Midjourney's Second Act: A Full-Body Scanner for Consumer Health

🧠 LAUNCH

Midjourney's Second Act: A Full-Body Scanner for Consumer Health

Midjourney — the bootstrapped image generation lab that scaled to millions without a dollar of VC — just announced its second product, and it's not what anyone expected. A consumer-grade full-body medical scanner positioned as casually as stepping on a bathroom scale. The pivot from generative art to medical imaging sounds wild until you remember that both are fundamentally about turning signals into high-fidelity visual representations. If Midjourney can bring the same accessibility-first approach to medical imaging that it brought to AI art, the diagnostic imaging market just got a new kind of competitor. Read more →

Claude Code Gets Artifacts — Preview React and HTML Without Leaving the Terminal. Artifacts — the interactive previews you know from claude.ai — now work directly in Claude Code. Preview React components, HTML pages, and data visualizations inline without context-switching to a browser. If you're building UI in Claude Code, this collapses the edit-preview loop into a single surface. (Desktop vs terminal comparison for the full breakdown.) Read more →

AWS Launches Continuum and Context at New York Summit. AWS dropped two new services at its New York Summit: Continuum (AI-powered security at machine speed) and Context (a knowledge graph purpose-built for agents), alongside updates to Kiro, DevOps Agent, Transform, Amazon Quick, and Bedrock AgentCore. The message is clear — AWS is betting that the infrastructure layer for enterprise agents is where the real lock-in happens. (10 likes) Read more →

🔧 TOOL

MCP Gets Enterprise-Grade Auth — Admins Can Now Centrally Manage Connector Access

The biggest friction point in enterprise MCP adoption just disappeared. Anthropic shipped centralized authorization management for MCP connectors — admins can now pre-authorize OAuth flows and control which connectors users can access, all from a single pane. No more per-user OAuth prompts, no more shadow IT connector sprawl. If your org has been holding off on MCP because of auth headaches, this is your green light. Read more →

Claude's Code Execution Tool Gets 90-Second Cell Budgeting. The updated code_execution_20260521 tool version lets Claude budget long-running computations against a 90-second per-cell time limit — no more silent timeouts on complex data processing. No beta header required, just update your tool version string. Read more →

Ollama v0.30.10 Brings Command A and North Models to Apple Silicon. Ollama's latest release adds support for Command A and the North model family via the MLX engine. If you've got an M-series Mac and want to run these models locally without spinning up GPU servers, ollama pull and go. Read more →

📝 TECHNIQUE

The Complete Guide to Steering Claude Code: CLAUDE.md, Skills, Hooks, and Subagents

Anthropic published the definitive reference for customizing Claude Code's behavior — covering CLAUDE.md files, skills, hooks, rules, and subagents in one comprehensive guide. This isn't a tutorial; it's the full surface area. If your team is using Claude Code in production and you're still relying on ad-hoc prompting, this is the document that turns it into a repeatable engineering workflow. Read more →

Beyond LoRA: HuggingFace Benchmarks the Fine-Tuning Methods You're Not Using. HuggingFace surveys LoRA alternatives head-to-head — DoRA (weight-decomposed adaptation), LoRA+, rsLoRA, and PiSSA (principal singular value adaptation). The takeaway: vanilla LoRA is no longer the automatic best choice, and the performance gaps on specific tasks are large enough to justify benchmarking alternatives before your next fine-tuning run. Read more →

Zero-Touch OAuth: How Enterprise MCP Deployments Kill the Auth Prompt. The MCP protocol blog details exactly how zero-touch OAuth works under the hood — admins pre-authorize connectors at the org level so end users never see an OAuth consent screen. Read this alongside Anthropic's managed auth announcement above for the full picture of how enterprise MCP auth is supposed to work now. (47 likes | 14 RTs) Read more →

🔬 RESEARCH

DeepMind Publishes a Concrete Roadmap for Securing AI Agents. Google DeepMind released an AI Control Roadmap that combines traditional safeguards with real-time monitoring for agentic systems. This is a practical security framework with specific implementation guidance, not another thought piece about hypothetical risks. If you're deploying agents with access to internal systems, this is your reading list. Read more →

Stop Using Generic Benchmarks — Build Agentic Evals for Your Own Tools. HuggingFace drops a framework for evaluating how well open models handle tool use with your specific tools — not the sanitized function-calling benchmarks that tell you nothing about production performance. The methodology lets teams build custom agentic evals that actually predict how a model will behave with their API surface. Read more →

MosaicLeaks: Your Research Agent Might Be Leaking Private Data. ServiceNow tested whether research agents leak sensitive information when given broad document access. The results are uncomfortable — agents that individually access harmless documents can reconstruct private data by combining fragments across sessions. The attack surface is real, and most teams haven't thought about it. Read more →

