Everything That Happened in AI Today Thursday, June 18

Midjourney apparently looked at the MRI machine and said, “what if this were faster, cheaper, wetter, and somehow a spa?”

Welcome to the Around the Horn Digest, the one page you need to sound dangerously informed at work tomorrow. The day’s strangest story came from Midjourney, which is now pitching full-body ultrasound scans like a futuristic bathhouse membership. Meanwhile, OpenAI recruited one of Google’s most important AI architects, life-science benchmarks got much more realistic, robot data became a funded business, and every coding-agent person on X seemed to be arguing about skills, filesystems, kernels, and whether SaaS is mostly a database wearing a blazer. Normal Thursday stuff, assuming your Thursdays now include medical tubs and autonomous lab agents. Let’s get into it.

Around the Horn — Thursday, June 18, 2026

The big news today was Midjourney deciding image generation was apparently too small a canvas. The company introduced Midjourney Medical and the Midjourney Scanner, a full-body ultrasonic computational tomography system that sends millions of sound waves through water to produce a detailed internal scan in about 60 seconds.

The wild part is the packaging: Midjourney wants the scan to feel as casual as a spa visit, with a first San Francisco “Midjourney Spa” planned for 2027. The long-term ambition is even weirder: 50,000 scanners worldwide capable of roughly a billion scans per month by 2031. That would turn medical imaging from a rare, expensive event into a recurring data stream. Useful for medicine, huge for AI training data, and a little bit like turning every checkup into an aquatic iPhone backup.

Several people immediately clocked the second-order move. Frequent, cheap scans could produce longitudinal health data that personal AI doctors can reason over, while the same focused-ultrasound hardware might eventually move from diagnosis into therapy. Regulators, doctors, insurers, and hypochondriacs with premium memberships are all going to have feelings about that.

🏆 TOP 5 NEWS (Around the Horn)

Noam Shazeer announced he is joining OpenAI to lead architecture research, with The Information noting the move takes a Transformer co-author and former Gemini technical co-lead back out of Google after the $2.7B Character.AI deal; the story was still rolling through X trending as Sam Altman welcomed him publicly.
OpenAI introduced LifeSciBench, an expert-authored benchmark of 750 real-world biological research tasks that tests how models handle incomplete evidence, conflicting results, experiment design, assay troubleshooting, translational risk, and research decisions under uncertainty; OpenAI’s thread framed it as a more realistic test for scientific assistants.
Anthropic’s foreign-access fight expanded after the White House ordered SK Telecom’s Claude Mythos access revoked over alleged China-ties concerns, WIRED reported officials wanted impossible-to-jailbreak guardrails before Fable 5 returns, Reuters said JPMorgan blocked Anthropic access for Hong Kong staff, and TechCrunch tied the blackout to foreign leaders’ fear that American AI can be shut off overnight.
Odyssey raised a $310M Series B from Natural Capital, Amazon, GV, AMD, IQT, and others to build world models that simulate the physical world for robotics, science, healthcare, education, and gaming.
Block rolled out Builderbot, an internal agent orchestration layer that works inside Slack across the company’s multi-service codebase, now accounts for about 15% of production code changes, and merges roughly 1,500 PRs per week while 100% of Block engineers regularly use AI.

Honorable Mentions

AWS announced Continuum and Context at its New York Summit as part of a broader push to help organizations coordinate agents, preserve context, and deploy them across real workplace operations instead of leaving them as isolated pilots.
Pew Research Center found 49% of U.S. adults have used AI chatbots, up from 33% in 2024, while 63% say AI is advancing too quickly.
PwC analyzed 1B+ job ads across 27 countries and found AI is splitting work into “professionalized” roles, where AI raises demand for judgment, creativity, and leadership, and “democratized” roles, where AI reduces the need for deep specialization; the most AI-exposed companies grew headcount 52% versus 36% for the least exposed, while top “superstar” firms saw 163% productivity growth.
Bernie Sanders proposed a nearly $7T public AI ownership fund funded by a one-time 50% stock tax on AI companies with more than $200M in annual sales, giving Americans voting power, more than $1K per person annually in potential dividends, and money for healthcare, education, and housing.
Uber said it will bring a premium robotaxi service to Houston in 2027 using Lucid EVs with Nuro’s self-driving system.

