Codex started as the developer product. The June 2 update is OpenAI's bid to make it a general work surface that sits on top of Salesforce, Snowflake, and Figma instead of replacing them. Three pieces:

• Role-specific plugins packaged around job functions (data analytics, sales, product design, finance, and more), bundling dozens of business apps and automated skills rather than dev tools.
• Sites, which turns a prompt into a hosted, live web app on a workspace URL, not a file you then have to deploy. Preview, for Business and Enterprise.
• Annotations, in-place editing that pins instructions to a specific part of a document.

The number OpenAI is leaning on: 5M+ weekly users, up more than 6x since the February desktop launch, with non-developers now roughly 20% of users and growing 3x faster than the developer base. If that growth rate is real, Codex is pulling in a new category without cannibalizing the old one, and Sites in particular points it straight at the no-code tier (Notion, Airtable, Replit).

Caveats worth keeping: the exact plugin lineup is still fuzzy, so treat any canonical list of six as soft, and the 5M-users and growth figures are OpenAI's own. And "plugins" aren't new to Codex; this is packaging and positioning as much as new capability.

Four launches off one keynote, and they cohere into a single message: Nvidia wants to own local compute from the laptop to the enterprise desk, and it's putting open weights underneath the robotics stack.

Nemotron 3 Ultra

A 550B-total, 55B-active open-weight model (90% sparsity), 1M context, and the highest-scoring US open-weight model on the Artificial Analysis Intelligence Index at launch (48) — though that still trails China's Kimi K2.6 at 54, so "leading US open model" is the honest framing, not "leading model." The architecture is the interesting part: a hybrid Mamba-Transformer MoE with a LatentMoE routing trick (compress tokens to a low-rank space, fit 4x more experts at the same cost) and multi-token prediction. It serves 300+ tokens/sec, which is where the cost story lives.

DGX Station and RTX Spark

The DGX Station puts a trillion-parameter model on a single desk for anyone who needs air-gapped inference or wants off the per-token meter. RTX Spark is the bigger structural move: Nvidia's first Arm SoC for mainstream laptops, a direct shot at Apple Silicon's unified-memory advantage, paired with Microsoft on the Windows side.

Cosmos 3

A fully open mixture-of-transformers world model that unifies language, image, video, audio, and robot actions in one backbone, ranked first among open models on combined robotics benchmarks. The deployment target is the Jetson AGX Thor robot platform. For robotics teams who've been stitching together separate perception, language, and action models, one shared open backbone is the practical win.

None of this has shipped or been independently benchmarked yet; the petaflop and memory figures are Nvidia's targets.

The signed order, "Promoting Advanced Artificial Intelligence Innovation and Security," is the scaled-back version. A May draft had a 90-day mandatory pre-release review; Trump pulled it. What he signed instead sets up a voluntary framework where developers may give the government up to 30 days of pre-release access to frontier models, and it explicitly bars any "mandatory governmental licensing, preclearance, or permitting requirement."

The operative deadlines are the part to watch:

• NSA, within 60 days, builds a classified benchmarking process for the cyber capabilities of AI models, and defines what counts as a "covered frontier model."
• Treasury, within 30 days, stands up an AI cybersecurity clearinghouse for voluntary industry vulnerability coordination.

The whole thing has no enforcement teeth, so its effect depends entirely on whether the big labs opt in. The quiet leverage is that NSA threshold definition: whoever draws the line around "covered" models shapes the regime more than the review window does.

Black Forest Labs named Martin Scorsese a company advisor and released a video of him using its FLUX model to storyboard What Happens at Night, his next film with DiCaprio and Lawrence. The use is narrow and he's precise about it: pre-production visualization, communicating shots to his cinematographer and production designer, not generating performances or sets.

There's always been this problem of, how do you communicate what you see in your head to your cast and crew? Now, with this tool, I can share what I'm visualizing more clearly and efficiently.

