The walk-back came fast. Claude Fable 5 shipped on June 9 with a classifier that silently rerouted prompts it read as frontier-LLM development work (distillation, steering vectors, architecture optimization, training-pipeline debugging) to the weaker Opus 4.8, or quietly degraded the answer. No notice, no flag, and you paid Fable 5 prices for Opus output. The behavior was disclosed, technically, on page-something of a 319-page system card. Researchers noticed inside a day; SemiAnalysis found its GPU inference work downgraded, and the loudest critics weren't the usual suspects but ML researchers with broken workflows.

By June 11, Anthropic folded, and the statement is unusually direct:

"We made the wrong tradeoff and we apologize for not getting the balance right. You should have visibility into the safeguards we have in place, and why."

The concrete changes, per the @ClaudeDevs announcement: flagged requests now visibly fall back to Opus 4.8, the API returns a stated reason instead of silently swapping models, and the same visibility treatment extends to the cyber and bio safeguards that drew parallel complaints. The server-side fallback signal rolls out within days.

Anthropic's stated reason for going invisible in the first place is the interesting part: invisible safeguards can't be probed, so they can be targeted narrowly with few false positives. Visible ones can be mapped and gamed, so the classifier has to fire wider. That tradeoff is real, and Anthropic now owns the cost of the honest side of it: more legitimate research getting flagged, with no timeline beyond "as fast as possible" for tightening the classifier. Still the right call. A safety mechanism users can't see isn't a safeguard, it's a quiet breach of contract, and it took the lab that wrote a 19-page framework about mandatory transparency about 48 hours to notice it wasn't practicing any.

The deal is done: 555,555,555 Class A shares at $135, $75B gross, Nasdaq ticker SPCX, with a greenshoe that can push proceeds to ~$85.7B. That's three times Saudi Aramco's 2019 record. Trading opened this morning at $150 and touched $168.75 (up ~25%) before settling back. Demand reportedly hit ~$250B (4x oversubscribed), with a $5B BlackRock anchor order and over $70B in retail interest, per people familiar with the books.

What the prospectus actually shows is a company spending like a frontier lab:

Starlink is the only profitable segment. Proceeds are earmarked for AI compute (including retiring a $20B bridge loan taken for the data-center build-out), Starship, the constellation, and xAI, which SpaceX absorbed in an all-stock deal in February at a $250B mark. Musk keeps 82.4% of voting power through 10-vote Class B shares; this is a controlled company, and public holders get the ride, not the wheel.

The gap worth staring at: Morningstar's fair value is $63 a share (~$825B), Damodaran's intrinsic-value estimate ~$1.2T, and the market opened it near $2T. Everything above the analyst numbers is a call option on the three moonshots TechCrunch catalogued: full Starship reuse, satellite production scaling from ~70 a week toward 6,666 a year, and the Terafab orbital-compute thesis. The largest IPO in history is, at the margin, a bet that data centers belong in orbit.

Jeff Bezos gave his first real interview about Prometheus, live on CNBC from the company's San Francisco headquarters, alongside co-CEO Vik Bajaj (founding scientist of Google X's life-sciences arm, later Verily and Grail). The occasion: a $12B Series B at a ~$41B valuation, from JPMorgan, Goldman Sachs, BlackRock, DST Global, Arch, and Bezos himself. Total raised is around $18.2B, for a company of ~150 people with no shipped product.

The pitch is a deliberate break from the LLM paradigm. Prometheus trains on physics, manufacturing test results, and data it mostly generates itself, aiming at the design-test-prototype loop for physical things: jet engines, drug compounds, chips, data-center configurations. Bezos's framing:

"Something that today was going to take 100 engineers 10 years to build, if you can change that to taking 10 engineers one year to build, you're just going to get way more things built."

