ThursdAI - The top AI news from the past week · From Weights & Biases, Join AI Evangelist Alex Volkov and a panel of experts to cover everything important that happened in the world of AI from the past week

📆 ThursdAI - Dec 11 - GPT 5.2 is HERE! Plus, LLMs in Space, MCP donated, Devstral surprises and more AI news!

·1 hr 37 min·5 clips

GPT-5.2 scores 100% on AIME 2025 and jumps to 52.9% on ArcAGI2.

1. ThursdAI's December 11 episode focuses on GPT-5.2, open-source model releases, space GPUs, and Google research agents. 2. Alex Volkov hosts the show with Wolfram, LJ, Nistan, Tahira, and LDJ, and their roles matter because they surface benchmarks, tooling, and live product updates. 3. The episode asks what changed this week in AI and whether GPT-5.2, Mistral Devstral, and Google's new agent APIs justify the attention. 4. OpenAI launched GPT-5.2 in the ChatGPT app and API, including GPT-5.2 Pro, and Alex says it is state-of-the-art across many benchmarks. 5. GPT-5.2 scored 100% on AIME 2025, 52.9% on ArcAGI2, and 70.9% on GDP Eval, which Alex ties to math, general reasoning, and economically useful work. 6. The episode quotes early testers like Ethan Mollick, Matt Schumer, Aaron Levie, Greg Kamrad, and Dan Shipper, who gave mixed but detailed reactions. 7. Mistral's Devstral and Mistral Vibe CLI drew repeated praise because listeners described them as strong small models that can run at home. 8. Essential AI's RNJ1 was introduced as a Frontier 8B model led by Ashish Vaswani, with claims about coding, agentic tasks, math, and STEM. 9. StarCloud's H100 satellite was discussed as the first training and inference setup for an LLM in space, using Andrej Karpathy's NanoGPT and a Shutterstock-like demo of SSH into orbit. 10. The hosts also covered GINA VLM, Z.ai's GLM 4.6 vision release, Perceptron's Isaac 0.2, and the Putnam math model race. 11. GPT-5.2 stood out because it hit a perfect AIME score and a large jump on ArcAGI2, which made the launch feel more consequential than a routine update. 12. The GDP Eval number mattered because Alex described it as a 1,300-task benchmark for legal briefs and financial analysis, not just abstract model tests. 13. Greg Kamrad's comment about the 'late stage of the S-curve' introduced a note of caution about how much room these gains may leave. 14. Dan Shipper's view that GPT-5.2 was 'mostly incremental' for day-to-day use added a second, more restrained reading of the release. 15. The Mistral discussion emphasized that open-source models are becoming practical for home use, not just research demos, especially on a 3090 or a high-end laptop. 16. The satellite segment argued that power, cooling, and land scarcity could eventually push compute off Earth, even as LDJ and Alex questioned GPU replacement and maintenance in orbit. 17. The tone is fast, live, and comment-driven, with frequent interruptions, side notes, and audience reactions from Twitter Spaces. 18. The format shifts between breaking-news recap, benchmark debate, and open-thread discussion rather than a single structured interview. 19. People tracking GPT-5.2, open-source models, and AI infrastructure. 20. Listeners wanting a single topic-focused deep dive without news overload.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from ThursdAI - The top AI news from the past week