Compute limits push training efficiency; China leans on distilling US models
trevposts · x · 2026-08-21
The author raises and partially answers questions about how compute constraints affect US and China AI progress: reduced compute supply adds even more incentive to make training runs compute-efficient. He also asks how much this affects China, given much of its current capability seems to come from distilling US models and possibly stealing weights now or in the future.
Related event: Thought Experiment: What If the US Paused Data Center Construction?(3 posts)→
More from AGI Musings
- Model Nuances Don't Matter as Posthuman Transition Approaches — danfaggella · 2026-08-21
- When Your Buyer Is an AI Agent: A Shift in B2B Commerce — rseroter · 2026-08-21
- Borretti essay: automating drudgery also automates meaningful work — same complexity class — zetalyrae · 2026-08-21
- nptacek: AI's most impactful parts live out-of-distribution — don't prune the weird corners of latent space — nptacek · 2026-08-21
- A post-work world may produce more weird, ambitious projects — VraserX · 2026-08-21
- Experts Call Zuckerberg's "AI for Everyone" Vision Laughable, Citing Overstated Data Center Benefits — ChinasaTOkolo · 2026-08-21