Everything That Happened in AI Today Thursday, June 18

OpenAI did the rarest thing in AI news: it turned a benchmark-shaped promise into 18 real pediatric diagnoses families had been waiting on.

Welcome to the Around the Horn Digest, everything that crossed our desk today, sorted into something you can actually use. The day’s center of gravity was unusually medical: OpenAI helped researchers reopen unsolved childhood disease cases, Google pushed AMIE from diagnosis into ongoing care, and GPT-5.4 started looking less like a chatbot and more like a lab collaborator. Meanwhile, the infrastructure layer was busy turning into a geopolitical choose-your-own-adventure: open-weight Chinese models got louder, Anthropic access drama kept echoing, and Amazon started eyeing Nvidia’s chip turf. Apparently “AI will change healthcare” now also comes with a memory shortage, robot dogs, and a pricing calculator. Let's get into it.

🆕 NEW From The Neuron

Sensor Tower's State of AI 2026 report shows AI assistants becoming the new shopping shelf: product discovery, ads, retailer referrals, and conversion are all starting to move into the chat box.
Adobe's creator survey says 87% of creators using creative AI see growth, but Corey's useful read is the tension underneath: output is cheap, trust and control are not.
Our guide to AI Skills, Projects, GPTs, Agents, Plugins, Connectors, Loops, and Codex Workflows is the map for anyone staring at the new agent alphabet soup and wondering which custom-AI thing they are actually supposed to use.

Around the Horn — Thursday, June 18, 2026

The big story today was OpenAI helping researchers from Boston Children's Hospital, Harvard, and OpenAI reopen 376 unsolved rare pediatric genetic-disease cases with o3 Deep Research. The system surfaced evidence-linked candidate diagnoses, then human experts reviewed the gene variants under clinical genetics standards and confirmed them through certified labs.

That process produced 18 new diagnoses, a 4.8% additional yield in cases that had already gone unsolved. That number sounds small until you remember the denominator: families who had exhausted ordinary diagnostic paths. For them, a few percentage points is not a leaderboard bump. It is a name for what has been happening, a path to care, and maybe the first useful answer after years of no answers.

The broader signal is that AI in medicine is moving from demo cases toward evidence assembly: reading the literature, connecting genes to symptoms, ranking variants, and giving specialists a better starting point. It is not replacing the lab or the doctor. It is compressing the search space so the people with medical authority can spend less time hunting and more time deciding.

🏆 TOP 5 NEWS (Around the Horn)

Google advanced AMIE from diagnosis into disease management, with Google Research and Google for Health saying the Gemini-backed system matched primary care physicians on complex ongoing-care conversations while improving treatment precision and guideline alignment.
Z.ai launched GLM-5.2 as an MIT-licensed open-weight frontier model (anyone can use or modify it) with 1M context (very long input memory), two reasoning modes, and stronger coding/agentic performance; it also appeared on Hugging Face, got more launch detail from Z.ai, the technical report, and the blog, drew major Agent Arena, chart, Artificial Analysis, and Vals AI results, got a technical read from Sebastian Raschka and his GLM-5.2 architecture note, and became easier to run locally through Unsloth docs, GGUF weights, Unsloth's launch post, plus KVCache.AI's memory analysis and KV Cache Calculator.
Anthropic reported Claude Opus 4.7 completed shared robotics tasks more than 18-37x faster than human teams from the original Project Fetch experiment while producing almost 10x less code; the earlier Anthropic post showed Claude helped non-roboticists program a robot dog faster and complete more tasks.
Noam Shazeer, co-inventor of the Transformer who left Google to build Character.AI and returned to DeepMind through a $2.7B deal, is reportedly joining OpenAI to work on new model architectures and evolve the Transformer itself (The Information, short link).
Amazon is reportedly in talks to sell its custom Trainium3 AI chips for other companies' data centers, turning its internal Nvidia alternative into a direct accelerator-market play.