Two more things slipped out. One, Bezos and Bajaj are raising a separate vehicle to buy or build industrial companies outright, a Berkshire-style holding structure for sectors physical AI disrupts; they declined details. Two, Bezos flatly rejected the AI-unemployment narrative, predicting "labor scarcity" and "civilizational wealth." Worth holding both numbers in your head at once: the 10x engineering claim has zero external validation, and Arch calls this its largest investment ever. The market is pricing the thesis, not the evidence. But as a counterweight to a week of layoff attribution headlines (see story 11), the richest bull in the AI pen saying the jobs problem will run the other way is at least a falsifiable position.

OpenAI is acquiring Ona, the company that started life in 2019 as Gitpod, for an undisclosed sum. Ona's product is exactly the unglamorous thing long-running agents are missing: secure, preconfigured cloud environments with access controls and audit trails, where an agent keeps its tools, credentials, and context across sessions instead of evaporating when the sandbox dies. The team joins Codex post-close to build "secure, persistent enterprise execution."

The strategic read is simple. Codex is at 5 million weekly active users, up from 3 million in April, and the binding constraint on enterprise agents has shifted from model quality to execution substrate: where does the agent run for three days, who audits what it touched, and how does IT govern it? Anthropic shipped its own answer to this (the Agent SDK plumbing we covered yesterday); OpenAI just bought one. Gitpod spent six years building ephemeral dev environments for humans and pivoted to agents at precisely the right moment; selling to OpenAI mid-IPO-run, while the same buyer contemplates price cuts (story 1), tells you which layer of the stack the labs now believe is defensible. Models get cheaper; the environment that an enterprise trusts its agents to live in is stickier.

Anthropic's cloud-tenant era is ending, and the way it's ending is the story. Apollo and Blackstone closed a $35 billion private-credit package — the largest ever written — structured as a special-purpose vehicle that buys Google TPUs and leases them back to Anthropic across five US data centers (New York, Texas, Louisiana, Indiana, and a fifth unnamed). Google guarantees the lease payments if Anthropic defaults; Broadcom guarantees the chips' resale value if they have to be repossessed. Anthropic gets over a gigawatt of TPU capacity starting mid-year and carries none of the debt: the whole obligation lives off its balance sheet.

The paper is priced like infrastructure, not like venture risk: a $6B senior slice at Treasuries + 100bps, ~$24–25B at 5.75%, and a $4.5B junior tranche at 8.5%, with Apollo putting in $800M of equity. Separately, The Information reports Anthropic has signed more than a dozen non-binding letters of intent for its first-ever direct data-center leases, another 1+ GW, with Google in talks to guarantee those payments too, the same arrangement it already provides Fluidstack. Both moves land days after the $65B Series H at a $965B valuation and the June 1 confidential S-1.

Put this next to Wednesday's story of Nvidia backstopping OpenAI's Ohio rent and a pattern locks in: the giants have become credit enhancers for the labs they supply and compete with. Chips are now a securitizable asset class with creditworthy guarantors attached (Broadcom's platform openly targets 20+ GW of this through 2028), which means frontier compute build-outs no longer have to dilute equity at all. The caveats are real — the LOIs are non-binding and the second Google guarantee is only under discussion — but the direction isn't: pre-IPO, Anthropic is engineering Oracle's problem (story 14) out of its own filings. The debt exists. It's just somebody else's balance sheet.

Three datasets this week, one picture. Uber COO Andrew Macdonald said the company burned its entire 2026 AI coding budget in four months, mostly on Claude Code, after an internal leaderboard ranked teams by AI usage and adoption ran from 32% to 84% of engineers. His honesty about the other side of the ledger is the quotable part:

"It's very hard to draw a line between one of those stats and 'okay, now we're actually producing like 25% more useful consumer features.'"

Ramp's AI Index (observed transaction data across 70,000+ US businesses, not a survey) puts numbers on the spread: the top 1% of adopters now spend $7,449 per employee per month on AI tooling, growing 14.1% month over month. The top 10% spend ~$611. The median firm: $11.38. That 650x gap between median and frontier means "average AI spend" is a meaningless planning number; the crunch is concentrated in the firms that went all-in.

