- Math score almost doubled.
This is a dev post classified by AGK Intelligence as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
- 2 neurons. - columns in layer 29. - 0.00012% of weights. - Math score almost doubled. Claude Opus 5.5 just performed targeted weight surgery on Qwen3.5-4B and boost performance. It identified the exact neurons causing reasoning loops and removed that direction editing only 0.00012% of the model. results: - Math accuracy: 59 to 92-94 up - Repetition loops: 22–25 to 0. - Zero human intervention. - No gradients, No training loop, Just 2 neuron columns edited by Opus 5.5 in 27 minutes for $1.03. https://x.com/zengjiajun_eth/status/2117409198686216574/photo/1
Posted by Md Ismail Šojal 🕷️ (64.7k followers) 1 h ago · 6 likes · 619 views · view the original post on X. Kept by the Dev Radar as AI dev tools.
More dev work like this
- TRENDING 📈 @jcrwhittington shares Latent Sparse Attention (LSA), an open research… — @thestartpod
- 𝗛𝗲𝗿𝗺𝗲𝘀 v0.22.7 just dropped. The 1 feature worth knowing: a lead bot runs the… — @BkashJosi
- Meet Qwen3.5-0.8B weight-space-programming: a tiny 0.8B model that learns by editing… — @HuggingModels
- i've shifted my work style pretty heavily in recent weeks to start using "review desk".… — @grichadev
- Running parallel agents? Stop copying the whole project for each one. — @DanKornas
- Full macOS running inside your terminal. — @0x0SojalSec
- Day 33/ — @theo
- Super interesting finding on optimizing agent harnesses. — @omarsar0
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.3k posts from 5k X accounts over the last 21 days, 9.2k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-03 23:59 UTC. Full method.