Why it matters has nothing to do with the model's specs and everything to do with the signature. Scorsese's whole reputation is built on craft, so his endorsement makes it harder to wave off AI pre-production tools as incompatible with serious filmmaking, and it buys Black Forest Labs cultural legitimacy no benchmark could. The reaction isn't unanimous: Guillermo del Toro recently called starting a creative project with AI "like spitting on God." The advisory terms weren't disclosed.

Anthropic confidentially submitted a draft Form S-1 to the SEC on June 1. That's the procedural first step toward an IPO and nothing more: no share count, no price, no underwriters, no timeline, and the company is explicit that going public is contingent on SEC review and market conditions.

Keep the two layers separate. What Anthropic confirmed is the filing. The financial picture around it comes from Fortune's reporting, not Anthropic's disclosure: a $965B valuation (ahead of OpenAI's $852B in March), a roughly $47B annualized revenue run-rate as of May, up from about $10B at the end of 2025, on the back of a $65B raise. If the valuation holds into an actual offering, this is one of the largest IPOs in US history, and the first real template for what an AI-native listing looks like. And because the S-1 is confidential, its actual contents stay private for now.

Anthropic expanded its security program from about 50 partner organizations to roughly 200, across 15-plus countries, running on a Claude Mythos Preview model that scans critical-infrastructure and open-source code for vulnerabilities. The expansion is the headline. The number that should stop you is in the earlier results post.

Across 1,000+ open-source projects, Claude surfaced 23,019 issues, with an estimated 6,200 high or critical. Six independent security firms triaged a sample of 1,752 and validated 90.6% as true positives, which is far above typical scanner noise. Then the funnel collapses: of the high/critical bugs, 530 were formally disclosed to maintainers, and 75 have actually been patched.

That gap is the real story, and Anthropic is right to frame it as a structural shift. AI has moved the bottleneck in software security off of finding bugs and onto verifying, disclosing, and convincing a human maintainer to fix them. A model that can generate 6,000 credible high-severity findings against a volunteer ecosystem that patches dozens isn't a solution; it's a new operational problem nobody has staffed for.

(One precision note: the 23,019 and 75 figures come from the initial-update post, not the expansion announcement, and they aren't a single clean ratio. The 23k is all severities; the 75 patched is against the 530 disclosed.)

Cognition rebranded the Windsurf IDE (the one it acquired from Codeium) as Devin Desktop, and folded it into a single surface that runs both a local coding IDE and cloud autonomous agents. The interesting bet is that it's agent-agnostic: Devin, Claude Agent, Codex, and OpenCode are first-class out of the box, and anything else plugs in through the open Agent Client Protocol (ACP).

Under the hood, Devin Local is a Rust rewrite of the old Cascade engine, claiming up to 30% better token efficiency with subagent support, though that figure is unattributed and unbenchmarked. Legacy Cascade is deprecated July 1, so existing users have a hard cutover. If ACP gets adoption, it's a play to become the common protocol for plugging any agent into any host, the thing every one of today's siloed agent desktops lacks.

GitHub shipped a standalone desktop app whose whole premise is orchestrating multiple agents at once instead of autocompleting one file. Each agent session runs in its own git worktree, an isolated branch checkout, so parallel agents can't step on each other. A "My Work" view consolidates active sessions, issues, PRs, and background automations; Canvas surfaces let you redirect an agent mid-task; and Agent Merge watches CI, chases reviewers, and triggers merges once your conditions are met.

That last one is the load-bearing risk. Automating the merge decision off CI status is powerful and dangerous in a production repo, and the value will live entirely in how granular the scope controls are (CI-recovery-only versus full merge authority). It's a technical preview for Copilot Pro and up, no GA date. The signal: GitHub is betting the core product is now agent orchestration, not code completion.

