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

GPT‑5.1’s New Brain, Grok’s 2M Context, Omnilingual ASR, and a Terminal UI That Sparks Joy

·1 hr 10 min·3 clips

Paul says Scribe v2 Real-Time transcribes speech in about 150 milliseconds across 90 languages.

1. ThursdAI on November 13, 2025 covers open-source AI, company releases, and two live demos from ElevenLabs and W&B. 2. Alex hosts the episode, and guests Paul from ElevenLabs and Dima from W&B matter because they show Scribe v2 Real-Time and 1DB LEET directly. 3. The episode asks what changed in AI this week, from benchmarks and multimodal models to terminal-native tools and speech transcription. 4. Terminal Bench v2 is presented as a harder agent benchmark with 89 terminal-focused tasks. 5. The hosts say Warp, Code CLI, and GPT-5 are near the top, with scores around 50 and 49. 6. They compare Terminal Bench to saturated benchmarks like MMLU and to older, less reproducible systems such as OSWorld. 7. Baidu’s Ernie 4.5 VL is discussed as a 3 billion-parameter visual model with claims against Gemini 2.5 Pro and GPT-5 High. 8. The hosts inspect chart results for MathVista, MathVision, and MMMU rather than accepting the headline claim at face value. 9. Holo 2 from the Age company appears as a new multimodal family for grounding and reasoning across web, desktop, and mobile. 10. Project Ala, renamed around a Leon AI partnership, centers on roughly 100,000 scientific papers and a visualization tool. 11. The project also releases L-animatron 12B and L-equine 14B for structured JSON summaries of scientific research. 12. Weibo’s Vibe Thinker 1.5 is described as a 1.5 billion-parameter Quen fine-tune with strong math and code benchmarks. 13. The hosts question AIME 24 claims because they treat leaked competition sets as less trustworthy signals. 14. Paul from ElevenLabs explains that Scribe v2 Real-Time is the speech-to-text side of voice agents, alongside text-to-speech and LLM brains. 15. He gives a latency figure of about 150 milliseconds and says the system supports more than 90 languages. 16. His demo at captions.events shows live transcription and automatic switching into Dutch without manual intervention. 17. The tone shifts between benchmark skepticism, release coverage, and live product demos. 18. The style is fast-paced, panel-driven, and heavy on specific charts, links, and implementation details. 19. Best for listeners tracking weekly AI releases, evals, and developer tooling. 20. Skip if you want a slow, narrative story instead of dense product news.

As heard by us

A brisk AI tools review that works best when Grok4Fast, Leet, and Scribe 2 Live move from claims into demos.

Thursday Eye turns a crowded AI news week into a guided pass through practical demos and fast-moving claims. Terminal Bench is framed as a new bar for agent evals, GPT-5.1 gets attention for a warmer voice and personality, and Grok4Fast is presented with a 2 million token…

Read the full review in PlayNext →

Why you'd press play

Want to see AI headlines turn into live demos, including Dutch speech switching and an offline terminal dashboard?

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