Honorable Mentions

Google DeepMind published an AI Control Roadmap for securing agents that may misread instructions or over-pursue goals, with Sébastien Krier sharing the roadmap PDF and the related Three Layers of Agent Security paper.
OpenAI reported GPT-5.4 helped drive a medicinal-chemistry workflow from literature review to validated lab result, proposing an improved Chan-Lam coupling that human chemists tested across 10,080 reactions and hand-validated in 14 representative cases.
Waymo recalled nearly 4,000 robotaxis after at least 13 incidents in which vehicles drove into highway sections closed for construction in Phoenix and San Francisco.
FERC ordered six regional grid operators to justify or overhaul rules within 60 days for connecting massive data-center loads quickly while allocating costs and protecting grid stability.

🍪 TOP TREATS TO TRY

Cursor moves your local agents to cloud machines so they keep working after your laptop closes, accept phone prompts, run in parallel, and return pull requests with demos via /in-cloud snapshots -- no pricing details in the batch.
Vercel Eve gives developers an open-source production-agent framework with durable execution, sandboxed compute, approvals, channels, tracing, and evals (tests for agent behavior), while Vercel's post framed it as a Next.js-like file structure for agents -- open-source.
Open Design gives you a local-first Claude Design/Cursor Design alternative for pointing, commenting, marking up, editing, capturing, and remixing UI designs with your own keys and local model paths, while a second Open Design post framed it as a canvas-based Claude Design alternative with design-system plugins, Claude Code sync, and tool connections (GitHub) -- open-source.
Google Vids turns slides into avatar-led videos with digital presenters, branded avatars, and voiceovers in 24 languages -- free to try.
Runway Recipes packages production-ready generative media workflows into one API call, including product ads from images and product swaps inside videos -- no pricing details in the batch.
Factory AutoWiki turns a codebase into structured, browsable engineering docs that update on every push, with example wikis for 100 popular open-source repos -- no pricing details in the batch.
Kimi Work Goal Mode runs your desktop agent 24/7 on long-horizon, multi-step workflows until the task is complete -- no pricing details in the batch.

🏢 Big Tech & Major Companies

OpenAI is reportedly preparing GPT-5.6 and GPT-5.6-Pro, with TestingCatalog saying the family could arrive as soon as next week and Chetaslua sharing early GPT-5.6 Pro web-dev outputs that showed stronger understanding but 20-40 minute generation times and still-rough frontend execution.
OpenAI committed $600K to the Rust Foundation through Platinum membership plus extra maintainer support, saying Rust has become critical to Astral's toolchain and Codex work while Predrag, maintainer of cargo-semver-checks, joins the Foundation board as OpenAI's representative.
OpenAI Developers hosted "Women who Code(x)," where participants shipped task agents, personal guides, and real projects with Codex.
Anthropic added Enterprise-Managed Auth for Model Context Protocol connectors (the standard that lets agents plug into outside tools and data), letting admins centrally authorize tools through an identity provider across Claude chat, Claude Code, and Cowork, with Okta beta support plus connectors from Asana, Atlassian, Canva, Figma, Granola, Linear, Slack, and Supabase (announcement, MCP details).
Google DeepMind is hiring Forward Deployed Engineers to embed with strategic partners, build evals, unblock Gemini integrations, optimize GenAI workloads, and feed high-fidelity partner signals back to model teams, with a Google Careers salary range of $174K-$253K plus bonus and equity.
Cerebras announced that Google's Gemma 4 multimodal model is in private preview on Cerebras Inference at more than 1,500 output tokens per second, roughly 15x faster than Claude Haiku at comparable quality, unlocking faster screenshot-to-insight, document-summary, and UI-patching workflows.
Apple may raise iPhone prices because AI-driven DRAM and NAND shortages have quadrupled memory costs since last year, a supply squeeze CEO Tim Cook reportedly called "unavoidable" and "unsustainable."
Intel rose 10% after President Trump said Apple agreed to design and build chips with Intel in the U.S., adding to earlier foundry momentum from Nvidia and Elon Musk's Terafab project.
ByteDance is reportedly Microsoft's biggest AI customer, on track to spend more than $1B a year mostly on OpenAI models via Azure, despite Washington treating Chinese AI companies as a threat and OpenAI/Anthropic refusing direct China sales.
Epic/Unreal Engine shipped Unreal Engine 5.8 with advanced terrain, vegetation, lighting, MetaHuman, and Model Context Protocol plugin updates for faster world and character creation, with Unreal's X post pushing the launch to developers.

