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 - Oct 30 - From ASI in a Decade to Home Humanoids: MiniMax M2's Speed Demon, OpenAI's Bold Roadmap, and 2026 Robot Revolution

·1 hr 37 min·5 clips

Skyler says MiniMax M2 is top-five on leaderboards, with 8% of Sonnet's price and twice the speed.

1. ThursdAI's October 30 Halloween edition centers on MiniMax M2, OpenAI's roadmap, Cursor 2.0, and home humanoids. 2. Alex Volkov hosts with Ryan and Skyler from MiniMax; Alex says he is an AI evangelist with Weights & Biases, Ryan covers agent tools, and Skyler leads engineering at MiniMax. 3. The episode asks which AI releases from the past week matter most for builders, especially around open models, agents, and robotics. 4. Alex says MiniMax M2 is the week's biggest open-source story because it is fast, cheap, and positioned as an open-weight model for agent work. 5. Skyler says M2 reached the top five on leaderboards and is around 8% of Sonnet's price with roughly twice the speed. 6. Skyler says M2 uses an MOE design with 200 billion parameters and 10 billion active parameters. 7. Skyler says the team has trained MOE architectures for almost two years and expects M2.1 in a few months. 8. Skyler says MiniMax reduced context to 200,000 tokens for performance first, while planning to return to 1 million context later. 9. The show also covers OpenAI's GPT OSS safeguard models, which Alex describes as policy-focused safety reasoners with 120 billion and 20 billion parameter versions. 10. The episode adds Kimi Linear 5.1, IBM Granite 4.0 Nano, and MiniMax's video and speech updates as shorter open-source and product follow-ups. 11. Alex calls M2 a response to the "impossible triangle" of performance, cost, and speed. 12. Skyler says MiniMax wants "intelligence with everyone," which explains the push for affordable access. 13. Skyler says agent tasks often take 20 or 30 rounds, so low TPS hurts real workflows even when benchmark scores look good. 14. Alex notes that M2 is practical to run on a single H200 in 4-bit or 5-bit precision, which matters for deployment. 15. Skyler says interleaved thinking support matters for agentic performance and that some apps still do not pass those tokens correctly. 16. Ryan says agent labs are now choosing different models for different sub-agents instead of relying on one default model selector. 17. The interview format is fast and technical, with Alex interrupting for clarifications and Skyler answering directly from the engineering team. 18. The overall tone is news-driven, with a live show structure that moves quickly from model releases to robotics and product demos. 19. Builders, model tinkerers, and AI infra teams will get the most from this episode. 20. Casual listeners who want one narrow AI topic may skip the dense release roundup.

As heard by us

A fast Halloween week recap that still finds the signal in the noise.

Thursday Eye turns a noisy Halloween week into a brisk recap of AI developments, moving from open source to OpenAI restructuring, then on to fast-moving agents and humanoid robots in the home.

Read the full review in PlayNext →

Why you'd press play

If you want this week's AI chaos decoded at full speed, press play.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

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