Blogger questions weight-space programming: just frontier models recalling old papers
builderjaydub · x · 2026-11-09
Builder builderjaydW questions the novelty of "weight-space programming", suggesting the apparent gains may just come from frontier models being good at recalling well-established methods from past papers. The tweet cites the ICLR 2023 paper "Mass-Editing Memory in a Transformer" (MEMIT, MIT CSAIL/Northeastern), which scaled direct knowledge editing to thousands of associations on GPT-J (6B) and GPT-NeoX (20B), hinting that some current weight-intervention work is repackaging such prior research.
More from Research
- Delta Intelligence shows single generalist policy driving 69-DoF humanoid at near-human speed — chris_j_paxton · 2026-11-09
- MaxxBench criticized as cargo-cult science: overfitting confounds claimed benchmaxxing — teortaxesTex · 2026-11-09
- VPR thread: when retrieval is strong enough, image matching re-ranking hurts — gabriberton · 2026-11-09
- VisionSIM, a Blender-based multi-sensor robot perception simulator, lands CoRL 2026 spotlight — CSProfKGD · 2026-11-09
- Theory framework explains when prover-conjecturer self-play lets LLMs self-improve at theorem proving — stanfordnlp · 2026-11-09
- MIT CSAIL: GPT-6 Astra edits 0.016% of Qwen's weights, lifts MATH-500 from 7.9% to 48.1% — winer666 · 2026-11-09