🍪 TOP TREATS TO TRY

ChatGPT scheduled tasks now have a dedicated management page and a faster, more reliable experience, with TestingCatalog noting the update replaces Pulse for Pro users who want recurring briefings from interests, past chats, and connected apps — included in eligible ChatGPT plans.
Claude Design now imports your design system, checks its own work against your components, supports steadier on-canvas editing, exports to PDF and PowerPoint, and connects to tools like Adobe, Canva, Miro, Replit, Vercel, and Wix, building on Anthropic’s April preview and Nate Parrott’s demo thread — available in public beta on Pro, Max, Team, and Enterprise.
Google Flow turns ideas into cinematic video, images, and custom creative tools with Veo 3.1, Gemini Omni, and Nano Banana support, while Flow by Google showed off the agent-style planning, storyboard, and editing workflow — free to try with a Google Account, paid plans available.
Ploy turns your website into an autonomous marketing system that handles site updates, brand, CMS, CRM, campaigns, analytics, SEO, and AI answer-engine optimization, with founder Bryant Chou’s launch thread noting a $27M seed and early use at Hex and Clay — free to try.
Palmier Pro gives AI video generation a native editing timeline, so you can generate images, video, and audio from models like Kling, Veo, and Grok Imagine, edit the clips, and export to Premiere or DaVinci; Marcos Rich’s launch post and Philipp Schmid’s demo showed the workflow — pricing details not listed.
Photon gives agents an iMessage App API, so they can send interactive menus, games, and mini UIs inside iMessage; Daniel Tian’s demo framed it as a way for text-based agents to ship actual interface moments inside a chat — pricing details not listed.
Honen turns docs, call recordings, kickoff videos, slide decks, or topics into courses with modules, lessons, projects, auto-grading, and a live tutor; Esan Durrani’s launch note reinforced the onboarding and training use case — free to start.

🏢 Big Tech & Major Companies

Apple is raising prices after Tim Cook said higher memory and storage costs are “unavoidable,” with AI demand, chip shortages, and suppliers prioritizing server chips making it harder for the company to keep absorbing component inflation.
Microsoft reportedly built a significant AI business in China by selling OpenAI models and cloud services to local companies, with ByteDance alone reportedly on track to spend more than $1B annually despite the U.S.-China rivalry over advanced AI.
Asian chip suppliers are seeing demand surge because data centers need much more than Nvidia GPUs: memory, boards, power components, and manufacturing capacity are all part of the boom.
AWS announced Continuum and Context at its New York Summit as part of a broader push to help organizations coordinate agents, preserve context, and deploy them across real workplace operations instead of leaving them as isolated pilots.
Unreal Engine 5.8 shipped with upgrades for bigger worlds, higher-fidelity characters, and faster lighting/rendering, with Unreal pushing the launch to developers.

💼 AI Productivity, Labor & Economics

Convey raised $38M from a16z to automate repetitive office work, while its product site and Gabe Pereyra’s launch note describe enterprise-grade digital teammates that run in the background across workplace tools.
XDOF emerged from stealth with $70M raised and 20 customers, building the data pipelines, collection tools, and annotation systems that frontier AI labs need when they would rather outsource the physically demanding work of gathering real-world robot manipulation data; its careers page shows the hiring push behind that data operation.
Clair Health raised $11M to build a noninvasive wearable with 10 biosensors and voice-biomarker analysis for hormone tracking, cycle phase classification, inflammation, bloating, energy, and perimenopause insights.
NeuralTrust raised $20M to expand its AI agent security platform across Europe.
Comand AI raised a €32M Series A to scale its Prevail command-and-control platform across NATO, using specialized agents for mission analysis, courses of action, and tactical control to compress the OODA loop (observe, orient, decide, act) while keeping humans in command; Saab joined as a strategic backer and integration partner.
Data center neighbors are fighting constant low-frequency noise from cooling systems, generators, and fans, with residents reporting health problems, sleep disruption, and property-value fears as AI infrastructure expands; the story also spilled into a large Reddit discussion about rural communities getting stuck beside the machinery behind “the cloud.”

