Midjourney's 60-second body scan 🤖, Gemma 4 in your browser 💻, frontier labs lose their grip 🤖

The AI image company announced Midjourney Medical, a new division building what it calls Ultrasonic CT. A full-body scan takes 60 seconds using only sound and water, with no radiation or magnets. Midjourney plans to deploy 50,000 scanners worldwide within six years to run a billion scans a month, starting with a flagship Midjourney Spa in San Francisco at the end of 2027.

OpenAI is stacking talent ahead of its public debut, hiring Google DeepMind's Noam Shazeer, a co-author of the original Transformer paper and a Gemini co-lead, along with former Trump White House AI policy official Dean Ball. Google had rehired Shazeer two years ago in a $2.7 billion deal for his startup Character AI. The moves read as OpenAI shoring up both research firepower and policy credibility pre-IPO.

After forcing Anthropic to suspend Fable 5 and Mythos 5 over a jailbreak, the White House is now working with the company on a framework to grade the severity of security flaws in new models and decide when the government should intervene. Officials and Anthropic had clashed over how dangerous the flaw really was. The shift to standards-setting is the first sign the export-control standoff is starting to ease.

AWS is in talks to sell its Trainium AI chips to other companies for their own data centres, breaking a long-standing refusal. CEO Andy Jassy has said a standalone chip business would run at roughly $50 billion a year, near Intel's revenue. Amazon held back because hosting the chips itself also pulls in storage, networking and security fees, but demand keeps outstripping supply, with each generation selling out almost instantly.

The Model Context Protocol's Enterprise-Managed Authorization extension is now stable, so companies can govern which MCP servers an agent reaches through their existing identity provider instead of clicking through OAuth for every server. One corporate sign-in provisions every connector, with no consent screen, using the new ID-JAG grant and Okta's Cross App Access. Anthropic and Microsoft support it first across Claude, Claude Code and VS Code.

Google made AI avatars in Google Vids free for anyone with a US Google account, with a global rollout to follow. The avatars act as digital presenters you can re-script anytime, and a Slides integration turns a deck into a narrated video automatically. Free users get ten generations a month and can build custom avatars with Nano Banana 2, with emotion steering coming later in June.

Following Project Glasswing, Cloudflare details how to build a security-scanning harness that treats frontier models as interchangeable parts. One model handles discovery and a different one does validation, so vulnerabilities get cross-checked by distinct logic. It traces flaws across cross-repo dependencies and filters thousands of raw candidates into a triaged queue of real fixes, the layer it calls more important than any single model.

PyTorch's Helion kernel language normally tunes performance through Bayesian search that grinds hundreds of compile-and-benchmark cycles per kernel. A new LLM-guided autotuner matches that performance while benchmarking roughly ten times fewer configurations in 6.7 times less wall-clock time. The result holds across Opus 4.8, GPT-5.5 and Sonnet 4.6, suggesting LLM-guided search is a practical, model-independent way to speed up kernel tuning at production quality.

Browser Use rebuilt its cloud browsers on Firecracker micro-VMs to hit three goals at once: fast startup, real isolation and low cost. Each session now gets its own lightweight VM on EC2, starting in under a second and costing $0.02 per browser-hour, down from $0.06. The writeup explains why ordinary VMs are too heavy when you spin up and discard thousands of disposable browser sessions a day.

Elastic walks through a persistent agent-memory system built on three Elasticsearch indices, treating a million-token context window as a scratchpad rather than real memory. It blends hybrid retrieval with a cross-encoder reranker, supersession to resolve contradictions, time decay and per-user document-level security. On a 168-question evaluation it averaged 0.89 recall at ten results with zero cross-tenant leaks, and the full implementation is on GitHub.

This piece makes the case that the two-year assumption that a few labs own AI's value is breaking on several fronts. Open-weight models like Qwen3 and Kimi have nearly closed the quality gap. Enterprises such as Uber and Priceline are now capping token budgets rather than defaulting to the priciest model, and a government order switching off Anthropic's best model overnight showed how fragile a single-vendor foundation is.

Verna reframes the AI story around what software gets built, not just faster developers. When building required scarce, expensive engineers and venture funding, only ideas large enough to justify the cost ever shipped. As that cost collapses, the winners become teachers, accountants, agents and other domain experts shipping narrow tools for problems they understand deeply, an era she calls Mom-and-Pop SaaS.

The post explains why teams full of more-productive people still aren't shipping faster. When an engineer has a model draft a tech brief, reviewers receive something several times longer that they must now fact-check line by line, because they can't tell which claims the author actually stands behind. The afternoon he saved gets quietly spent by everyone who reads it, since a document has one writer and many readers.

RTK has 60,000 GitHub stars on a promise to cut token usage tenfold. This teardown shows the viral savings stat only measures how much raw terminal output it strips, not your real invoice, which is driven by file reads, repo context and reasoning tokens. Worse, the agent never knows text was compressed, so a dropped stack-trace line fails silently, and RTK publishes no task-success benchmarks.

Quanta lays out why sequence-only models like AlphaGenome struggle to capture how genes actually work. Regulation depends on the genome's shifting three-dimensional structure, its chromatin loops, folding domains and fluid transcription hubs, not just the linear code these models train on. As one researcher puts it, the field has stopped thinking of the genome as a flat strand of DNA.

cocoindex-code is a lightweight, embedded code-search CLI that indexes your repository by its abstract syntax tree rather than plain text. That lets a coding agent pull the exact code it needs instead of dumping whole files into context. The project says it saves around 70% of tokens and speeds agents up, runs locally in Python, and is trending with roughly 2,100 GitHub stars.

Clear is a v0.1 programming language built agent-first, where the specification and the running implementation are the same file, so documentation and behaviour can't drift apart. Plain-English keywords like ability, rule and data describe what a system should do, and clear run executes that spec directly. The bet is that a language designed for agents to read and run turns out to be more legible for humans too.

The webml-community built a Hugging Face Space that runs Google's Gemma 4 entirely in your browser, GPU-accelerated through hand-written WebGPU kernels with no server and no install. Open the page and the model loads and runs locally on your own hardware, so your prompts never leave the device. It is a concrete demo of where private, zero-backend LLM inference is heading for web apps.