In 2024 the largest single AI training run will use roughly 5x10^25 FLOP.
Accurate ●●●●●
“So, the most compute spent on a single training run is something like 5x10^25 FLOPs.”
- Said on
- 6 August 2021 · 2024 section of his year-by-year AI scenario post
- Source
- What 2026 looks like · LessWrong
- Should come true
- 2024
- How sure they sounded
- "something like"; stated in his best-guess scenario
- Categories
- AITechnology
- Importance
- 3 of 5 · Frontier compute scale is a key driver of AI progress
Outcome
Accurate Happened 31 December 2024.
Epoch AI estimates the largest runs through 2024 at roughly this scale: Gemini 1.0 Ultra ~5e25 FLOP (Dec 2023) and Llama 3.1 405B ~3.8e25 FLOP (Jul 2024). Larger ~1e26+ runs (e.g. Grok 3) were released only in 2025. The surrounding claim that labs stopped training new models was wrong, but the recorded compute figure held.
History
- 2026-10-17 07:27:18 · created by claude
- 2026-10-17 07:27:18 · categorized by claude
- 2026-10-17 07:28:09 · scored by claude