🤖 AI Agents & Infrastructure

Ben Holmes showed how to move from manual Claude Code prompting to automated loop engineering, using a YouTube walkthrough, the Codex app, Oz, and Matt Pocock’s skill patterns to scope issues, generate specs, implement, and open PRs for human review.
DAIR.AI highlighted PreAct, a computer-using agent paper that turns a successful screen task into a reusable state-machine program (a checklist-like script of screen checks and actions), replaying repeat tasks 8.5–13x faster without calling a language model at every step, then handing control back to the full agent when the screen no longer matches expectations; the paper also adds a store-time evaluator that only keeps programs proven to solve the task again on a clean run.
OpenAI Codex now supports OSS mode for local providers and third-party endpoints, with Thibault Sottiaux noting open-source models can run through the Codex app, CLI, and SDK.
Ollama said GLM-5.2 and Kimi-K2.7-Code can now run in Codex using local Ollama launch commands.
Vercel Labs released V, an open-source personal-agent template built on Eve that works across iMessage, Slack, and web, with built-in GitHub and Linear tools plus long-term memory with user-approved saves.
mistral.rs added native Agent Skills inside a Rust inference engine, and Michael Buehler’s thread argued small local models can now run scientific workflows with persistent Python sessions, code execution, and tool use on-device.
Google Gemma showed Gemma 4 26B orchestrating 10 local sub-agents to code an SVG gallery at 100+ tokens/sec, with a cookbook example and a second Gemma post for concurrent multi-agent workflows.
Omma launched Omma Canvas, a massive parallel agent workspace where you can run up to 100 agents at once to build real HTML/CSS/JS sites and web apps, keep each frame scoped with its own version history, enforce design.md or a custom design system, and publish results standalone or move them into Omma chat for domains and settings.
Pipecat 1.4.0 introduced Pipecat Evals, a behavioral evaluation framework for voice and multimodal agents: developers can define scenarios in YAML or code, simulate user audio, transcribe agent responses, mix modalities, and run tests from the CLI or library so coding agents can edit the voice agent, run evals, read results, and iterate until the scenario passes.

💻 AI Coding & Developer Tools

turbopuffer open-sourced alyze, a Rust tokenization and analysis pipeline for full-text search that powers its production tokenizer and runs up to roughly 4x faster than the previous Tantivy-based approach.
OpenRouter shipped a Cost Simulator that compares what the same generations would cost across different models using effective pricing.
Lightricks released LTX Trainer, LTX-2.3 docs for open audio-video generation, and a Hugging Face collection of free IC-LoRAs for relighting, colorization, deblurring, compression cleanup, water effects, beard removal, and in/outpainting.
Open Design released 80+ slide templates for its open-source Claude Design alternative, a local-first desktop app with templates, skills, design systems, exports, and support for many coding CLIs.
Greenlight is an open-source pre-submission compliance scanner that helps developers catch App Store issues before sending an app to Apple.
Spandan Madan reverse-engineered Claude Code’s tool system, describing a uniform 30-method interface, 43+ tools, a seven-phase dispatch pipeline, permissions, hooks, concurrency scheduling, streaming execution, and fail-closed behavior that turns the model into a runtime instead of a chatbot.
Xenova used Fable 5 to write custom WebGPU kernels (browser GPU code) for Gemma 4 inference, climbing from 84 to 255 tokens/sec after Anthropic rolled back invisible LLM-development safeguards, then released the demo and kernels so people could run the optimized model locally in the browser before global Fable 5 access was suspended the next day.
Kedyt demonstrated a Codex “Build iOS Apps” plugin that runs apps in an in-app browser, opens SwiftUI previews, and hot-reloads edits so iOS developers can stop doing copy-paste, build, screenshot loops.

