OpenAI Lets ChatGPT Subscribers Code Inside Zed — No Codex Sub Required
🧠 LAUNCH
OpenAI Lets ChatGPT Subscribers Code Inside Zed — No Codex Sub Required
Your ChatGPT plan now works directly inside the Zed editor's agent — same usage limits, no separate Codex subscription. This is OpenAI deliberately lowering the barrier to AI-assisted coding by meeting developers where they already are, rather than forcing them into a new product. If you've been Zed-curious, the cost of trying just dropped to zero. (2,265 likes | 127 RTs) Read more →
A 30B Model With Only 3B Active Parameters Hits Olympiad Gold in Math and Physics
A new 30B-A3B mixture-of-experts reasoning model reaches gold-medal level on both physics and math Olympiad evaluations — with only 3B parameters active per forward pass. That means frontier-class reasoning running on consumer hardware. The MoE architecture routes each token to a small subset of expert sub-networks, so you get the knowledge of 30B params at the inference cost of 3B. This changes who gets to run serious reasoning models locally. (703 likes | 87 RTs) Read more →
Microsoft Drops Lens: A 3.8B Text-to-Image Model Built for Training Efficiency. Lens is a 3.8B-parameter text-to-image model designed for high-quality output at a fraction of the training compute. If you need image generation without burning a small fortune on GPU hours, this is your open-weight starting point. (165 likes | 17 RTs) Read more →
xAI Finishes Training Grok V9 at 1.5 Trillion Parameters. Grok V9 clocks in at 1.5T parameters — 3x larger than V8 — and Musk says early results show a "gigantic" gap, even before fine-tuning on coding data. Another frontier contender muscling into the coding agent race; benchmarks will tell us if the parameter count translates to actual capability. (34 likes | 12 RTs) Read more →
🔧 TOOL
Codex Goes Mobile Inside ChatGPT — Setup Guides Are Already Circulating. Developer walkthroughs are popping up for running Codex directly in the ChatGPT mobile app — setup, workflow tips, the works. The mobile coding agent pattern just went from novelty to documented practice. (630 likes | 54 RTs) Read more →
Claude Code v2.1.143 Adds Plugin Dependency Enforcement and Worktree Isolation. Claude Code v2.1.143 ships plugin dependency enforcement, projected context cost in the plugin marketplace, and background worktree isolation. If you're running multi-agent or plugin-heavy workflows, this tightens the reliability story considerably. Read more →
Supabase Ships an Official Plugin for AI Coding Agents. The Supabase plugin bundles MCP tools and Supabase-specific skills that work across Codex, Claude Code, Cursor, and Gemini CLI. Database-as-a-service just became agent-native — add it and your coding agent can scaffold secure Supabase apps out of the box. (82 likes | 11 RTs) Read more →
📝 TECHNIQUE
Hard Numbers: MCP Uses 10x More Tokens Than Code-First Agent Design on the Same API
Someone finally ran the experiment. Testing against Monday.com's GraphQL API: the SDK-native approach completed in 1 step using 15k tokens. The MCP server took 4 steps and 158k tokens — over 10x the cost for the same result. This isn't a theoretical argument anymore. If you're building agent tooling, benchmark your MCP integrations against direct SDK calls before assuming the standard protocol is the efficient path. (136 likes | 12 RTs) Read more →
Simon Willison: Coding Agents Make "Port It, Then Port It Back" Economically Rational. Simon Willison points out that coding agents have made throwaway ports cheap enough to be a real strategy — port a native mobile app to React Native to test a hypothesis, then port it back. The economics of rewriting code just fundamentally shifted when the labor cost approaches zero. (248 likes | 10 RTs) Read more →
The Small But Meaningful Gaps Between Claude Code and Codex CLIs. A developer documents the practical differences: different skill names and shortcuts, different built-ins (docs, browser testing), subtle feature gaps. If your team is choosing between the two, this side-by-side saves you the discovery time. (45 likes | 2 RTs) Read more →
🔬 RESEARCH
UK AISI Confirms Mythos Preview Is First Model to Solve Both Cyber Ranges End-to-End. The UK's AI Safety Institute found that Mythos Preview is the first model to complete both their cybersecurity evaluation ranges from start to finish — no model had ever done it before. A significant capability jump that will reshape how we think about AI in offensive and defensive security. (1,079 likes | 51 RTs) Read more →
LeCun's Useful Framework: LLMs Excel Where Language IS the Reasoning, Not Just the Description. Yann LeCun sharpens his LLM critique: these models are strongest in domains where language is the actual substrate of reasoning — math proofs, code, logic. They struggle where language merely describes a process that happens in another medium, like physical manipulation or spatial reasoning. A practical heuristic for deciding where to deploy LLMs vs. other approaches. (642 likes | 91 RTs) Read more →
