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 - Mar 26 - AGI here? Gemini Flash Live, Turbo Quant, RIP SORA, ARC-AGI-3, 4 voice AI drops & more AI news

·1 hr 40 min·5 clips

Jensen says that for one broad definition of AGI, NVIDIA has it today.

1. ThursdAI on March 26 focuses on the week’s AI news, with Alex Volkov framing the episode around breaking updates already arriving that morning. 2. Alex Volkov hosts with LDJ, Wolf from Ravenwolf, Yam, and Daniel Han from Onsloth, who matters here because he explains Onsloth Studio and quantization work. 3. The episode asks what changed in AI workflows, what TurboQuant means for KV cache, and whether Jensen Huang’s AGI comment changes the conversation. 4. Alex says cloud code can now control a Mac, with computer-use features that click buttons on the user’s machine. 5. The hosts note reports that cloud code users are burning through weekly quotas quickly after asking only a few questions. 6. OpenAI’s Sora app, Sora API, and Sora ChatGPT video features are described as discontinued. 7. ARC-AGI-3 is discussed as a public benchmark where humans get 100% on the tasks and frontier labs are under 1%. 8. Google Research’s TurboQuant is presented as a KV-cache compression method with a claimed 6x compression and 8x speedup. 9. The show’s tools segment also covers Modular, Mojo, and a new Flux 2 optimization that runs under 300 milliseconds. 10. The hosts mention a supply-chain attack in the LiteLLM Python package and say SSH keys and cloud credentials were exfiltrated. 11. The voice block is the biggest release cluster, with Mistral Voxtral TTS, Cohere Transcribe, Google Gemini 3.1 Flash Live, Irodori TTS, and Tencent Kovo all named. 12. Google also launched Lyria 3 Pro for full three-minute music tracks with structural control. 13. Alex says Gemini 3.1 Flash Live is intended for real-time voice agents and vision agents without separate TTS and LLM components. 14. Daniel Han explains Onsloth Studio as a UI for fine tuning, inference, code execution, web search, and tool calling. 15. Daniel says Onsloth supports Weights & Biases logging and can sync training runs in real time. 16. Daniel describes quantization as rounding model numbers into fewer bits and says the goal is to reduce size with minimal accuracy loss. 17. LDJ argues TurboQuant is not a fresh invention and says the underlying paper was published a year ago. 18. Yam says every compression method involves trade-offs and no single paper removes the need for RAM or changes the limits of inference. 19. The tone is fast, technical, and conversational, with Alex steering the show through news, demos, and live reactions. 20. Open-source AI users, agent builders, and model hackers would get the most from this episode, while listeners wanting a nontechnical recap may skip it.

As heard by us

A crowded AI news roundtable that makes inference efficiency the main event.

Thursday Eye moves briskly through a dense stretch of AI updates, then pauses long enough to make TurboQuant and KVCache compression feel concrete instead of hand-wavy.

Read the full review in PlayNext →

Why you'd press play

Press play for a loud AI catch-up on TurboQuant, KVCache, and the week’s market whiplash.

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