Hilton specifies noise test: N(0, 0.0025²) weights typically cost 1-10 accuracy points
JacobHHilton · x · 2026-10-18
Following up on his mechanistic faithfulness proposal, Jacob Hilton gives concrete parameters: adding N(0, 0.0025²) noise to the weights is a good test point — this level of noise typically degrades accuracy by 1-10 percentage points — and other noise levels would also be informative for checking whether a hand-built model can mechanistically predict accuracy degradation.
Related event: Researcher Hand-Rebuilds 16-Neuron Network, Beating Original Accuracy(3 posts)→
More from Research
- AuroraASR-1M: A 1MB Speech Recognition Model Running 436x Real-Time on One CPU Thread — Tall_Abrocoma_3533 · 2026-10-18
- Lattice Gaussian distributions: closed-form KL divergence via Riemann theta functions — FrnkNlsn · 2026-10-18
- Megabake compiler hits 5.5x torch.compile speedup on SmolLM — ricklamers · 2026-10-18
- DeepMind's AlphaProtein Novo designs enzymes for chemistry nature never evolved — anthara_ai · 2026-10-18
- mlabonne's llm-course passes 83k GitHub stars with roadmaps and Colab notebooks — mlabonne · 2026-10-18
- Kaggle LLM attribution contest: GLM text is flagged as Claude 44% of the time — HarveenChadha · 2026-10-18