Ontario Auditors Find AI Medical Note-Takers Routinely Get Basic Clinical Facts Wrong. Ontario's auditors tested AI-powered clinical note-taking tools and found they regularly botch basic facts — wrong medications, incorrect dosages, fabricated symptoms. A sobering counterweight to the AI healthcare success stories, and a reminder that verification layers aren't optional in high-stakes domains. (40 likes | 7 RTs) Read more →
💡 INSIGHT
Anthropic Commits $200M to the Gates Foundation for Global Health, Education, and Agriculture
Anthropic pledges $200 million in grants, Claude credits, and technical support to the Gates Foundation — covering global health, life sciences, education, agriculture, and economic mobility. This is the largest philanthropic commitment by any AI lab, and it's a deliberate positioning move: safety-focused AI as global infrastructure, not just a Silicon Valley product. Whether it reshapes outcomes in malaria research or crop yield optimization will take years to judge, but the signal is immediate. (2,094 likes | 191 RTs) Read more →
OpenAI Merges ChatGPT, Codex, and API Teams Under Brockman's Unified-App Vision. With Fidji Simo on leave, Greg Brockman takes the reins of a reorganized OpenAI product org — merging ChatGPT, Codex, and API teams into a single unified-app strategy. Every OpenAI product is converging into one surface. This is an organizational bet that the lines between chat, code, and API are artificial. Read more →
Anthropic Publishes Legal Industry Deployment Case Study for Claude. Anthropic drops a detailed deployment case study for Claude in the legal vertical — concrete patterns for regulated industries, not just benchmarks. If you're deploying AI in legal, healthcare, or finance, the compliance and workflow integration patterns here are directly transferable. Read more →
Mollick: Codex Still Treats Non-Coders as Less Competent, Not Differently Competent. Ethan Mollick argues that Codex — pitched as an everything app — still defaults to a developer-centric interface that treats non-technical users as simply less capable, rather than designing genuinely different interaction models for different expertise levels. A pointed UX critique for anyone building AI tools beyond the developer audience. (475 likes | 9 RTs) Read more →
🏗️ BUILD
Watch a Neural Net Learn Snake in Real Time — Interactive PPO Visualization. An in-browser, real-time visualization of PPO reinforcement learning training a neural net to play Snake. You can watch the agent go from random wandering to strategic play, adjusting hyperparameters as it trains. Bookmark this for the next time you need to explain RL to someone who learns by seeing. (110 likes | 28 RTs) Read more →
🎓 MODEL LITERACY
Mixture of Experts (MoE) and Active Parameters: Today's headline — a 30B-parameter model hitting Olympiad gold with only 3B active parameters — is a perfect example of why "how many parameters?" is no longer the right question. In a Mixture of Experts architecture, the model is divided into many specialized "expert" sub-networks, but each input token is routed to only a small subset of them. So a 30B-total, 3B-active model has the stored knowledge of 30 billion parameters but the inference cost of 3 billion — meaning it can run on a single consumer GPU while matching models that need a full server rack. When evaluating any model going forward, ask two numbers: total parameters (knowledge capacity) and active parameters (actual compute cost). The gap between them is where MoE earns its keep.
⚡ QUICK LINKS
- Hashimoto Calls Out "AI Psychosis": HashiCorp founder says entire companies are making irrational bets under AI hype. (612 likes | 283 RTs) Link
- Abridge Hits 100M Doctor Visits: AI medical scribe saves clinicians 10-20 hours/week at scale. Link
- Codex vs. Claude Business Model War: Latent Space breaks down the competitive landscape and Anthropic's metered pricing. Link
- Next Week's Showdown: GPT 5.6 vs Gemini 3.2 announcements expected — clear your calendar. (200 likes | 5 RTs) Link
🎯 PICK OF THE DAY
The 10x token gap between MCP and code-first agents is the number everyone building agent tooling needs to see. When someone actually benchmarked MCP against a direct SDK approach on the same GraphQL API, the result wasn't close — 158k tokens vs. 15k, four steps vs. one. That's not a rounding error; it's an order of magnitude. The industry's rush to standardize tool interfaces through MCP makes sense from an interoperability standpoint — one protocol, many tools, plug-and-play. But interoperability and efficiency are in tension. MCP's dynamic tool discovery and schema negotiation add conversational overhead that compounds with every API call. Code-first agent designs, where the tool interface is baked directly into the SDK, skip that overhead entirely because the agent already knows exactly how to call the function. This isn't a case for abandoning MCP — standardization has real value. But it's a case for being honest about the tradeoff: you're paying 10x in tokens for the flexibility of not writing custom integrations. For high-frequency agent workflows, that math doesn't work. The smart play is MCP for long-tail integrations you use occasionally, and tight SDK bindings for the APIs your agents hit thousands of times a day. Read more →
Until next time ✌️