Training from scratch on a single H100 hits 76% on ARC-AGI-1 in ~4 hours
GregKamradt · x · 2026-09-15
Developer kschweig shares that a simple autoregressive transformer with a few twists, trained from scratch and running inference in a bit more than 4 hours on a single H100, scores 76% on ARC-AGI-1. A throughput-optimized recipe reaches 44% (TRM-level performance) in just 17 minutes. Full details are in the attached thread.
More from Infra
- Leaker Claims Nvidia RTX Rubin 6090 Launching Next Year — max_paperclips · 2026-09-15
- Pareta routes cheap LLM tasks to small models, 620x cheaper than GPT-5.5 — D33B · 2026-09-15
- Prefill and Decode: why asking an LLM for three takeaways from a long document still takes minutes — dotey · 2026-09-15
- From Ollama to vLLM: a roadmap for scaling LLM deployment — kalyan_kpl · 2026-09-15
- Kimi K3 is live and free on NVIDIA NIM with OpenAI-compatible API — airesearch12 · 2026-09-15
- Dual Radeon AI Pro R9700 vs. Two Used RTX 3090s at $1600 Each for Local LLM Inference — Current-Ticket4214 · 2026-09-15