TBPN: what AI ‘doomers’ actually propose for slowing down
TBPN ● Covered by 85 sources
AI ‘doomers’ now have a real slowdown plan, not just warnings. It’s a compute clampdown: audits, chip counts, and limits on frontier training.
Based on reporting by TBPN — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
The AI safety debate is getting more concrete. Instead of vague calls to “slow down,” Daniel Kokotajlo has put out a detailed proposal on ai-2040.com that tries to specify how a slowdown could actually be enforced, from data centers down to the chips inside them.
The core idea is an “AI Pause” on new frontier training and big AI R&D experiments. If a data center is running more than 10,000 H100 equivalents, the plan says it should be limited to inference on current models, with workloads checked by an independent auditor, likely the government. Major countries would also have to publish AI-compute inventories so people can see what chips exist, what’s being manufactured, and who controls it.
The proposal gets very hands-on. Major data center owners and semiconductor-supply-chain companies would need to hand over sales records. Foreign inspectors would do routine physical chip counts. Big chip transfers would only go to registered, auditable counterparties. On the networking side, the plan calls for removing high-bandwidth east-west connections inside data centers so large distributed training runs become much harder, while inference still works.
For new AI R&D sites, the essay imagines something closer to a state security project than a normal server farm: new facilities built from scratch, with airgapped communications, very limited personnel, and heavy verification. It also wants the outside research-information channel capped at 1 MB/s, so stealing giant model files becomes conspicuous or impossible.
Then there’s the weirdly practical stuff. Frontier weights moving between an R&D site and an inference site would be stored on physical devices encrypted separately by the U.S. and China, then escorted by representatives of both countries. The plan even favors making frontier models larger than compute-optimal, on the theory that a 100-TB file is simply harder to steal than a 1-TB one.
The broader goal is not to freeze AI forever. The proposal aims for gradual scaling toward top-human-expert capability around 2035, then a roughly five-year pause for alignment and control work, and only then a push toward superintelligence around 2040. That’s the real point here: not “stop,” but brake hard enough to buy time.
My take — AI-written commentary, not fact-checked reporting
The uncomfortable truth is that the AI safety crowd has moved past slogans and into infrastructure. That’s healthier than doom-posting, but it also means the fight is no longer about vibes; it’s about who gets to inspect the chips and write the rules. Once the argument turns into compute caps and auditor access, the industry starts sounding a lot less like a revolution and a lot more like regulated utilities with better branding.
Read more about this at: TBPN