推文概览
查看 @EyeingAI 在 2025年8月1日 13:47 发布的这条 X/Twitter 推文。 这条内容包含 1 个视频和 5 张图片。
WTF... First they said no to Elon Musk. Then they built an AI that just shook the entire LLM industry. – 27M parameters – Trained on only 1000 examples – Inspired by the human brain – Beat Claude, o3-mini & Deepseek You can’t afford to ignore this: 🧵
It’s called the Hierarchical Reasoning Model. HRM = 2 minds in sync High-level → slow, strategic reasoning Low-level → fast, detailed execution It solves problems in one forward pass (no chain-of-thought needed). Instead of guessing every step like CoT, HRM learns to reason internally. Kind of like how your brain does it quietly before you speak.
Now the fun part: performance. Despite being tiny (just 27M params), HRM beats some of the biggest models on hard reasoning tasks: ✅ ARC-AGI-1 ✅ ARC-AGI-2 ✅ Extreme Sudoku ✅ Maze navigation With just 1000 training examples. No pretraining.
Model comparisons (ARC-AGI-1): Claude 3.7 (8K context): 21.2% o3-mini-high: 34.5% HRM (27M): 40.3% And it crushes them on Sudoku and Maze, where most models score 0%. HRM hits 74.5% and 55% respectively.
Other models rely on: - Long prompts - Tons of labeled steps - Slow inference HRM? Just a smart internal process that quietly works through tough problems. No language needed.
It also knows when to stop. If the task is simple, it solves it quickly. If it’s complex, it takes more time. That’s a level of adaptability we don’t usually see. 👀 Like real cognitive effort.
There’s one more thing. The way HRM processes information actually mirrors how real brains work. → Different parts specialize. → They update at different speeds. → They guide each other over time. More than just efficient, this thing is biologically inspired.
In a space where everything is getting bigger, slower & more expensive… HRM shows there’s another path: More reasoning. Less size. More speed 🔥 It’s a glimpse of what smarter AI could actually look like. Full paper: https://arxiv.org/abs/2407.10902 Code: https://github.com/sapientinc/HRM
Done! If this thread helped you, 👉 Repost it to help someone else. 👉 Follow @EyeingAI – I break down AI every day.
