or open-source models may want to track whether this gap holds up under further testing. The main open question is whether the gains reflect real new capability or mostly bug fixes, as some skeptics argue. Watch for the full retest results and the Qwen log analysis before drawing conclusions.
Signal Intelligence
Signal Strength 80%
Technical label WATCH
Public Interest 46
Category OPEN SOURCE
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 51
Practical Impact Score 28
Novelty Interest Score 94
Consequence Score 10
Curiosity Score 36
Shareability Score 41
Why this is here
VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).
