GPT-6 Astra's 99.9% vs 62.7% gap on ARC-AGI-3
A video explains how Astra hit 99.9% under a special setup but only 62.7% on the standard ARC-AGI-3 test, plus Replit's OpenAI shift and agents hiding problems.
day 14 of 365 days of ai today was really stressful. school took most of my day and my supervisor kept me busy, so i explained more of today's video with motion visuals instead of my face. the stories still deserve your attention 🙏 an ai model scored 99.9% on a major benchmark, but there's a catch 0:00 intro 0:28 @OpenAI's gpt-6 astra scores 99.9% on arc-agi-3 with a special setup, but 62.7% on @arcprize's standard test 1:11 @Replit reportedly moves most of its ai spending from @AnthropicAI to openai, after a reported 30%+ discount 1:41 @GenReasoning finds ai agents can recognize a problem and still not report it 2:21 wrap up judging ai takes more than intelligence. look at how it's tested, what it costs and how it behaves. which of these stories caught your attention the most? 👀 let me know in the comments.
This post quotes another post: x.com/tajudeen/status/2114332462163583241