💼 AI Productivity, Labor & Economics

Elena Verna argues AI is creating a "Mom-and-Pop SaaS" era where domain experts such as teachers, coaches, recruiters, and small-business owners can build and sell specialized software for tiny markets without VC funding or massive scale.
Epoch AI argued that measuring AI's ability to automate AI R&D needs a sharper task map than broad occupation stats, and its full piece proposes an O*NET-like taxonomy of 60+ frontier research tasks rated for automation potential.
Viktor positions itself as an AI employee inside Slack and Microsoft Teams that connects to 3,200+ tools for reports, dashboards, code, and campaigns; Fryd framed the Teams launch as distribution through existing work chat rather than a new prompting app.
Fuse is building robots for skilled trades such as electricians and plumbers, opening a Birmingham training center, and planning to use real physical data from owned worksites for fast deployment while hiring a founding robotics team.

🤖 AI Agents & Infrastructure

Cloudflare opened Agents SDK primitives as a runtime any framework can target and launched Flue as the first compatible framework, adding durable execution through Fibers, dynamic workflows, durable filesystem, secure code execution in isolates, one-click deployment, and dashboard state streaming; Cloudflare's launch post said agents are rolling into the dashboard.
Cua launched background Linux computer-use agents that can drive real desktop apps through command-line or tool-connector interfaces while the user's desktop remains usable; the open-source stack includes Python SDK, cua-driver, sandboxes, and benchmarks, while Cua's technical post explains how it reads app structure, sends inputs, shows a separate agent cursor, and recovers across different app toolkits.
Aravind Srinivas argues context graphs will be the best way for businesses to deploy agentic harnesses because they create a self-improving "god-mode view" over tacit knowledge currently scattered across Slack, call notes, and SaaS tools.
Hyperagent, built by the Airtable team, showed three Gemini Omni API demos: a landscaping proposal from a lot video, an animated professor explaining dashboard findings, and an 8-bit platformer-style morning briefing generated from calendar, email/chat, and market intel.
DeLM turns decentralized parallel agents into reusable progress by letting them claim subtasks from a shared queue, read verified context, reason locally, and write compact updates, delivering up to 10.5 percentage-point gains on SWE-bench Verified (a test of whether models can fix real GitHub bugs) and about 50% lower task cost; VentureBeat highlighted the Stanford framework, and the paper was referenced as arXiv:2606.10662.
SkillMigrator stores web-agent skills as transferable interaction patterns tied to page-layout sketches so agents can reuse skills across sites with similar structure, reducing LLM action count by 8-10% on WebArena and Mind2Web; DAIR.AI highlighted the layout-based routing angle.
Compositional Skill Routing decomposes complex agent requests into atomic subtasks, retrieves a matching skill for each with a search index, and composes executable plans with a dependency-aware planner; elvis highlighted the paper and its CompSkillBench benchmark of 300 multi-skill queries over 2,209 real skills.
Covenant said it is competing on training infrastructure, with PULSESync reducing 7B-parameter reinforcement-learning weight-sync communication by roughly 100x losslessly because most updates disappear after the BF16 cast (a rounding step that stores model numbers more cheaply).
Nous Research launched Teams in Nous Portal, letting organizations share one credit pool under a billing owner with per-member usage visibility, spend caps, invitations, removals, and multi-team account switching.
Google Cloud introduced the Open Knowledge Format as an open, vendor-neutral way to package enterprise context into portable markdown files with YAML frontmatter (metadata at the top of a file); the GitHub directory includes the spec and samples, while Google Cloud Tech framed it as reusable agent context that travels across tools.
Apodex launched a self-evolving heavy-duty solver for hard research questions with no existing answer, with the product site positioning it around deep research and verifiable steps while Sheri Yuo argued ordinary best-of-N/rejection-sampling loops break down when evaluation criteria do not already exist.

