ARC-AGI-3 上 99.9% 与 62.7% 两个分数之争
播客讨论黄仁勋「AGI 已至」的说法：Astra 在自家口径拿 99.9%，在厂商中立测试框架上只有 62.7%，差别在于是否保留推理状态。
🚨 Episode 90 of The Intelligence Journal is here. 99.9% or 62.7%: The Fine Print Behind "AGI Has Arrived" 🧾⚡ @JensenHuang posted three words that moved markets: AGI has arrived. He credited GPT-6 Astra, trained on over 100,000 Grace Blackwell systems, with 400,000 more GPUs coming online next. The headline number is 99.9% on ARC-AGI-3. The number nobody put in the headline is 62.7%, which is what Astra scored on the provider-neutral harness built for fair comparison. Same model, same benchmark, two very different stories 🔍 Join us as we break down: - What separates the two scores, and why preserving opaque reasoning state between requests is the whole ballgame. - Why ARC Prize themselves say saturating this benchmark would not be proof of AGI, and why the people who built the test are more cautious than the people selling the GPUs. - What it means that OpenAI never claimed AGI, and the loudest declaration came from the company that sells the compute. 🗓️ Date: Sep 7, 2026 ⏰ Time: 3:00 PM UTC 📍 Set your reminder: https://x.com/i/spaces/1DGleVopppOJL?s=20