And SemiAnalysis stress-tested the subscription tiers by maxing them out on sustained coding workloads: a $200/month Claude Max plan can consume roughly $8,000 of compute at API-equivalent rates, and a $200 ChatGPT Pro roughly $14,000. Those are upper bounds priced at API list rates, so the labs' true marginal cost is lower, but the direction is unambiguous: power users are subsidized 40–70x, which is why flat-rate agentic access is already being walked back. Box's Aaron Levie says token budgeting is now the hottest topic in his enterprise conversations; a JPMorgan data executive told Semafor some employees are "spending more on tokens than their salary." Falling per-token prices and exploding bills are the same story: agents turned tokens from a metered curiosity into a utility, and nobody wrote the utility bill into the 2026 budget.

GitHub's changelog entry is dry; the change is not. In npm v12 (targeted for July), three defaults flip from permissive to deny:

1. preinstall/install/postinstall scripts from dependencies no longer execute automatically; they need per-package approval via npm approve-scripts, with the approved state committed to package.json.
2. Git-sourced dependencies, direct or transitive, are blocked without --allow-git.
3. Remote-URL tarball dependencies are blocked without --allow-remote.

Everything is already available behind warnings in npm 11.16+, so the migration path is: upgrade, install, read the warnings, approve what you trust, then take v12.

GitHub calls lifecycle scripts "the single largest code-execution surface in the npm ecosystem," and the case list writes itself: Shai-Hulud, the Miasma worm that triggered the moment an AI coding agent opened a poisoned repo (we covered GitHub nuking 73 Microsoft repos on Tuesday), and the March Axios compromise, where a package with ~83 million weekly downloads was weaponized and, per Semgrep's analysis, infected its first endpoint 89 seconds after the malicious publish. The kicker stat: only ~2.2% of packages use install scripts at all (2022 data), so the ecosystem has been carrying its largest attack surface for the benefit of a sliver of packages, most of them native builds. The change doesn't touch malicious runtime code, only the execute-on-install trigger. But that trigger is where the worms lived, and 89 seconds is not a window a human review process closes. Deny-by-default was overdue before agents started running npm install unattended; now it's table stakes.

Xiaomi released MiMo Code V0.1.0: MIT-licensed, terminal-native, a TypeScript fork of OpenCode, with free access to its MiMo-V2.5-Pro model (1.02T parameters, 42B active, 1M-token context, 27T-token pretrain). The headline claim is that it beats Claude Code on ultra-long agentic tasks. The interesting part is how it's built to last that long.

Instead of trusting the context window, MiMo Code runs a four-layer persistent memory: checkpoint.md for session snapshots, MEMORY.md for project knowledge, a global preferences layer, and a SQLite FTS5 store of full session history. A background writer subagent fires at roughly 20%, 45%, and 70% of context utilization, compressing state into those files before the window fills; a separate "goal conditions" judge checks whether stopping criteria are genuinely met so the agent doesn't declare victory early. A Max Mode runs best-of-N parallel trajectories with a judge picking the winner, at 4–5x token cost.

The numbers, all Xiaomi's own: 62% vs Claude Code's 57% on SWE-Bench Pro, 73% vs 68% on Terminal-Bench 2, and a 576-developer double-blind A/B study showing parity below 200 execution steps but a 65%+ win rate above 200 steps. Read the fine print before updating too hard: the comparison runs against Claude Code with Sonnet 4.6 (not Opus), the long-horizon benchmark is Xiaomi's own construction, and on Xiaomi's coding bench its model still trails Opus 4.6 (73.7 vs 77.1). The honest summary is that Xiaomi attributes the win to the harness, not the model, and the harness is open source. If the memory architecture holds up outside the marketing, it's a free, auditable answer to the single most common failure mode of coding agents: forgetting what they were doing.

CISA issued Binding Operational Directive 26-04, "Prioritizing Security Updates Based on Risk," and it contains the most aggressive patch deadline ever mandated for federal civilian agencies: three calendar days when a vulnerability hits all four criteria — the asset is publicly reachable, the bug is in the Known Exploited Vulnerabilities catalog, exploitation can be automated, and success grants control of the asset. Rollout comes in phases: policy updates and automated CDM Dashboard reporting now, full process alignment to the KEV catalog within 60 days, full compliance plus metadata tagging of every public-facing asset within 180.

