Model Mayhem: OpenAI’s 5.6 and Meta’s Muse Spark 1.1 | Diet TBPN

What you'll learn

• 1 (00:05) **Model Mayhem: A Wave of New AI Models** - The hosts introduce a flurry of new model releases, including xAI's Grok 4.5, Meta's Muse Spark, and OpenAI's GPT-5.6, framing the "slow summer" as anything but for the AI race.
• 2 (02:58) **GPT-5.6's Arc AGI Score: A Massive Leap** - The hosts analyze GPT-5.6's performance on the Arc AGI V3 benchmark, scoring 7.78%, a huge jump from GPT-4.8's 1.5%, signaling progress in generalization and spatial reasoning.
• 3 (04:26) **GPT-5.6 Launch Games: Fun and Functional** - The hosts discuss the interactive mini-games included in the GPT-5.6 blog post, such as a sailing game, which demonstrate the model's ability to create polished, vibe-coded software.
• 4 (06:37) **The Future of Entertainment is Interactive** - The hosts explore how AI is enabling the creation of mini-games and simulators, shifting from static memes to interactive experiences that can be distributed on platforms like Steam.
• 5 (08:05) **GPT-5.6 Cracks a Magic Trick: A Test of AGI** - The hosts share a story from Stanley Tang of DoorDash, who claimed GPT-5.6 was the first model to figure out a bulletproof magic trick, framing it as a milestone for first-principles reasoning.
• 6 (09:06) **Pareto Frontier: Fable vs. GPT-5.6** - The hosts discuss a comparison between Fable and GPT-5.6, noting that while GPT-5.6 is faster and more affordable, Fable still excels on the hardest problems, illustrating a healthy competitive landscape.
• 7 (10:38) **AI 2040 and the Meaning of Model Numbers** - The hosts briefly discuss the release of "AI 2040," a sequel to "AI 2027," which advocates for a slowdown in AI development, and then delve into the confusion around model numbering.

+ Full timestamped outline available in the app