💡 INSIGHT

Noam Shazeer Joins OpenAI — The Transformer's Co-Inventor Switches Sides

Noam Shazeer — co-inventor of the Transformer architecture, creator of the attention mechanism, and founder of Character.AI — just announced he's joining OpenAI. This is one of the most consequential talent moves in AI history. Shazeer doesn't just understand Transformers theoretically; he built the thing. His presence at OpenAI signals that the next generation of architecture improvements will come from someone who knows exactly where the current design's limits are. (260 likes | 236 RTs) Read more →

GLM-5.2's Frontend Coding Dominance and the IndexShare Decoding Trick. Latent Space digs into why GLM-5.2 is crushing frontend coding benchmarks and unpacks IndexShare, a speculative decoding technique that's quietly making open models faster at inference. The technical breakdown goes deeper than yesterday's benchmark headlines — worth reading if you're evaluating GLM-5.2 for production use. Read more →

Mollick: AI Interfaces Aren't Intuitive — You've Just Internalized the Tricks. Ethan Mollick nails a critical UX blind spot: power users have absorbed dozens of AI interaction patterns — specific phrasing, workarounds, mental models — that feel "obvious" but are completely invisible to everyone else. If you're building AI products and only testing with your team, you're testing with people who've already learned the tricks. (576 likes | 30 RTs) Read more →

🏗️ BUILD

last30days-skill: Cross-Platform Research Agent Hits 44K Stars. A drop-in agent skill that synthesizes recent discussion across Reddit, X, YouTube, Hacker News, Polymarket, and the open web into a grounded summary. The 44K star count isn't hype — it reflects genuine demand for cross-platform research automation that goes beyond single-source search. Add it to your agent toolkit and stop manually cross-referencing platforms. (43,974 likes | 3,617 RTs) Read more →

GLM-5.2 GGUF Quantizations Drop — Run the New Open-Weight Leader Locally. Unsloth already has GGUF quantizations of GLM-5.2 trending on HuggingFace. If you want to benchmark the model that's topping frontend coding leaderboards on your own hardware before committing to an API, the local path is open. (120 likes | 305 downloads) Read more →

🎓 MODEL LITERACY

Low-Rank Adaptation Variants (DoRA, PiSSA, rsLoRA): LoRA fine-tunes large models cheaply by injecting small trainable matrices into frozen layers — but it's no longer the only game in town. DoRA (Weight-Decomposed Low-Rank Adaptation) splits weight updates into magnitude and direction components, often outperforming vanilla LoRA on tasks that need precise directional adjustments. PiSSA (Principal Singular Value Adaptation) initializes the low-rank matrices using the principal components of the original weights rather than random values, giving training a head start. rsLoRA scales the learning rate by rank, fixing a subtle issue where higher-rank adaptations learn too slowly. The practical upshot: PEFT method selection is now a real engineering decision — defaulting to vanilla LoRA without benchmarking alternatives means you might be leaving performance on the table.

⚡ QUICK LINKS

• Anthropic Python SDK v0.111.0: Refusal-fallback middleware tagging + code execution tool support. Link
• Anthropic TS SDK v0.105.0: Lazy partial tool JSON parsing — faster streaming for tool-heavy agents. Link
• OpenRouter Robot Combat Arena: Claude vs Grok in real-time robot combat simulations. Genuinely fun, surprisingly informative. (142 likes | 116 RTs) Link
• LangGraph 1.2.6: Fixes nested subgraph checkpoint bug from 1.2.3 — update immediately if affected. Link
• "Done" Is the Most Expensive Word Your Agent Says: Same task, two runs — one trusts "done," one verifies. Night and day. (6 likes | 1 RTs) Link

🎯 PICK OF THE DAY

Noam Shazeer joining OpenAI is the talent signal that reshapes the architecture race. The co-inventor of the Transformer — the person who literally wrote "Attention Is All You Need" — just walked into OpenAI's office. This isn't a typical executive hire or acqui-hire shuffle. Shazeer built the attention mechanism that every frontier model runs on, then founded Character.AI and proved that novel architectures could scale consumer products. His move reveals something important: the Transformer architecture race is far from over. The next frontier isn't replacing Transformers — it's radically reimagining them from the inside, and Shazeer is perhaps the one person on Earth best positioned to do that. For the broader ecosystem, the talent consolidation at frontier labs is accelerating faster than the open-source community can absorb. When the person who invented the architecture joins the company with the most compute, the gap between frontier and open models risks widening at exactly the moment we need it to narrow. (260 likes | 236 RTs) Read more →

Until next time ✌️