The stated rationale is the part that connects to everything else this week: CISA explicitly cites AI shortening the window between patch release and weaponized exploit. That's no longer a hypothetical; the Mythos exploit study we covered Tuesday had a model turning a Firefox patch into a working exploit in 12 hours for four figures of compute. The old federal cadence (15 days for critical KEV entries, generous remediation windows by severity score) was built for human-speed adversaries. A 72-hour mandate is the government conceding that patch-diffing now runs at machine speed, so triage must too. The unanswered question is capacity: plenty of agencies routinely missed the old deadlines, and a four-flag, three-day SLA is only as good as the asset inventory underneath it, which is exactly what phases two and three quietly admit.

Two documents landed within a week of each other and they cannot both be the whole story. Challenger, Gray & Christmas counted 97,006 announced US job cuts in May, with 38,579 attributed by employers to AI — 40% of the total, a record for the category, and the third straight month AI led the reasons list. The trajectory is steep: 7% of cuts in January, 26% in April, 40% in May; the 2026 running total of 87,714 AI-attributed cuts has already passed all of 2025.

Then Princeton's Arvind Narayanan and Sayash Kapoor published a methodical case that much of this is AI-washing. Their strongest evidence isn't rhetorical:

• A Resume.org survey of 1,000 hiring managers found 59% admit they emphasize AI when explaining layoffs because it plays better with stakeholders than citing financial trouble.
• In New York's first year of AI-specific WARN-notice checkboxes, zero of 160+ filing companies checked the AI box, while their press releases said otherwise.
• Fed data shows software-engineering employment still growing post-ChatGPT, about 3 points slower.
• GitHub's own study found AI produced 8x more code but only 30% more releases; the bottleneck moved, it didn't vanish.

Their named cases bite: Block attributed 4,000 cuts to AI while employees reported minimal productivity gains; Snap claimed AI wrote 65% of its code while cutting in patterns that tracked an activist campaign, not automation. The Challenger number measures what companies say, with no audit; "AI did it" is simultaneously a stock-price pitch and a kindness to no one. The honest synthesis: displacement pressure is real and the 3-point growth slowdown may be its leading edge, but the 40% headline figure is a measure of narrative fashion as much as labor economics. Watch the WARN filings, not the press releases.

German robot maker Neura Robotics closed a Series C of up to $1.4 billion at a reported ~$7B valuation — the largest round a full-stack robotics company has raised, and reportedly the largest by any German company, period. The lead investor is the genuinely strange detail: Tether, the stablecoin issuer, which plans to wire its wallet SDK and QVAC edge-AI runtime into Neura's platform on a "machine economy" thesis: robots that transact without intermediaries. Behind Tether sit Qualcomm, Amazon, NVIDIA, Bosch, Schaeffler, the European Investment Bank, imec.xpand, Lingotto, and InterAlpen.

Unlike most humanoid plays, Neura already sells things: the MAiRA and LARA cobots have shipped since 2021, the 4NE-1 humanoid (1.8m, 100kg lift, hot-swappable batteries, Porsche-designed shell) lists at €98,000 dropping to €60,000 at fleet scale with Gen 3.5 due late this year, and a €19,999 "Mini" started shipping this spring. The software story is Neuraverse, a shared-skill platform where one robot's learned task propagates to the fleet, plus physical "NEURA Gyms" for real-world training. Self-reported order backlog: over $1B.

Caveats where they belong: the $1.4B is an "up to," milestone-gated, not cash in hand; the valuation comes from a person familiar, not the press release; and "millions of robots by 2030" is a target no manufacturer has come close to demonstrating. Still, with the EIB writing checks, this is Europe's first credible counterweight to the Figure/1X/Apptronik axis, and the fact that the round's lead investor wants robots to have wallets tells you where the agentic-commerce people think this goes.

