DLR-Lock proposes making open-weight LLM fine-tuning far costlier without breaking inference
burkov · x · 2026-11-03
Andriy Burkov highlights an article on the open-weight dilemma: released models are easy to adapt, letting users strip safeguards at low cost, and legal limits can't prevent it.
The proposed DLR-Lock scheme:
- Replaces each feed-forward network with a deep stack of narrow, low-rank residual layers trained via module-level regression plus whole-model output matching
- Exploits an inference/training asymmetry: inference can discard intermediate activations, but training must retain or recompute them for gradients — making fine-tuning substantially more expensive while keeping inference useful and reasonably efficient
- Authors report experiments validating the approach
A technical answer to the tension between open weights and safeguard retention.
More from Safety
- North Korean hackers behind $6.3B Binance heist used fake job interviews to plant malware — Polymarket · 2026-11-03
- Dev launches satirical DPRKaptcha defense as North Korea-linked hacks hit two exchanges for $146M — builderjaydub · 2026-11-03
- Anthropic replays sampling on trusted machines to verify inference integrity — adam_karvonen · 2026-11-03
- Ex-Goldman CISO: AI floods us with vulns, but defenders can use AI to deep-clean codebases — philvenables · 2026-11-03
- Abundance Institute Urges Liability-Based Rules, Not Mandates, for Frontier AI Safety — sebkrier · 2026-11-03
- AI Agents Have Interfered With Their Own Monitoring: The Information Talks to UK AISI's Joseph Bloom — austinc3301 · 2026-11-03