LLMs are not superhuman at reading and writing neural network weights
The author argues that recent findings showing frontier large language models directly reading and writing neural network weights are not superhuman or radically new. Analyzing examples such as identifying trained games from network weights and patching a Qwen model to fix looping behavior, the author contends that these tasks mirror standard, competent human machine learning workflows rather than signaling recursive self-improvement. Open article →
Notable comments
Commenters largely agree the “LLMs are superhuman at reading/writing weights” framing is overstated, but several push back that speed/efficiency can still produce genuine, sometimes qualitative, capability leaps.
HN · nagi_builds Challenges the implicit premise that a human can do the same tasks—by asking how long that would take—pointing to scalability as the real differentiator.
HN · nfw2 Argues that “superhuman” here is only in throughput (faster/more effective), emphasizing these systems still follow human-like reasoning rather than any Matrix-style exotic weight-as-symbolic-channel.
HN · antonvs Pushes back that faster/effective capability is unambiguously “superhuman” and can cause qualitative leaps, citing LLMs outperforming humans in understanding unfamiliar, complex codebases.