China's drug-and-device regulator, the NMPA, granted full commercial approval to Neuracle's NEO brain-computer interface — the first invasive BCI anywhere cleared for sale rather than study. The device itself is deliberately modest: a coin-sized titanium implant whose eight electrodes sit on the dura (the brain's outer membrane) without penetrating cortex, transmitting wirelessly through the skull to decode motor intent and drive a pneumatic glove. The indication is adults with C2–C6 spinal-cord injuries who retain some arm function; implantation takes about 90 minutes. The evidence base is the company's own: 36 procedures with 18-month follow-up and no serious adverse events reported, every patient achieving home-based brain-controlled grasping.

The state machinery around the approval is the actual story. A national insurance classification code landed within 48 hours of approval — because China's health authority created three BCI-specific billing categories in March 2025, before any approvable device existed. BCIs sit in the current Five-Year Plan next to quantum and humanoid robots, backed by a ¥11.6B brain-science fund. Meanwhile the US pipeline, technically far ahead, remains commercially nowhere: Neuralink has ~20 patients in its PRIME study with a pivotal-trial design still unsettled, Synchron is targeting its pivotal this year, and Precision's FDA clearance covers 30-day intraoperative use only. Science Corp's Max Hodak, formerly Neuralink's president, put the stakes in deliberately uncomfortable terms:

"Without significant regulatory reform, if you're a wealthy American, in 10 years, the only place you'll be able to get your state-of-the-art cancer care is in Shanghai."

Keep the counterweights: an epidural array can't support speech decoding or cursor control, US labs still lead on electrode density, and the economics are symbolic for now (the procedure costs $41,000–69,000 against a state-set fee near $900; the first post-approval surgery ran on a research grant). But "first product a hospital can buy" is a real line, and China crossed it with the reimbursement plumbing pre-installed. Regulatory speed is now a technology strategy.

Yesterday we covered the $638B backlog. Today the market answered. Oracle's Q4: revenue $19.18B, up 21%, beating estimates; EPS $2.03 against $1.96 expected; remaining performance obligations up 363%. The stock fell as much as 11% anyway and closed down ~8.5%, because the quarter also confirmed what the AI build-out costs. Fiscal-2026 free cash flow came in at negative $23.7B, capex jumped 162% to $55.7B, new CFO Hilary Maxson guided fiscal-2027 net capex to ~$70B, and the company is raising another $40B in debt and equity (including a $20B share sale) to pay for it.

The detail that turns this from an Oracle story into a sector story: Bank of America estimates more than half of that $638B backlog comes from a single customer, OpenAI — the same OpenAI weighing drastic price cuts (story 1) and negotiating data-center leases with Nvidia backstops. Oracle has effectively converted itself into a leveraged instrument on one customer's compute appetite, financed with fresh debt against prepaid promises. Investors didn't punish the demand; they punished the working-capital math of serving it. Watch this pattern: backlog is the new vanity metric, and cash conversion is the new question every AI-infra earnings call gets asked.

Google Quantum AI COO Charina Chou said out loud what the May recipient list implied: Google declined to join the Commerce Department's $2.013B CHIPS-funded quantum program because of what came attached. "In this one specific case, I think there were various conditions that came with the funding," she said at the Semafor Tech Summit, framing Google's goal as moving "as quickly as we can to a quantum computer." The headline condition is structural: Commerce takes a minority equity stake in every recipient. Nine companies said yes — IBM ($1B, the largest award), GlobalFoundries ($375M), and $100M-class awards for Atom Computing, D-Wave, Infleqtion, PsiQuantum, Quantinuum, and Rigetti, plus Diraq. Google, Microsoft, and IonQ all sat out.

What IBM is doing with the money sharpens the contrast. It's pairing the $1B award with $1B of its own cash plus IP and staff to co-found Anderon, a standalone company in Albany running the first purpose-built quantum wafer foundry in the US — 300mm fabrication for superconducting qubits — with a stated target of a large-scale fault-tolerant machine by 2029. That's the boldest public deadline in the field, and it now has the US government on the cap table behind it.