💻 AI Coding & Developer Tools

HumanLayer launched an agentic IDE and collaboration platform for engineering teams, while HumanLayer gives teams task management, versioned artifacts, worktrees (separate code workspaces), Research-Plan-Implement workflows, and real-time human-agent collaboration, free for small teams and $100/user/month for Pro with BYOK (bring your own API key) and unlimited usage.
Poolside released Laguna M.1 weights on Hugging Face under Apache 2.0, with Laguna M.1 offering 226B parameters and 256K context for long-horizon work; pool is Poolside's terminal/editor coding agent with editor integration, slash commands, Model Context Protocol tools, AGENTS.md instruction files, rewind, and non-interactive execution.
ML Intern open-sourced an agent that automates the post-training research loop across paper research, citation walking, dataset handling, GPU-sandbox training (training runs on isolated graphics-chip machines), evals, and iteration, and Lewis Tunstall said it is out of beta after 12K+ uses, 300M+ generated tokens, GLM-5.2 support, local models, dataset uploads, YOLO autonomous mode, and deep Hugging Face integration.
Codex CLI 0.141.0 added authenticated end-to-end encrypted relay channels for remote executors, cross-platform remote execution, per-thread Model Context Protocol activation for selected executor plugins, cached tool search, lower latency/memory in tool-heavy sessions, and a bounded prompt-image cache.
OpenAI Codex can run the App, CLI, and SDK with open-source models through advanced settings and local providers, not only OpenAI models (official docs, short link).
Claude Code 2.1.181 added /config key=value, a CLAUDE_CLIENT_PRESENCE_FILE env var to suppress mobile pushes, Apple Events sandbox opt-in, prompt-caching and network-drive fixes, Bun 1.4, improved streaming, subagent-panel updates, and system-prompt changes.
Mistral AI is reportedly preparing a Code section for Vibe, an Apps section for deployable shareable AI artifacts, and a desktop app (TestingCatalog).
Interface Craft is Josh Puckett's lifetime-access library for high-care product design with frameworks, walkthroughs, interactive videos, skill files, and Claude Code/v0 collaboration methods; Josh Puckett also launched Means & Methods, a 100+ topic interface-design collection with 11 chapters and interactive examples.
Emil Kowalski published emil-design-eng and review-animations skill files for building and reviewing animations across Claude Code, Codex, Cursor, and related tools; his launch post says the two skills complement each other and were inspired by Cursor's code-quality review skill.
Unreal Engine open-sourced Lore, giving binary-heavy teams a version-control system with content-addressed storage, Merkle-tree repository state, immutable revision chains, deduplication, cryptographic integrity, lightweight branches, and sparse workspace hydration.