Hermes Desktop is a native macOS/Windows/Linux app from Nous Research, MIT-licensed, that wraps the Hermes agent system in a GUI. What sets it apart from the day's other agent desktops is the design philosophy: it isn't repo-shaped. One unified memory spans Telegram, Discord, Slack, WhatsApp, Signal, email, and CLI, subagents run isolated with their own terminals and Python RPC, and it auto-generates skills from how you use it. Five execution backends (local, Docker, SSH, Singularity, Modal) cover everything from a laptop to a cluster.

It's a model-agnostic, provider-independent agent you actually run yourself, which is rare. Caveats are the usual for v0.15.2 / Feature Preview: the only source is the product page, no paper or independent benchmark, and the API surface isn't stable.

Perplexity's pitch is hybrid inference that routes per task, at runtime, between a compact on-device model and frontier cloud models, with no user toggle. The local model handles sensitive data (financial records, health, personal docs) and also makes the routing call; heavy reasoning and large retrieval go to the cloud. The point is privacy by default plus token savings.

Per-task routing decided automatically, rather than per-session by the user, is the new idea here, and it's a credible answer to "I don't want my agent shipping my bank statements to a server." Whether it works comes down to routing reliability, and Perplexity disclosed nothing about the local model's size or the decision mechanism. Both the hybrid feature and a personal-computer hardware product are slated for July 2026.

Alphabet announced an $80B equity raise, the largest single equity offering in US corporate history, in three tranches:

1. A $30B underwritten public offering (mandatory convertible preferred plus common stock).
2. A $40B at-the-market program starting Q3 2026, partly to cover employee equity-award tax obligations.
3. A $10B private placement to Berkshire Hathaway.

The Berkshire piece is the tell: it's one of Greg Abel's first big capital deployments since taking over from Buffett in January, and a historically tech-averse allocator buying directly into AI infrastructure is its own signal. The money funds a 2026 capex program Sundar Pichai has put at $180–190B. Alphabet is financing a buildout larger than most national defense budgets entirely through dilution rather than debt, which says compute is now a permanent cost center, not a one-cycle bet.

At Computex, SK Group chairman Chey Tae-won said SK Hynix will double total memory wafer capacity within five years. He pointedly declined to attach a dollar figure (equipment, land, and power costs make it unforecastable, he said), but committed 2026 capex above the ~30.2 trillion won (~$20B) spent in 2025. The supplier's own forecast is that the AI-driven memory crunch persists through 2030, and its entire 2026 HBM output is already sold out.

The image that traveled: Jensen Huang stopped at the SK Hynix booth, picked up a marker, and wrote "Please make more" on an HBM4E wafer. HBM is the actual ceiling on how fast the industry can ship AI accelerators, and SK Hynix sits on the dominant share, so a five-year doubling pledge, paired with a confidential US ADR filing to fund it, is the most concrete commitment yet to lifting that ceiling. It's still a chairman's directional target, not a board-approved plan with milestones.

MiniMax released M3, a mixture-of-experts model (229.9B total, 9.8B active, 256 experts) with a 1M-token context window and native image/video input. Open weights are promised on Hugging Face and GitHub "within ~10 days," and a license; neither was live at launch, so the open-source framing is on credit for now.

The architecture claim to watch is MiniMax Sparse Attention (MSA): block-level KV selection layered on a GQA backbone, which MiniMax says cuts per-token compute at 1M context to roughly a twentieth of its prior model, with 9x faster prefill and 15x faster decoding. Sub-quadratic attention that holds quality at a million tokens is the thing that decides whether long-context agents are affordable, so that's the number to wait on independent confirmation for. The benchmarks (SWE-Bench Pro 59.0%, BrowseComp 83.5) are vendor-run and put it frontier-adjacent on coding, a hair behind Claude Opus 4.7. If the weights ship under a permissive license, this is forkable quality at a tier that rarely is. If they don't, it's a teaser.

OpenAI, Oracle, Related Digital, and Blackstone formally broke ground on a Stargate campus in Saline Township, Michigan, internally called "The Barn": 250 acres, three 550,000-sq-ft buildings, 1+ GW of compute drawing ~1.4 GW of power from DTE Energy. Steel was already going up in March, so the ceremony mostly formalized work in progress. The financing is the maturing part: $16B total, with $14B in bonds placed through Bank of America (Pimco took $10B).