🔬 AI Research & Models

LoopCoder-v2 introduced 7B Parallel Loop Transformer code models trained on 18T tokens, with Dorothy Du’s thread explaining why two loops help and three loops regress, making “only loop once” the key design lesson.
AMIE advanced Google’s medical dialogue work into disease management, matching 21 primary care physicians on specialist-graded multi-visit reasoning across 100 virtual cases, scoring better on treatment precision and guideline alignment, and outperforming doctors on harder medication questions in the RxQA benchmark.
OpenAI and Molecule.one showed a near-autonomous AI chemist using GPT-5.4, Maria AI, and Maria Lab to improve a difficult Chan-Lam reaction; Piotr Kudła’s thread said the system proposed TEMPO and later 4-hydroxy-TEMPO as unexpected effective additives, while humans still executed the reactions in a high-throughput lab.
NVIDIA ENPIRE showed agentic robot policy self-improvement in real environments, using coding agents, automatic outcome checks, multi-agent algorithm evolution, and self-resetting tasks to improve dexterous manipulation like pin insertion and zip-tie cutting, reaching up to 99% pass@8 success with little human supervision.
DAIR.AI surfaced From Trainee to Trainer, a reinforcement-learning paper (trial-and-error training) where an LLM studies its own failure trajectories and proposes the next training environment itself; the paper introduced a controllable multi-agent FrozenLake testbed and found a Qwen3-4B checkpoint outperformed larger proprietary models and fixed-environment baselines at designing better next stages.
cuTile Rust brought Rust-style compile-time data-race safety to GPU kernel programming; Jared Roesch and Melih Elibol’s writeup said safe GEMM hit about 2 PFlop/s on B200, roughly 92% of dense f16 peak, with element-wise kernels near 91% of peak memory bandwidth.
Functional Attention reframed attention as functional correspondences between learned bases rather than pairwise token matching, and Simon Weber’s thread said it gives linear complexity with strong results on PDEs and 3D segmentation.
MiniT2I showed a minimal open-source text-to-image system can work directly on pixels with no auxiliary losses, using academia-scale compute while approaching state-of-the-art GenEval performance.
MolmoMotion released an open language-guided 3D motion forecasting model for predicting how object points move, with models, dataset, paper and AllenAI’s launch all public.
QUEST is an open family of 2B to 35B deep research agents trained with fully synthetic rubric-tree tasks, with paper, code, models, demo and the OSU NLP announcement all released.
DreamReasoner-8B is an open block-diffusion reasoning model for math and code, with Zirui Wu’s announcement, SGLang support and inference update showing efficient parallel block generation.
Neural Cellular Automata got a browser-based JAX / WebGPU demo, with Eric Zhang’s post showing high-performance ML experiments running in pure JavaScript.
Papers with Code continued surfacing trending AI research papers with code, datasets, methods, and evaluation leaderboards.

🏛️ AI Policy, Governance & Safety

The White House reportedly wants Anthropic to block all jailbreaks before rereleasing Fable 5, a standard outside security experts say no current model provider can guarantee.
SK Telecom sat at the center of Anthropic’s Claude Mythos export-control controversy after U.S. officials demanded access be revoked over alleged China ties.
JPMorgan reportedly cut off Hong Kong staff from Anthropic models, showing how quickly model-access rules are becoming a compliance issue for global companies.
World leaders at the G7 said they want American AI capabilities but fear U.S. companies or regulators can cut them off overnight.
Residential proxy networks are becoming a cyber threat as nation-state hackers route attacks through ordinary home electronics to mask their traffic.

🛠️ AI Tools & Products

MOSS-TTS Local Transformer v1.5 clones voices and speaks 30+ languages at 48 kHz in a local setup, with the Mosi playground available for testing.
Rumik launched silk mulberry 1.5, a multilingual voice model pitched as top-tier quality at more than 95% lower cost.
Midjourney released a big-batch draft mode for V8.1 that creates 24 lower-resolution images for half the price of a standard four-image job, then lets you vary favorites into full-resolution outputs.
Happy Oyster released HappyOyster 1.0, a real-time open-ended world model where you can create and explore worlds through text, voice, or images, chat with virtual companions, rewrite stories, move with WASD, parkour, wingsuit, raise virtual pets, and find hidden interactions across surreal environments like ocean ruins or dreamscapes — free daily credits until July 17.
Killed by OpenAI is Ben Hylak’s public graveyard for OpenAI products, features, and APIs labeled alive, killed, or zombie, complete with a dithered memorial-wall design and an $11 domain purchase that turned product deprecation into an actual cemetery joke.
Tacit Labs launched as an applied research lab for AI and biology, with Nicole Fitzgerald saying the team is building long-horizon evaluations that test whether AI systems can make linked decisions across drug discovery and development by combining human reasoning, biology foundation models, raw lab readouts, and lab infrastructure.
Orion is a computer-using agent for lab automation that controls scientific software through the screen and terminal instead of predefined APIs; the bioRxiv preprint and Ma Chang’s thread say it can learn tools like QuPath from tutorials, run multimodal microscopy and whole-slide-image tasks, and produced 52 cell-painting reports in a 100-hour run, with 22 judged plausible mechanistic hypotheses by human reviewers.
ABC released an open-source stack for bimanual manipulation, including ABC-130K robot demonstrations, simulation data, model training code, and evaluation protocols.
Beni is a two-wheeled self-balancing camera robot that can track subjects, traverse terrain, jump, flip, record, and is planned for Kickstarter.
NotebookLM kept trending after its more powerful Google AI Ultra experience fully rolled out, TestingCatalog spotted upcoming Personal Preferences and AI Editing for notes, users shared lecture-note workflows, and Similarweb estimated 252.9M May visits, up 289.78% year over year.

