Mechinterp's unsung root: nearly all weight-editing work traces back to one VGEL post
max_paperclips · x · 2026-11-03
A circulating X thread claims that mechinterp research — including the whole "editing weights" line — is downstream of posts by voooooogel (VGEL). The linked image is her 2023 blog post I made a transformer by hand (no training!), where she manually assigned every weight of a GPT-2-style decoder-only transformer to predict a simple sequence — no training, no pretrained weights — and it worked.
The original poster jokes that if VGEL had been a doomer, her impact score would exceed Neel Nanda's and Callum Hubinger's combined.
More from Fun
- X adds Promote button in composer, letting users pay to boost posts at publish time — tekbog · 2026-11-03
- GPU sellers push buying, inference sellers push agent swarms, banks just ban AI — tekbog · 2026-11-03
- Sloptoberfest community voting opens, winners announced November 7 — voooooogel · 2026-11-03
- Google Antigravity's Argon model hits repeated outages, user jokes it explains the stock drop — AIandDesign · 2026-11-03
- Can you name all three? An ML architecture guessing quiz from easy to hard — cneuralnetwork · 2026-11-03
- Beff jokes Palantir will ship a frontier-topping open source model in weeks — beffjezos · 2026-11-03