🔬 AI Research & Models

Artificial Analysis said benchmarking Claude Fable 5 on its full Intelligence Index cost about $6.2K, the most expensive model it has run, driven by 2x flagship pricing, heavy cached-token usage, and reasoning usage.
TestingCatalog reported Microsoft is evaluating GLM, MiniMax, Kimi, and other open models for Copilot Cowork to make models interchangeable, separate harness from model, reduce cost, and explore local execution.
Opus Magnum Bench tests frontier models on Zachtronics puzzle-game tasks through a Python REPL (a text command interface) instead of vision, probing spatial reasoning, rotation, concurrency, and optimization; Scaling01 argued GLM-5.2's strong result on a fresh benchmark is more meaningful because it is less likely to reflect benchmark overfitting.
LoopCoder-v2 showed a parallel loop transformer code model gains heavily from one internal rethinking loop, lifting SWE-bench Verified from 43.0 to 64.4 and Multi-SWE from 14.0 to 31.0, while later loops regress because repetitive updates and positional mismatch outweigh refinement gains; Rohan Paul summarized the same diminishing-return effect.
Looped World Models introduce looped architectures for world modeling, iteratively refining latent environment states through a parameter-shared transformer block with adaptive computation and up to 100x parameter efficiency; alphaXiv highlighted the paper's iterative latent-depth scaling axis.
Variable-Width Transformers argue models may waste capacity when every layer has the same width, with a wider-at-the-start-and-end, narrower-in-the-middle design that reportedly beats parameter-matched dense and mixture-of-experts models while using fewer FLOPs (compute operations) and less KV cache (the memory used to remember long context).
Latent Thought Flow moves reasoning into continuous latent trajectories instead of fixed token-by-token chains, using continuous GFlowNets (a training method for sampling good solution paths) so models can stop early on easy problems and spend more compute on hard ones.
Sakana AI introduced Doc-to-LoRA and Text-to-LoRA, hypernetwork methods that generate task- or document-specific LoRA adapters in under a second so models can internalize long documents or specialize without expensive fine-tuning or heavy context stuffing; Sakana also linked earlier work and discussion through this post, a DLCT recording, and ML Collective.
Liquid AI released LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M, two multilingual retrieval models for 11 languages with as-low-as 1.5 ms enterprise-stack latency, under 2 ms GPU latency, and under 10 ms CPU/laptop/edge latency; the technical blog adds that the embedding version prioritizes speed and tiny indexes while ColBERT adds word-level matching for higher accuracy.
Future Probes research says latent LLM features can predict future behaviors better than final-answer probes, enabling Future Probe Controlled Generation, a text-level steering method that samples candidate sentences and selects the one most likely to produce the desired later behavior; the project page includes paper links and interactive demos.
Trust Region Masking, also listed as an ICML 2026 poster, masks entire long-horizon LLM reinforcement-learning sequences when they violate derived trust-region bounds, aiming to reduce training mismatch and make model improvement more stable.
NEURRATOR generates natural-language descriptions of what individual neurons or neuron groups appear to encode during visual perception, mapping mouse visual-cortex spike trains into visual-language model space to narrate single-cell semantics.
Toshitake Asabuki shared a bioRxiv preprint asking how recurrent neural networks can combine independently learned computations without training on their combinations, with co-authors Osako Yuma and Aineias.
Jacob X. Li argues current agents handle new domains shallowly and proposes "Machine Studying" as the problem of developing real expertise from a document corpus alone, measured as a shift in the quality/cost curve; Jacob Li's thread framed StudyBench as a way to test whether agents can study like people.
Sonia Joseph argues world models and interpretability are the same causal-discovery problem: foundation models learn latent variables and dynamics for velocity, object permanence, causality, and physical interactions, and interpretability should identify those internal dials.
SemiAnalysis explained wide expert parallelism for Mixture-of-Experts model deployments (models that activate only part of themselves per request), where expert weights are distributed across multiple GPUs so each GPU loads only a small slice of weights, boosting memory bandwidth, throughput, performance per dollar, and performance per watt.

