AI Safety through Programming?
summary
Argues that programming ideas can make LLMs safer by: a tiny if-then rule language for domain specs (bank transfers, Singapore traffic and MRT rules) that triggers runtime checks; a layer-removal causal scan that flags hallucinations and jailbreaks; runtime verification proofs for simple algebraic invariants; experiments showing rule checks can stop law-violating outputs and that layer scans detect false answers, at a cost of roughly doubling inference time on Llama 2 but with cheaper one-pass consistency variants.
AI-written summary · about
chapters (8)
- 0:00Welcome
- 3:15Programming versus large language models
- 12:37AI safety challenges
- 22:51Domain-specific specs
- 31:56Runtime verification
- 41:51Incompleteness problem
- 50:22Conclusion
- 51:03Q&A
related talks
47mHarnessing LLMs for Next-Level Security Analysis
01h 29mLLM Safety, Alignment, and Generalization
32mSafer Code with LLMs: Scala Types Silver Bullet?
33mVerifiably Safe Agentic Compute
26mHow Bad Actors AbUse LLMs The Mechanisms That Enable Misuse and What We Can Do About
40mSafe Coding with LLMs - Verifiable Scala Silver Bullet
33mCorrectness Too Cheap To Meter: Formal Verification and LLMs
29mBeyond the Hype: Building Trustworthy and Reliable LLM Applications with Guardrails