AI index steepens 5x after late 2024 as compute shifts from pretraining to inference
ProfBuehlerMIT · x · 2026-08-04
AI performance curve steepens sharply after late 2024, suggesting a shift in where compute goes
The post argues that the ArtificialAnlysis Intelligence Index versus Kurzweil-style compute-per-$1k metrics shows a clear breakpoint in late 2024. After that point, the slope becomes almost 5× steeper.
- The key claim is that the underlying hardware price-performance curve itself did not change materially.
- What changed was how compute is being spent: pretraining alone is giving way to post-training, RL, and inference.
- Because the metric only tracks dollars per computation, this shift is invisible on the hardware axis and only shows up as a sharper AI capability curve.
- The post frames this as a new kind of substrate transition: not new hardware, but a new locus of scaling.
More from Infra
- chapter-tgz adds chapter boundaries to tar.gz for O(1) skipping and parallel reads — charliermarsh · 2026-08-04
- Community-built MiniMax H3 weights target 12GB–24GB GPUs with INT4 and NVFP4 variants — VoidAsuka · 2026-08-04
- Used R940 and two RTX 3090s run DeepSeek V4-Flash at 33 tok/s — AbbreviationsSad5582 · 2026-08-04
- NVIDIA says long-context serving speed is set by architecture choices before training — NVIDIAAI · 2026-08-04
- MiniMax H3 full bf16 run hits 95 GB VRAM and finishes a 15s clip in 2 hours — Moarkush · 2026-08-04
- Windows reset restores RTX 3070 throughput for local Qwen3.6-35B inference — campaigner_ · 2026-08-04