🏛️ AI Policy, Governance & Safety

Axios reported the Trump administration's AI team is forming through ad hoc coordination among Commerce, Treasury, the Chief of Staff's office, and National Economic Council staff, while Axios separately argued the administration has created a shadow AI policy through export controls, voluntary frameworks, executive actions, and procurement rules despite promising deregulation.
Dean Ball, who helped shape the Trump administration's early AI policies, is joining OpenAI to lead a Strategic Futures team focused on frontier AI policy, risk, governance, and institutions; Chubby argued the move gives OpenAI a government-relations advantage while Anthropic faces U.S. government issues.
Kimmonismus argued Anthropic's access cuts have produced a major PR win for open source and especially Chinese models like GLM-5.2 because companies and governments now want sovereign models they can run locally without sudden platform cutoff risk; TestingCatalog and SynthwaveDD added that Anthropic's international managing director reportedly said at a Seoul press conference that Fable 5 access was expected to return in the coming days.
Guardrails Alliance is a tech-worker-backed super PAC trying to raise $15M this cycle for pro-AI-safety legislation, positioning its small-donor "people in the trenches" base against Big Tech's $100M+ Leading the Future PAC.

🛠️ AI Tools & Products

fal released LTX 2.3 LoRA Trainers (small custom add-ons that tune a media model for a specific style or task) with 23 endpoints for custom media generation and editing, then posted direct trainer links including audio-to-video, video-to-video, image-to-video, spatial outpainting, and text-to-audio, with a follow-up fal thread bundling the suite.
xAI said Grok TTS leads Vapi's blind Humanness Index at 96/100, only 4 points below a real human voice, and Vapi is running the live blind-voting voice-model leaderboard.
Ludeo turns gameplay moments into accessible, shareable, ready-to-play experiences for game developers, with no pricing details surfaced in the batch.
Claude Design can now import design systems from GitHub, design files, or raw uploads, validate outputs against brand rules, and sync design-to-code handoffs while lowering token usage for long sessions.
This Google Vids project link appeared to be private or auth-gated in the pasted context, with only the generic "Hello, Grant. Let's start creating" interface text available, so it needs transcript or access before editorial use.
LM Studio previewed private frontier-scale inference by running Kimi K2.6 across four Mac Studios and streaming secure access through LM Link to a MacBook and iPhone.
Unreal Engine released the MetaHuman Animator Markerless Motion Capture plugin on Fab, letting creators generate full-body animation from ordinary video with no suits, markers, mocap stage, cloud processing, or credits; Unreal's post added the launch announcement layer.
Perplexity Brain is a research-preview memory system for Perplexity Max subscribers that builds context graphs from tasks and sources to improve correctness, recall, and cost on context-heavy work (blog).
Linear's agent-written project updates draft polished summaries from issue activity and linked Slack discussions.
John Capobianco demonstrated a custom vision agent that captures webcam frames, uses Gemini plus MCP for reasoning and tool calling, applies style transfer through Nano Banana while preserving identity, turns the result into an 8-second Veo 3 video, and supports real-time ASL interpretation; related YouTube description links included GEAR resources and a GitHub repo.

🤖 Robotics & Embodied AI

Mahi Shafiullah introduced "Do as I Do," a method for mapping chaotic real-world human videos to dexterous robot actions so models can learn from abundant internet video instead of only clean teleoperation data.
Harsh Gupta introduced LUCID, which learns embodiment-agnostic intent models from unstructured human videos and pairs them with a generalist sensorimotor policy for zero-shot dexterous manipulation across tasks, objects, scenes, and robot embodiments with no real robot data.
Qingxu Zhu showed MindOn robots using one model across humanoid and dual-arm bodies to complete a shared logistics task, trained only on human-centric learning rather than robot-collected data.
Sangdoo Yun and team argued VLA models do not need per-task retraining: a frozen retrieval-conditioned policy can handle new tasks by adding cheap human-hand demos to the retrieval pool, with gains on PushT, RoboTwin, and real robots.

