AI safety researchers propose VCAD: audited compute disclosures plus reallocating compute away from capabilities R&D
thlarsen · x · 2026-11-18
AI safety researcher thlarsen released a concrete near-term AI slowdown proposal:
- Verified Compute Allocation Disclosures (VCAD): AI companies disclose how they use their compute, with third-party auditors verifying the claims, making slowdowns credible.
- Compute Reallocation: shift compute from capabilities R&D toward serving customers and safety work, which actually slows capability progress.
The author argues the current 'safety case pacing' approach fails: there's no consensus standard for a valid safety case (so no consistent auditing), and delaying a training run without pausing R&D barely changes the capabilities trajectory. Compute reallocation is specific, verifiable, and more durable. Written quickly given AIFP's recent timelines update, with more analysis to follow.
More from Safety
- Boardroom meme on agentic attacks: the 'just improve cybersecurity' guy gets thrown out the window — kingslyj · 2026-11-18
- Follow-up on the VCAD AI slowdown proposal: written quickly amid urgency, more analysis coming — thlarsen · 2026-11-18
- KVM guest-to-host escape CVE-2026-99312 lets cloud VMs read 127GB of host RAM — kingslyj · 2026-11-18
- nginx2root: 5 nginx bugs chained from unauthenticated visitor to root, patch now — evilsocket · 2026-11-18
- Nvidia halts shipments of US-approved AI chips to China, Reuters reports — minchoi · 2026-11-18
- Pinker, HF co-founder Wolf join rare cross-camp statement slamming Anthropic's safety branding ahead of IPO — wfithian · 2026-11-18