OpenAI Launches Codex Subagents

#1 📝 Simon Willison

Use subagents and custom agents in Codex - OpenAI Codex announced general availability of subagents and support for custom agents, enabling patterns similar to Claude Code's subagents (explorer, worker, default) and TOML-defined custom agents. The post notes widespread platform support for subagents and provides links to documentation across multiple providers.

#2 📝 Simon Willison

Introducing Mistral Small 4 - Mistral released a new Apache-2 licensed 119B parameter Mixture-of-Experts model called Mistral Small 4 that unifies capabilities previously spread across their flagship models and supports selectable "reasoning_effort" modes. The model is available as a large download on Hugging Face and has been tested via the Mistral API.

#3 𝕏

Andrew Ng launched Context Hub (chub), an open‐source CLI tool that’s already racked up 6K GitHub stars and grown from under 100 to over 1,000 API documents thanks to community contributions and an agentic document writer.

#4 𝕏

Mistral AI announced a strategic partnership with NVIDIA to co-develop frontier open-source AI models, combining Mistral’s cutting-edge model architecture and full-stack AI offering with NVIDIA’s leading compute infrastructure and development tools.

#5 𝕏

NVIDIA AI has expanded its partnership with Google Cloud to co-engineer the core infrastructure foundation needed to power the next generation of agentic AI.

#6 𝕏

NVIDIA AI unveils the latest AI and accelerated computing breakthroughs in Jensen Huang’s live #NVIDIAGTC 2026 keynote.

#7 𝕏

Philipp Schmid wrote a developer guide for Nano Banana 2 with the Gemini Interactions API, walking through four use cases: text-to-image photorealistic Kyoto travel poster generation, Web Search grounding with real landmark facts, Image Search for accurate photos, and referen...

#8 𝕏

Logan Kilpatrick explains that the Gemini API offers two cost-control levers—global billing account caps to cap overall spend and user-set spend caps to limit individual usage—detailing how each works to manage billing.

#9 ▶️

Building iOS apps with Claude Code as a non-coder

How I AI Podcast

Also covered by: @Claire Vo

#10 ▶️

Can Claude Code Learn To Draw In MS PAINT?

All About AI

#11 𝕏

clem 🤗 reveals open-source AI can be trained for <$2K (text classification), <$7K (image embeddings), <$100K (OCR), or <$500K (machine translation) versus ~$300M for GPT-4.5, arguing that smaller, task-specific models often suffice over massive ones.

#12 𝕏

claire vo 🖤 assigns AI models to dev roles—Codex as senior engineer/spec writer, Devin as implementer, Bugbot for QA, Cursor+Opus for design/PM, and CC as a versatile utility player.

#13 𝕏

Santiago launched PixVerse R1, a real-time world model that streams infinite, continuous visuals and adapts instantly to user input instead of generating fixed clips.

#14 𝕏

Aravind Srinivas announces that Computer can now use the local browser Comet as a tool, enabling it to perform any task without connectors or MCPs. He points out this capability gives Computer a unique market-leading advantage unmatched by any other tool.

#15 𝕏

Peter Yang says PMs must write specs for AI agents rather than engineers and rapidly master core AI skills or risk obsolescence. He even proposes token spend should eclipse salaries and warns that waterfall methodologies won’t survive the AI revolution.

#16 𝕏

OpenAI’s Head of Health Dr. Nate Gross and Health AI Research Lead Karan Singhal join Andrew Mayne to unveil new AI models and products aimed at solving concrete healthcare challenges for patients and doctors.

#17 𝕏

Yann LeCun highlights FAIR’s flagship “embed the world” project—initially deployed internally as Filament—which was later open-sourced as PyTorch-BigGraph to pioneer universal embeddings using more primitive techniques than today.

#18 𝕏

Yann LeCun says his 2023 Model S with Full Self-Driving is handy but officially only Level 2 autonomy—nowhere near Level 5. He points to https://motherfrunker.ca/fsd/ for a deeper dive.

#19 𝕏

Mustafa Suleyman says Microsoft Copilot Health didn’t diagnose him but guided him to request a critical medical test no doctor had ordered in 20 years, highlighting AI’s potential to transform patient care.

#20 𝕏

Rowan Cheung reports that Dr. Prokar Dasgupta in London remotely operated a 4-armed surgical robot with a 3D camera over fiber-optic (plus 5G backup) to perform a prostatectomy 1,500 miles away in Spain with just 0.06 s lag.

#21 in

Marily Nika, Ph.D warns that a rogue Chipotle burrito-bot demo exposed how AI products fail without steering guardrails. She’s teaming with Aman Khan and Tal Raviv for live OpenClaw & MCP builds to teach true AI Product Sense.

#22 𝕏

Mistral AI partners with Nvidia as a founding member of the Nemotron Coalition, launching their first joint project to accelerate open-frontier AI models using Nvidia’s GPU-optimized NeMo toolkit.

#23 𝕏

Philipp Schmid published a step-by-step Nano Banana 2 developer guide covering Debian flashing, eMMC/Wi-Fi setup, GPU driver install and peripheral demos (MIPI CSI-2 camera, I2C/SPI). It also walks through Python/C SDK usage with TFLite benchmarks showing ~20 FPS on CIFAR-10.

#24 𝕏

claire vo 🖤 demonstrates how non-engineers can build iOS apps by running dual Claude Code agents that provide step-by-step workflows, feature-prioritization prompts and daily review routines.

Also covered by: @Claire Vo

#25 in

Peter Yang warns PMs that specs are now read by AI agents—not engineers—so mastering advanced AI skills beyond basic ChatGPT prompts is essential. He even argues that if agents deliver 10× more work, token budgets could eclipse salaries, and waterfall processes won’t survive.