Source · Tier B · PreprintSource · What does it take to catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring
What does it take to catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring
Y. Shavit. 2023. arXiv.
| Original | https://arxiv.org/abs/2303.11341 |
|---|---|
| DOI | 10.48550/arXiv.2303.11341 |
| arXiv | 2303.11341 |
| Version | Read arXiv v2 (30 May 2023). |
| Accessed | 2026-10-20 |
| Imported from | hodgkins-ai-verification-papers@c71e59ff0e8e |
| Note | Listed under "Proof of learning and training" in the Hodgkins bibliography (CC BY 4.0). |
Cited by
- Cap frontier trainingGoal
- Training-transcript verification (proof-of-learning)Mechanism
- Communication between compute groups is boundedClaim
- Chips are where they are declared to beClaim
- Compute stock is at most a declared amountClaim
- This compute runs inference, not trainingClaim
- There is no undeclared relevant computeClaim
- A training run stayed within declared limitsClaim
- Evidence bindingConcept
- FLOP accountingConcept
- Inference and training workloadsConcept
- Numerical nondeterminismConcept
- Positive and negative claimsConcept
- ProverConcept
- RecomputationConcept
- Root of trustConcept
- Sampling and assuranceConcept
- Tamper evidence and tamper resistanceConcept
- Threat modelConcept
- Undeclared computeConcept
- VerifierConcept