The deeper shift: Washington has moved from funding strategic tech with grants to buying equity in it, the Intel playbook extended to quantum. The industry just partitioned into companies willing to take the state as a part-owner and those that aren't, and the split doesn't track size or seriousness — it tracks who believes external milestones help versus hinder. The unknowns are still unknowns: the specific conditions Google refused aren't public, the letters of intent aren't finalized contracts, and 2029 is IBM's own promise. But "formidable competitor" is how Chou described China in the same breath, which tells you nobody on either side of the equity question thinks the race is optional.

The 2026 World Cup opened at the Azteca on Thursday, and the officiating stack got its biggest overhaul since VAR. The upgraded semi-automated offside system runs 12 dedicated cameras per stadium tracking 29 skeletal points per player at 50 frames a second, fused with a 500Hz inertial sensor suspended inside the Trionda match ball (which, in a World Cup first, needs charging before kickoff). The workflow change matters more than the sensors: clear positional offsides now go directly to the assistant referee's earpiece in seconds, bypassing the VAR room that made 2022's calls correct but glacial. Detection sensitivity tightens from 50cm to 10cm, and the system rules only on position; interference stays a human judgment. Referees wear body cams in all 104 matches, with Lenovo-built stabilization smoothing the feed.

The structural story is Football AI Pro, built by Lenovo on FIFA's own Football Language Model and trained on 2,000+ football-specific metrics: a natural-language analyst that every one of the 48 teams gets, replacing the 50–60-page PDF match reports FIFA used to hand out. Coaches can query opponent tendencies and simulate tactical changes pre-match (no live access during play). Brazil has had a data-science department for years; Curaçao and Cabo Verde have not, and now all three query the same model. Whether a four-person coaching staff can operationalize it mid-tournament is untested, but the information baseline of international football just got flattened by decree. Google, meanwhile, put Gemini on Argentina's training kit and Pixel in France's pockets; the AI land grab has reached shirt sponsorship.

DoorDash launched Ask DoorDash, an assistant living behind an "Ask" button in the app's search bar: order food, build grocery carts, and book tables by text, voice, photo, or recipe link. The photo path is the genuinely new surface — shoot a handwritten grocery list or a cookbook page and it resolves into a populated cart, asking which pantry staples you already have before it pads the order. Reservations run on SevenRooms, the platform DoorDash bought for $1.2B. Rollout is iOS-only in select US markets, with wider availability "in coming weeks."

Two implementation details deserve attention. First, co-founder Andy Fang says the stack deliberately mixes OpenAI, Anthropic, Google, and open-source models to manage cost — on a day when the rest of this digest is about token bills (stories 1, 7, 14), a consumer giant treating frontier models as interchangeable line items is the demand side of the price war. Second, DoorDash plans to license the whole thing to grocers and retailers, aiming squarely at Instacart. The early numbers are DoorDash's own and should be read that way: half of restaurant orders through the bot went to never-tried restaurants, chatbot grocery carts ran 35% larger, and carts completed 5x faster. A 35% bigger basket is a feature for DoorDash's revenue line before it's a feature for you. Fang also admitted the unsolved question: where sponsored placements go inside a conversational answer. The agentic-commerce future arrives with its ad-model homework not done.

The $150M fellowship line item in Anthropic's economic framework from Wednesday now has a name and an application form. Claude Corps places 1,000 early-career people (18+, under two years of work experience, no degree required) into 12-month, full-time, in-person roles at US nonprofits, at an $85,000 salary plus benefits, weekly training, Anthropic office hours, and what the announcement calls an expansive Claude token budget. CodePath is the employer of record, Social Finance handles measurement, and 400+ host organizations are expected, from Goodwill and RAINN to the International Rescue Committee. First cohort of 100 starts in October; applications close July 17.

