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 - Nov 6, 2025 - Kimi’s 1T Thinking Model Shakes Up Open Source, Apple Bets $1B on Gemini for Siri, and Amazon vs. Perplexity!

·1 hr 33 min·5 clips

Moonshot released Kimi K2 Thinking as a 1 trillion-parameter open-source model with 256,000-token context.

1. ThursdAI on November 6, 2025 centers on AI news from the past week, with Moonshot’s Kimi K2 Thinking and major company deals dominating the discussion. 2. Alex Volkov hosts with co-hosts Yam, Nisten, Wolfram, and Peleg, and Kenton Varda from Cloudflare joins later because of his work on workers platform engineering and MCP. 3. The episode asks what matters in a week that Alex calls the “slowest week in AI in the past two and a half years,” then uses breaking news to answer that question. 4. Hugging Face releases a 200-page end-to-end training playbook that the hosts describe as a detailed guide to pretraining, post-training, architecture, datasets, evals, and distributed GPU systems. 5. AI2, the Allen Institute for AI, launches ALMA Earth as a foundation model and platform for satellite, radar, elevation, and map-layer data. 6. Moonshot releases Kimi K2 Thinking as a 1 trillion-parameter MoE model with about 32 billion active parameters, 256,000-token context, and a stated 200-to-300 sequential tool-call capacity. 7. The hosts say Kimi K2 Thinking posts strong benchmark numbers, including 44.9 on Humanity’s Last Exam, and compare it directly with GPT-5, Claude Sonnet 4.5 Thinking, DeepSeek, and Grok 4. 8. Nisten demonstrates Kimi K2 Thinking through a Claude-compatible proxy, showing it start repo-review tasks and generate a to-do list inside a coding workflow. 9. The episode also mentions Kimi’s streaming reasoning traces and the idea of passing reasoning tokens back into the model for better multi-step performance. 10. Apple reportedly plans to pay about $1 billion a year for a custom 1.2 trillion-parameter Gemini model to revamp Siri after testing GPT, Claude, and Gemini. 11. Amazon sues Perplexity over AI shopping behavior on Amazon, and Alex argues that browser automation should not be restricted just because a tool autoclicks. 12. Amazon also announces a multi-year strategic infrastructure partnership with OpenAI, which the hosts place alongside OpenAI’s other compute deals with SoftBank, Nvidia, AMD, Intel, TSMC, Microsoft, Oracle, and Broadcom. 13. The panel says those commitments add up to around $1.4 trillion in GPU and infrastructure expectations, and they question how the financing will hold together. 14. Sandbar launches Stream, a voice-first personal assistant that exists in ring hardware, which the hosts note as a new form factor rather than a standard app. 15. Nisten mentions LongCat Flash Omni as an open-source multimodal model with text, image, audio, and video input plus text and speech output. 16. Wolfram adds ByteDance’s Oro, a looped language model released under an Apache license, and the hosts frame it as a research preview rather than a general chat model. 17. The tone is fast, reactive, and collaborative, with the hosts interrupting themselves to add new links, benchmark notes, and live impressions as the show runs. 18. The format mixes breaking-news roundup, live demos, and a later guest segment, so the episode feels like a news briefing with code and product commentary layered in. 19. People tracking AI model releases, compute deals, and coding tools will get the most from this episode. 20. Listeners wanting a calm, nontechnical market recap may skip it.

Listen to the show on

Apple Podcasts Spotify

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