đ Learn AI With Me For Free – https://www.skool.com/the-aigrid-community-1726
đSubscribe To My Newsletter – https://aigrid.beehiiv.com/subscribe
Get your Free AGI Preparedness Guide – https://theaigrid.kit.com/agi
đ€ Follow Me on Twitter https://twitter.com/TheAiGrid
GPT-6.2 Astra launched to some of the harshest reviews of any OpenAI model. Theo called it the worst model OpenAI has ever released, Matt Shumer said it was essentially unusable, and Every called it a disappointing release. But the same model just became the first to score 100% on ARC-AGI-3, sits at the top of Epoch’s Capabilities Index, and is the first model OpenAI has released at the High threshold for AI self-improvement. In this video: why there’s "AI smart" and "human smart", what OpenAI admitted in the GPT-6.2 system card (blocker deception, broken search tool, evaluation awareness), why reinforcement learning keeps producing brilliant models that don’t listen, and why you should test new models on your own benchmarks instead of trusting launch-day reviews.
â±ïž Chapters
0:00 Intro
1:02 The launch everyone hated
5:09 GPT-6.2 Astra is genuinely smart
10:56 AI smart vs human smart
16:07 What OpenAI admitted in the system card
20:30 Why this keeps happening
26:57 Test it on your own benchmarks
31:14 The pace is still insane
đ„ Video clips courtesy of: OpenAI (GPT-6.2 Astra launch films and demos, Building 3D video sets with GPT-6.2 Astra, GPT-6 Astra films, Codex), DesignArena (3D Design comparison), Sequoia Capital (Andrej Karpathy: From Vibe Coding to Agentic Engineering: https://www.youtube.com/watch?v=96jN2OCOfLs), Dwarkesh Patel (Andrej Karpathy: https://www.youtube.com/watch?v=lXUZvyajciY; Ilya Sutskever: https://www.youtube.com/watch?v=aR20FWCCjAs), Anthropic, Google
đŠ Posts featured: @theo, @mattshumer_, @danshipper, @deedydas, @DaveShapi, @mckaywrigley, @koltregaskes, @signulll, @ValsAI, @voxelbench, @miu21590, @ZryMiller, @arcprize, @EpochAIResearch, @JoshPurtell, @HistoryGPT, @TheRealAdamG, @AlexFinn, @beffjezos, @mweinbach, @AdamHoltererer, @scaling01, @thsottiaux, @Conor_D_Dart, @iruletheworldmo
đ° Sources
1. Introducing GPT-6.2 Astra and Sol (OpenAI): https://openai.com/index/introducing-gpt-6-2-astra-sol/
2. GPT-6.2 Astra and Sol System Card (OpenAI): https://deploymentsafety.openai.com/gpt-6-2-astra-sol
3. Theo, "Why I hate GPT-6.2 Astra" (X): https://x.com/theo
4. Matt Shumer on GPT-6.2 Astra / Sol (X): https://x.com/mattshumer_
5. Dan Shipper / Every on Astra 6.2 and Sol 6.2 (X): https://x.com/danshipper
6. ARC Prize, GPT-6.2 Astra on ARC-AGI (Verified) (X): https://x.com/arcprize
7. Epoch AI, GPT-6.2 Astra on the Epoch Capabilities Index and serial reasoning steps (X): https://x.com/EpochAIResearch
8. Epoch Capabilities Index (Epoch AI): https://epoch.ai/eci
9. Vals Index, GPT-6.2 Astra and Sol (X): https://x.com/ValsAI
10. VoxelBench on GPT-6.2 error rates (X): https://x.com/voxelbench
11. vechen, "Astra 6.2 in Codex is 4.7x slower than Astra 6" (X): https://x.com/miu21590
12. Mark Humphries on early access to GPT-6.2 Astra (X): https://x.com/HistoryGPT
13. Josh Purtell on the autoresearch benchmark (X): https://x.com/JoshPurtell
14. Kol Tregaskes on putting the GPT-6.2 releases into perspective (X): https://x.com/koltregaskes
15. Max Weinbach on using the GPT-6.2 models (X): https://x.com/mweinbach
16. Adam Holter, GPT-6 Astra to GPT-6.2 Astra behaviour evals chart (X): https://x.com/AdamHoltererer
17. Lisan al Gaib on evaluation awareness and CoT monitoring (X): https://x.com/scaling01
18. Tibo on the GPT-6.2 models being polarizing (X): https://x.com/thsottiaux
19. signĂŒll on the negative reactions (X): https://x.com/signulll
20. Conor Dart, "Ignore the negative reviews until YOU have tested it yourself" (X): https://x.com/Conor_D_Dart
21. Alex Finn benchmark, via Adam.GPT (X): https://x.com/TheRealAdamG
22. DesignArena 3D Design comparison: https://www.designarena.ai
23. Andrej Karpathy on verifiability and jagged intelligence (Sequoia Capital): https://www.youtube.com/watch?v=96jN2OCOfLs
24. Andrej Karpathy on reinforcement learning (Dwarkesh Patel): https://www.youtube.com/watch?v=lXUZvyajciY
25. Ilya Sutskever on the gap between evals and real-world performance (Dwarkesh Patel): https://www.youtube.com/watch?v=aR20FWCCjAs

