Anthropic S-1 data: training compute share rebounds to 73%, sparking alignment-ceiling worry
robleclerc · x · 2026-10-27
Quoted data from Anthropic's S-1 shows the training vs inference compute spend split: 94%/6% in 2023, 70%/30% in 2024, inference peaking in H1 2025 (56%/44%), then back to 73%/27% in H1 2026. The author infers that if inference efficiencies keep compounding faster than training gains, the terminal state of compute demand may be "almost all training" — contrary to the consensus all-inference view.
Rob LeClerc adds an alignment worry: are models becoming harder to align? If so, an "alignment ceiling" could arrive before AGI, confining the most advanced models to highly contained environments; he also claims OpenAI had to redo 6.1 over alignment issues.
Related event: Anthropic S-1 Reveals Compute Split and Training Infrastructure Edge(2 posts)→
More from AGI Musings
- Apollo chief economist: AI has delivered zero financial payoff outside Big Tech — rseroter · 2026-10-27
- OpenAI acquires Orchestra Bio to bring clinical-trial monitoring agents into drug development — dl_weekly · 2026-10-27
- Experts Clash: Today's Models Dwarf GPT-4, But Journal Delays Lock Out New LLMs — DeryaTR_ · 2026-10-27
- Venkatesh Rao essay: On Learning to Ride Your Brain — jger15 · 2026-10-27
- Veteran engineer on what the AI-coding generation will never feel: 36-hour sprints and elusive bugs — bendee983 · 2026-10-27
- AI systems are not, and will never be, our friends, argues AI ethicist — LuizaJarovsky · 2026-10-27