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 - Feb 5 - Opus 4.6 was #1 for ONE HOUR before GPT 5.3 Codex, Voxtral transcription, Codex app, Qwen Coder Next & the Agentic Internet

·1 hr 38 min·5 clips

OpenAI's Codex app is a standalone Mac app for managing multiple coding agents in parallel.

1. ThursdAI on February 5, 2026 focuses on a week of AI news built around Opus 4.6, GPT 5.3 Codex, voice models, and the agentic internet. 2. Alex Volkov hosts the episode, and VB from OpenAI joins later because the panel wants first-hand context on the new Codex app and model release. 3. The episode asks what this week's releases mean for coding agents, transcription, video generation, and the security of tool-using AI systems. 4. Alex says Anthropic dropped Opus 4.6 and OpenAI released GPT 5.3 Codex roughly an hour later. 5. The panel compares those two models live on the show and says they were competing on the same benchmarks. 6. Alex describes OpenAI's Codex app as a standalone Mac OS command center for managing multiple coding agents in parallel. 7. VB is expected to talk about the Codex app, which Alex says had already reached 1 million active users and 600,000 downloads. 8. The show also mentions OpenAI's frontier agentic marketplace for enterprise agents working together in a safer workflow. 9. The open-source segment starts with ZAI's GLM OCR 0.9B, which the panel says ranks first on the OmniDoc benchmark. 10. Wolfram says GLM OCR handles formula recognition, table parsing, and information extraction, and is available through VLLM, SGLang, and Ollama. 11. The panel also covers Shanghai AI Lab's Intern S1 Pro, a 1 trillion parameter model for scientific reasoning across chemistry, biology, materials, and earth sciences. 12. Yam notes that the Intern S1 Pro benchmark chart shows a large gap versus models such as GPT-5.2 and Gemini 3 Pro on scientific tasks. 13. Step 3.5 Flash from StepFun is presented as a sparse MOE with 196 billion total parameters and 11 billion active. 14. The panel emphasizes StepFun's 100 to 350 tokens per second claim and compares its coding and reasoning scores to other open models. 15. Quen 3 Coder Next is described as an 80 billion MOE coding agent with 3 billion active parameters, 256k context, and YARN expansion to 1 million tokens. 16. LDJ says that the Quen model's smaller active footprint and context window make it a strong fit for local orchestration and larger coding workflows. 17. Alex spends time on the agentic internet theme and says OpenClaw has exploded, with more than 100,000 and then 160,000 GitHub stars mentioned in the discussion. 18. The tone is fast, live, and panel-driven, with Alex repeatedly steering between TLDR summaries, guest setup, and interruptions for demos. 19. Listeners interested in coding agents, model benchmarks, and AI tooling will get the most value. 20. Listeners who want a slow, single-topic interview will probably 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