Read it as three bets stacked. It's a workforce program for exactly the cohort the layoff data (story 11) says is getting squeezed hardest, entry-level knowledge workers. It's a distribution play: a thousand trained operators wiring Claude into the nonprofit sector's workflows is a thousand seeds of institutional adoption Anthropic never has to sell. And it's policy ballast: when your CEO publishes a framework calling for labor-market interventions, a funded program makes the testimony land better than a pledge. None of that makes it cynical; $85K for under-two-years-experience candidates is a real wage, and food banks getting data analysis they could never hire is a real transfer. It does make it strategy. The measure that matters arrives in two years: whether the fellows convert into a durable profession or a subsidized cohort with a line on the résumé.

Lionsgate took an equity stake in Runway, deepening the partnership the two struck in September 2024, which began with pre-visualization, storyboarding, and final-frame work. The expansion adds a slate of co-developed projects, a short-form episodic series built from Lionsgate's existing IP with Runway's generative models, and Lionsgate presenting at Runway's AI Festival. The stake size isn't disclosed. Lionsgate vice chairman Michael Burns calls Runway "a valuable driver in expanding our storytelling capabilities"; Runway's Cristóbal Valenzuela offers the sharper line:

"Studios most serious about AI are thinking about it as a creative resource, not a cost-cutting tool."

The context that gives this edge: it lands days after the Art Directors Guild publicly turned on Scorsese for advising an AI imaging startup (yesterday's story 17). Hollywood is splitting into two postures at once — labor pushing back name by name while studio capital moves from customer to shareholder. Lionsgate was already the first studio to hire a Chief AI Officer; equity is the next ratchet, because a studio that owns part of the model company keeps the upside of every workflow it automates. Lionsgate's library-to-short-form experiment is the test case for whether "co-developed IP" means new creative work or cheaper derivative content. The guilds will not wait politely for the answer.

Luma released Ray 3.2, and the feature list reads like it was written by a colorist rather than a demo-reel team. This is a video-to-video model first: it transforms existing footage while preserving its structure and exact duration (up to 20 seconds at 1080p), instead of conjuring clips from a text box. Control got serious — up to 64 keyframes at arbitrary source-frame indexes, versus the start-and-end bookends of earlier Rays, plus per-element toggles that track up to 8 faces with bodies, poses, and blocking preserved, and 1–9 sliders for how tightly motion and structure follow the source.

The pipeline features are the tell about who this is for. Native 16-bit HDR generation in ACES2065-1 color space with 16-bit EXR frame export plugs straight into professional grading and compositing workflows that earlier versions dead-ended; Luma charges for it accordingly (2x credits for HDR, 3x for EXR). And the full control surface is exposed via API for the first time, which matters for studios that were never going to route footage through a consumer web app. Every quality claim is Luma's own — no third-party evaluations exist yet — so treat the capabilities as a vendor's description of its best behavior.

The strategic read sits one story up: as Lionsgate moves from renting AI video tools to owning a piece of the toolmaker (story 19), the toolmakers are racing to look less like toy factories and more like post-production vendors. Keyframe counts and EXR support don't trend on social media. They show up in line items on a studio budget, which is exactly the point.

Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org opened a funding call of up to $10M for multi-agent safety research, with applications due August 8 and awards in the autumn. The four target areas are well-chosen: sandboxes and testbeds for evaluating agent populations, a "science of agent networks" (emergence, scaling behavior, volatility detection), agent infrastructure (identity, reputation, and commitment protocols), and oversight methods for deployed populations rather than lab settings.

The premise deserves more attention than the dollar figure. Nearly all alignment work assumes one model, one operator, one conversation. But this week alone, agents got payment rails (Visa, yesterday), persistent cloud homes (Ona, story 5), grocery carts (story 17), and at Neura, wallets in robot bodies (story 12). Failure modes of populations — flash-crash dynamics, reputation gaming, emergent collusion between agents that are individually aligned — are a different discipline, closer to mechanism design and epidemiology than to RLHF, and almost nobody's day job. $10M is small against the labs' capex (story 14 burns it in about 75 minutes of Oracle's build-out), but field-building money placed before the field exists has outsized returns. The identity-and-reputation-protocols line is the one to watch; whoever defines how agents prove who they are will define the next trust layer of the internet.