📊 Fundraising & Deals Roundup

DeepSeek — $7.4B funding round valuing the company above $50B, reportedly with a no-poaching clause for investors.
Baseten — $1.5B raise at $11B-$13B valuation for cheap model inference.
Manus — $2B reported buyback effort from original Chinese investors after Meta's deal was ordered reversed.
General Intuition — $300M raise in talks at around $2B valuation for embodied AI/world models.
HIVE — $220M Bell sovereign cloud contract using 2,304 NVIDIA Grace Blackwell GPUs.
Prem AI — $100M Series A in progress at a $500M valuation for private enterprise AI.
Gradial — $65M Series C at a $675M valuation for agentic enterprise marketing (company post).
Telepatia — $42M total, including a $33M a16z-led Series A, for doctor-facing AI across Latin America.
Architect Labs — $24M seed to build AI systems for designing and provably verifying custom chips, with more context from Reuters, the company blog, careers, short link, and Architect Labs on X.

🎙️ Interviews, Panels & Podcasts

Yacine Mahdid hosted a discussion with Hugging Face's Joel Niklaus about synthetic pretraining data, with AVB adding that Niklaus's Synthetic Data Playbook inspired the text-albumentations library (short link).
Andrew Ng's VocalBridge course sits between courseware and practical agent tooling, focused on voice agents, outbound calls, live transcripts, and evaluation.

💡 Industry Commentary & Analysis

Ethan Mollick observed that if leaked financials hold, OpenAI may already be profitable serving customers at 40%+ gross margins, while training remains extremely expensive and automating AI research could make training efficiency the bigger lever (FT reporting).
The Pragmatic Engineer argues Meta's AI push has turned a once high-autonomy engineering culture into a demoralized cost center by moving thousands of engineers into repetitive labeling, tracking developer activity, tying performance to token counts, and weakening critical infrastructure.
12 Grams of Carbon argues HTML may be a surprisingly flexible medium for agent-generated visual work, including graphics and video experiments built with Claude Code.
Matthew Gault uses Adrian de Wynter's Age of Empires II argument to warn that chat interfaces make people over-attribute human-like traits to systems whose primitives can exist in silly, non-sentient environments.
Matthew Berman shared a practical experiment with LLM-as-a-judge loops, arguing soft stopping criteria such as "simple enough" or "fast enough" can guide iterative improvement on open-ended tasks even when there is no deterministic answer key.
Qiuyang Mang and co-authors found that on a two-week AtCoder Heuristic Contest, AI coding agents sprint early and plateau within 24 hours while top humans keep improving for days; the full blog and Frontier-CS platform frame repeated-sampling Elo as a way to distinguish brute-force tries from true long-horizon strategy.
Constructor theory reframes physics around possible and impossible transformations instead of only equations of motion, with ArXivIQ explaining the broader technical context and the arXiv paper proposing tests around non-classical mediators, irreversibility, life, and probability.
This X trending page was included in the source batch but is not a stable post, thread, product, paper, or announcement; treat it as a pointer to find the actual surfaced post before using it editorially.

Previous Around the Horn Digests

Catch up on everything you missed:

Wednesday, June 17: SpaceX reportedly pushed deeper into AI coding with Cursor, CoreWeave trained DeepSeek-V3 in two minutes, and Anthropic met the White House.
Monday, June 15: Anthropic's Fable and Mythos fight spilled into policy and markets, Salesforce bought deeper into agents, and chip deal rumors got spicy.
Thursday, June 11: OpenAI acquired Ona, Anthropic faced a Claude Fable backlash, SpaceX priced a record IPO, and agent tools kept multiplying.
Monday, June 8: Apple rebuilt Siri at WWDC, OpenAI confidentially filed for an IPO, Anthropic showed Mythos exploiting fresh flaws, and NVIDIA expanded its AI factory push.
Monday, June 1: NVIDIA turned the PC into an agent computer, Anthropic filed confidentially for an IPO, MiniMax released M3, and Bernie Sanders proposed public ownership of AI labs.

That's a Wrap

That's 90+ stories and source links from today alone. If you made it to the bottom, you now have enough context to explain rare-disease AI, open-weight model politics, and why your next product demo may have an avatar presenter.

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 (Friday, June 19, 2026)