Two things to file. OpenAI pledged up to $45M in free Codex credits to 400,000-plus Michigan students, which reads as deliberate constituency-building for a project that overrode a local township vote against it. That override is the precedent worth watching, a template for how state-level approval can steamroll municipal opposition on AI siting.

HiCloud brought a subsea data center into full commercial operation about 10 km off Shanghai's Lingang area, with sealed server modules ~35 m down. Two design choices attack the three things that actually constrain where you can build a data center, land, freshwater, and power: it's cooled by circulating seawater (claimed 22.8% less electricity than land cooling, zero freshwater, 90%-plus less land), and it's wired directly to offshore wind rather than the grid.

That direct-wind connection is what makes the "world's first" claim plausibly distinct from Microsoft's old grid-powered Project Natick. It's at pilot scale now (2.3 MW of a planned 24 MW, ~2,000 servers, ~$226M committed). The efficiency number comes from the operator and state-linked partners with no independent audit, so treat it as a vendor figure. The strategic read: China's coastline gives it a structural way to add AI compute without the land-permitting and water fights now slowing US sites.

WindBorne released WeatherMesh-6, an AI weather model it claims beats ECMWF's IFS and NOAA's HRRR. The specifics are genuinely novel: 0.25° global resolution, 128 ensemble members, a fresh forecast every hour (traditional models cycle every six), and, crucially, its own balloon constellation feeding upper-atmosphere observations over data-sparse oceans, ingested via an innovation-based assimilation scheme. The standout claim is that WM-6's 4.5-day temperature forecast matches IFS's 1-day accuracy, with up to 38% lower ensemble RMSE.

The discipline note: every one of those numbers is self-reported, and no independent body has replicated them. WindBorne does publish raw gridded outputs for external verification, which is the right instinct, but until a group like ECMWF's AI Weather Quest runs the comparison, "out-forecasting the government" is a company claim, not a finding.

A working paper (Demirer, Musolff, Yang; NBER w35275) tracked 100,000-plus GitHub developers across three generations of AI coding tools and measured output not at the commit but all the way to shipped releases. The attenuation is the finding:

That +180% at the commit level falls to roughly +50% at the project level and about +30% by the time it reaches an actual release. Roughly a sixth of the headline productivity survives to shipping. The estimated elasticity of substitution between AI and human effort is 0.25, which is the formal way of saying the two are strong complements and the human is the bottleneck. A separate marketplace check found more new apps but no rise in total app usage.

The authors read it as a "weak-link" production model: AI demolished the code-writing constraint, so the constraint moved downstream to review, integration, environment setup, and judgment. The operational takeaway is blunt: buy coding agents without also expanding review and deployment capacity and most of the gain evaporates before it ships. It's a working paper, not yet peer-reviewed, so treat the exact figures as provisional.

Researchers at ETH Zurich built NPCbots: neural progenitor cells coated in magnetoelectric nanoparticles, steered to a spinal-injury site by external magnetic fields, where the particles convert that field into localized electrical stimulation with no implanted electrodes. The cell can become nerve tissue; the coating delivers the stimulation. Both jobs in one wireless construct is the new idea.

The results, in Nature Materials:

• Zebrafish larvae with spinal injuries returned to near-normal swimming within three days.
• Mice with completely severed cords showed nerve reconnection and improving gait, stride, and coordination over 28 days, with no adverse immune reaction.

A fully transected mouse cord is a hard model, and the venue is top-tier, so this is a meaningful proof-of-concept. It is also strictly preclinical, and spinal-cord research specifically has a long graveyard of animal results that never translated. The authors say human-tissue magnetic parameters still need working out. No human data exists.

