Just sitting and reflecting on this GPT-6 Astra graph. GPT-6 Astra set a new SOTA on ARC-AGI. 🧩 ARC-AGI-1 → 98.5% 🧠 ARC-AGI-2 → 95.0% 🤖 ARC-AGI-3 → 99.95% But the ARC-AGI-3 result has a weird twist. ARC-AGI-3 drops Astra into a completely NEW interactive environment with no instructions. It has to ... 👀 explore 🧠 figure out the rules 🗺️ build a world model 📌 remember what it learned 🎯 develop a strategy 🔄 adapt when it's wrong ARC tested ~500 humans to establish how efficiently people solve these environments. GPT-6 Astra with ARC's standard agent harness: 62.7% Astra using the new Provider Adapter that preserves reasoning state between requests + compacts it: 99.95% 🤯 Give a powerful model a persistent working state + good compaction, and suddenly the same intelligence can behave VERY differently over a long task. That's going to matter for Local AI too.