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 - Opus 1M, Jensen declares OpenClaw as the new Linux, GPT 5.4 Mini & Nano, Minimax 2.7, Composer 2 & more AI news

·1 hr 32 min·5 clips

Jensen spends 10 minutes on OpenClaw and calls every company’s OpenClaw strategy essential.

1. Thursday AI on March 19 centers on Anthropic’s 1 million-token context window, OpenAI’s coding model releases, and NVIDIA’s OpenClaw comments. 2. Alex Volkov hosts from Colorado with Wolf from Ravenwolf and LDJ, and he identifies himself as an “AI adventurist with weights and biases.” 3. The episode is built to cover the week’s AI news, with a strong focus on long context, agent tooling, and model releases. 4. Anthropic makes the 1 million-token context window the default in Cloud Code, while the hosts note that auto-compaction still happens around 170,000 tokens. 5. OpenAI releases GPT-5.4 Mini and GPT-5.4 Nano for coding and computer use, and the team links those launches to Codex subagents. 6. NVIDIA CEO Jensen Huang spends about 10 minutes on OpenClaw at GTC, calling every company’s OpenClaw strategy important and introducing NemoClaw for enterprise security. 7. Minimax releases M2.7 as a follow-up to MX2.5, and the hosts describe it as a self-evolving model with 56% on SweepBench Pro. 8. Google’s new vibe-coding experience appears in Google AI Studio, and Logan Kilpatrick is named as the source of the announcement. 9. Mistral releases Mistral Small 4 with 119 billion total parameters, 6 billion active per token, multimodal support, and Apache 2 licensing. 10. H company releases Holotron 12B, which the hosts describe as an open-source hybrid SSM for computer-use agents with 8.9 thousand tokens per second and 35% to 80% WebVoyager gains. 11. Wolfram says Mistral Small 4 is the worst model in his current OpenClaw benchmark list, averaging about 17%. 12. Wolfram also explains that Terminal Bench 2.0 agent tasks include things like configuring a Git server or training a model inside a sandbox. 13. Onslaught Studio is presented as a web UI for local model serving, training, and inference, with claims of 2x speed and 70% less RAM. 14. The hosts highlight that Onslaught Studio supports up to 500 models, including text, vision, audio, embeddings, and self-healing tool calling. 15. OpenAI announces an agreement to acquire Astral, the company behind Ruff and UV, and the Astral team joins the Codex team. 16. The hosts compare that move with Anthropic’s Bun acquisition and debate whether big labs owning key developer tools helps or hurts the ecosystem. 17. The tone is news-heavy and collaborative, with multiple speakers trading quick reactions and benchmark context across segments. 18. The format is a fast roundup with digressions into live benchmark discussion, tool demos, and platform announcements. 19. People who track AI model releases, coding agents, and open-source tooling will get the most value from this episode. 20. People looking for a single-topic deep dive or a lightly edited interview 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