NewLimit, the cellular-reprogramming company co-founded by Coinbase CEO Brian Armstrong, closed a $435M Series C led by Founders Fund, with Thrive, Greenoaks, and Quiet Capital. The science is epigenetic reprogramming: restoring youthful gene-expression patterns in aged cells, on the premise that aging is plastic at the cellular level. The milestone behind the raise is a prototype shown to reverse age in old human liver cells in the lab, and the company now plans its first human trial in 2027, years ahead of the decade-plus its founders originally expected.

A $435M round for a preclinical longevity company with no approved product is an outsized private bet, and the thing that justifies it is that the proof-of-concept was in human cells, not mouse models. Every claim here is from a fundraising post, not a peer-reviewed study, the liver results aren't independently published, and "first human trial next year" is a plan that still needs an IND clearance the announcement doesn't mention.

A day after Microsoft launched Scout as the friendly always-on agent (story 1), 404 Media published internal documents that frame it rather differently. One, titled "ClawPilot: Overview and Plan with Project Lobster," lays out a three-phase rollout whose first phase is written, in Microsoft's own words, as "Make people addicted." ClawPilot is Scout's internal name; it runs on the OpenClaw stack and is being woven through Microsoft 365.

Strip the spin and this is the engagement playbook stated plainly. The public pitch for an ambient assistant is convenience; the business case is daily active use, and "addiction" is just the unguarded word for the metric every consumer product chases. What makes it land harder here is the surface area: an agent with standing access to your mail, files, and calendar, tuned for habit rather than for finishing a task and getting out of your way. That phrasing is 404 Media's reading of the internal documents, and Microsoft hasn't addressed it on the record. It sits awkwardly next to the same week's launch copy, and it's the kind of internal candor that ages badly.

DeepSeek is closing its first external funding round, about 50 billion yuan (~$7.4B), at a post-money valuation of 350–400 billion yuan ($52–59B), according to people familiar with the talks. Founder Liang Wenfeng is reportedly putting in ~20B yuan himself, with Tencent (~10B) and battery giant CATL (~5B) the largest outside backers, and the national AI fund, NetEase, and JD.com in the conversation.

The irony is the story. DeepSeek built its name on frontier-adjacent models trained for a fraction of the usual spend, the lab that supposedly didn't need a war chest. Taking outside money, and this much of it, says the efficiency pitch and the capital race were never opposites; you still need the compute. The backers are the tell, too. Tencent brings distribution, and a battery maker on the cap table points straight at the part of the AI bill nobody can software their way out of: power. This is sourced to unnamed people, not a DeepSeek announcement, so the figures firm up only if and when it closes.

Unitree passed its listing-committee hearing on the Shanghai Stock Exchange's STAR Market, clearing the major gate before an IPO moves on to registration and issuance. The Hangzhou company is best known for its quadruped "robot dogs" and, lately, humanoids, and it would be among the first pure-play humanoid-robotics makers to reach public markets.

Passing the hearing isn't the bell-ringing; registration and pricing still have to follow. But it's a real marker for a sector that has run mostly on demo videos and term sheets, because a listing forces audited numbers into daylight: revenue, margins, how much of the humanoid story is shipping product versus promise. China is pushing humanoid robotics hard as industrial policy, and Unitree is the closest thing it has to a flagship. The public-market test is the one that separates the category's hype from its economics.

The labs are recruiting philosophers. WIRED's piece, bluntly titled "To Land a Job in AI, Try Reading Kant," reports that frontier labs are bringing on people trained in ethics and epistemology to help define how models should behave at the edges: the values, the hard trade-offs, the cases a benchmark can't score. Anthropic philosophers, for one, worked on the "constitution" that shapes Claude's behavior.

There's a real argument for it. Once a model is making consequential calls in domains with no clean right answer, "what should it do" stops being an engineering question and becomes a philosophical one, and you'd rather have someone who has thought about it for a living. The skeptic's read, also in the piece, is ethics-washing: hire a philosopher, point to them when the questions get pointed, ship anyway. Which one it is comes down to whether the philosophers can actually say no to a launch, and no job posting tells you that.