📊 Fundraising & Deals Roundup

Odyssey — $310M Series B for world models that simulate the physical world.
XDOF — $70M for robot training data collection and annotation infrastructure.
Convey — $38M Series A led by a16z for autonomous digital teammates.
Comand AI — €32M Series A for AI-native command-and-control software across NATO, with Saab as a strategic backer.
Ploy — $27M seed led by Y Combinator and First Round for autonomous marketing infrastructure.
NeuralTrust — $20M to expand an AI agent security platform.
Clair Health — $11M to build noninvasive hormone-tracking wearables.

🎬 Creative, Media & Demos

Vibe Jam 2026 named its winners, with a polished capybara delivery game taking first place; Pieter Levels later spotlighted the retro PS1-style winner, its Rio setting, fake in-game apps, and AI-generated soundtrack.
LTX-2.3 and its trainer give creators open audio-video generation plus LoRA training, while the Creative Lab gives practical effects like relighting and beard removal.
Google Flow sits in the creative stack as Google’s studio for video, images, and custom tools, with Flow’s X demo emphasizing cinematic workflows and storyboarding.
Dom is Hassaan Raza and Tavus’s working Knowledge Navigator-style assistant prototype with real-time perception, screen control, memory, personality, and tool use, reviving Apple’s 1987 personal-agent dream in working software.
Midjourney Medical was the day’s strangest creative-company pivot, with Midjourney, Nick Dobos, Nick St. Pierre, Sebastian Caliri, and Andrew Curran turning the scanner into a vision of cheap, frequent, full-body imaging: Caliri said a mini version already produced higher-quality-than-MRI images in under a minute at almost no running cost and argued the same ultrasound tank could ablate tissue or stimulate nerves, while Curran connected cheap scans to abundance and argued longitudinal imaging gets much more valuable once personal medical AI can reason over it.

💡 Industry Commentary & Analysis

Peter Wang argued reliable vertical agents should use filesystem and bash primitives instead of bespoke spreadsheet-like tools, because real files make workflows cheaper, rerunnable, auditable, and better at handling messy context.
Matt Pocock pointed out that agent skill systems need separate tiers for user-invocable, skill-invocable, and model-invocable skills, then released a skills update that cuts description-token costs by 63%.
mem0 argued the bottleneck for long-running autonomous loops is memory, because agents drift, repeat work, and reinforce early mistakes when durable semantic memory is treated as an afterthought.
Ravid Shwartz Ziv broke down IndexShare from GLM-5.2, a trick for making long-context models cheaper by reusing one small attention indexer across four layers; instead of making every token look at every other token, the indexer picks the most relevant tokens to check, then the next layers reuse that selection because the important tokens usually barely change, cutting FLOPs per token by 2.9x at 1M context.
Rahul said English-to-code systems should be evaluated like idea-to-correct-code machines, where risk-aware review and empirical verification matter more than reading every generated line; Boris Cherny agreed from the Claude Code side that the work now becomes guardrails, an advanced model, and a verifier loop, while the X trend captured the broader code-review debate in the Fable 5 era.
Yaroslav Bulatov flipped Artificial Analysis data to show the shrinking lag between closed frontier models and open-source performance.
Deli AutoResearch is a fully autonomous research framework defined by one SKILL.md file, with papers and a self-play story showing how it iterated a survey through 16 review rounds and 285B-parameter experiments.
Eric Jing of Genspark argued most SaaS is just a rigid UI over a database, and their AgentBase-style tools should let users talk directly to their data and generate the interface they need.
rsms posted about Playbit’s near-zero input-to-display latency work on macOS, where a rectangle followed the cursor with almost no visible lag after optimizations around Metal, presentation modes, vsync, and event handling, outperforming Figma on the same hardware.

That’s a Wrap

That’s 100+ stories, tools, papers, demos, and takes from one very wet day in AI. If you made it to the bottom, you now know more about ultrasound spas, robot data janitors, and Claude Code internals than anyone at your next “quick sync.” Use this power gently.

For the daily version, make sure you’re subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don’t have to.

See you tomorrow.

P.S: Know someone who’d find this useful? Forward this to them and tell them to subscribe here.

Everything That Happened in AI Today (Thursday, June 18, 2026)