Oriel — AI & technology worth noticing, every day.

Adventures in Weight Space Programming

Web page · https://withbeaker.ai/blog/weight-space-programming

First seen by Oriel on .

AI summary

This post reports that Beaker can program Qwen3.5-0.8B’s weights without running the model, using a single-neuron edit to reduce “Wait” loops and substantially improve MATH-500 accuracy.

  • In about an hour, Beaker identified layer 22 neuron 573; doubling its output raised MATH-500 accuracy from 7.4% to 37.4%.
  • The edit lowers the probability of “Wait,” reducing token-limit failures; a plain “Wait” token ban achieved 37.6%, suggesting no new math ability.
  • Combining Beaker’s edits with Takehi et al.’s larger edit reached 58.8%, versus 51.1% for Takehi et al.’s edit alone.
  • GPT-6.1 Sol performed comparably to GPT-6 Astra at much lower cost, with reported runs costing $4–16 versus $28–79.
  • With model reads, Sol reached 39.5% and Astra 34.8%; Beaker never generated answers on demand in these setups.

#weight-space programming #model editing #language models #qwen3.5

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Points

Score reported by the source.

  • Hacker News: 8, latest observation ; +3 over observed period since

Comments

Comment count reported by the source.

  • Hacker News: 1, latest observation ; 0 over observed period since

Aliases

SourceExternal IDURL
Hacker News50459411https://www.withbeaker.ai/blog/weight-space-programming Read discussion