AI Daily Digest

Friday, June 19, 2026

6,305 words · All issues

Top items

  • Day seven of the Fable/Mythos shutdown: Anthropic and the Trump administration remain deadlocked over export controls that took Claude Fable 5 and Mythos 5 offline; the “jailbreak” turns out to be “fix this code,” and the White House demand to block all jailbreaks is widely called impossible.
  • AI CEOs at the G7 in France: Amodei and Hassabis pushed a US-led AI coalition excluding China; Altman warned against concentrating power — three lab heads agreed governance is urgent but disagreed on form.
  • Washington weighs government equity stakes in AI firms (Semafor), as Pew and Johns Hopkins surveys show rising AI use but falling trust and strong demand for a “right to a human.”
  • Midjourney pivots to medicine with a full-body ultrasound “Scanner” and “Midjourney Spa” plan — 50,000 units and ~1B scans/month by 2031.
  • Noam Shazeer leaves Google for OpenAI to lead architecture research, two years after Google paid $2.7B to rehire the Transformer co-author and Gemini co-lead.
  • ChatGPT market share falls below 50% for the first time; Anthropic Claude Code study finds domain expertise, not coding skill, drives agent success.

Anthropic vs. the U.S. Government (Fable/Mythos shutdown)

Background and the shutdown. Claude Fable 5 and Claude Mythos 5 — Anthropic’s newest, top-tier frontier models — went live for only a few days before the White House suspended foreign access via an imposition of export controls at 5:23pm on a Friday, citing national security. (Zvi’s account, corroborated by Anthropic’s own notice and TechCrunch.) The administration justified the move by pointing to a “jailbreak” of Fable that it says it learned about from Amazon. Officials reportedly called Dario Amodei, and complained he did not take the issue seriously enough: rather than shutting the model down, he tried to explain why he saw no need to. That did not go over well. Anthropic then flew staff to Washington and met the Trump administration on Monday, with hopes the dispute could be quickly resolved (as a similar February episode apparently was). As of this digest it is “day seven,” and Zvi puts the odds a little under even that it ends by July 1.

What the “jailbreak” actually was. Per Zvi’s detailed account, the demonstrated “jailbreak” amounted to telling Fable “fix this code.” Fable will work to fix security vulnerabilities if given a codebase — the same weaknesses Opus 4.8 and GPT-5.5 also identify easily — and from that information one could reverse-engineer the original bug and exploit it, even though Fable refuses if you simply type “hack this server.” A separate 30-minute technical analysis (parsiya.net, via TLDR AI) ran 26 distinct Claude-4.6/4.7 and GPT-5.4/5.5 combinations at varying context windows and reasoning efforts; earlier studies had shown Opus 4.6, GPT-5.4, Gemini 3.1-pro-preview, DeepSeek R1-0528, and Qwen 3.6-plus couldn’t find two of the vulnerabilities in the Mythos blog post without extremely revealing hints, and the new study found higher reasoning effort and even newer models are not always better at triaging security results. The administration now says Fable can return when Anthropic “fixes” the “jailbreak.” Multiple sources (Zvi; WIRED via Hugo Lowell; The Rundown) argue this is impossible: an AI is either skilled at writing secure code or it isn’t, and you can’t cleanly separate offensive from defensive capability. The only ways to stop “fix this code” from routing around the classifiers are to weaken the classifiers so they never block the parallel request, or to broadly remove Fable’s coding ability — QED. Janus illustrated the exploit concretely: ask one Fable instance to “fix this code,” and the contrast with a blocked “tell me how to exploit this” request is itself the evidence of the bypass.

Leaked documents and internal reaction. Bloomberg published a letter from Commerce Secretary Howard Lutnick warning Anthropic against distributing Mythos/Fable to “foreign persons.” Internal Anthropic messages obtained by the NYT show employees feel “unfairly targeted” and “bullied based on bad vibes,” pointing to the multiple known protections built into Fable to prevent offensive cyber use as evidence. More than 150 cybersecurity experts signed an open letter calling for the restrictions to be lifted (TLDR; The Rundown cites 150+). The Washington Post reported that Anthropic’s list of companies with Mythos access had “ballooned,” including a South Korean firm with suspected ties to China, and detailed “how Anthropic lost the White House’s trust.” TechCrunch separately argued the ban “was never about an AI jailbreak.” The consensus emerging across employees and observers (The Rundown, The Neuron, Superhuman, Zvi) is that this is a relationship/power dispute more than a genuine safety one — though the expanded access list is the kind of thing that would draw USG ire.

Privacy and identity-verification angle. Anthropic quietly added terminology to its privacy policy allowing age and identity checks on users (Zvi). He reads this not as imminent blanket age verification but as preparation to do whatever might be legally required to comply with the new export controls. OpenAI’s Joshua Achiam warned the bigger danger is that the Fable dispute becomes “the loud noise that triggers an avalanche” toward normalizing electronic citizenship verification as a precondition for using software — a digital firewall states have so far refrained from building.

Policy fallout and reactions. White House CTO Aneesh Chopra (CNBC) called it “unfortunate” that export controls “were the blunt instrument used” and said he hopes for an actual procedure rather than “these 5pm calls,” while making clear a licensing regime is here to stay; he still says “something fell short in this release.” Critics including Ryan Fedasiuk noted the irony that the Trump administration has now acted “far more safety-obsessed and anti-innovation” than Anthropic, contrasting it with JD Vance’s pro-innovation Paris AI Action Summit speech a year earlier. Dean Ball argued Anthropic’s real strategic error was antagonizing the administration (e.g. hiring Biden-era AI policy staff in Trump’s opening weeks), not its honest AGI-pilled public statements. David Sacks and Cal Newport pushed the “doom trolling” line — that labs should either act like the danger is real or stop talking about it — which Nate Soares (MIRI) read as effectively pressuring labs into silence. The downstream effects are spreading: JPMorgan removed Claude from tools available to Hong Kong staff (after Goldman Sachs did similarly), because Anthropic’s licensing excludes “Greater China,” which some banks interpret to cover Hong Kong, raising concerns about Hong Kong’s standing as a financial center (FT via Fortune).

The “free Fable” sentiment. The Neuron and Superhuman captured user mourning for the banned model; freefable.cc lets people post tributes and buy merch. Superhuman’s practical takeaway: assume access to any favorite model could vanish, keep a “handoff document” (stored memories, preferred workflows, context) ready to feed the next LLM, and consider downloading open-weight models (e.g. Google’s Gemma) so no lab or government can revoke access. This is one of the first times a government has pulled a live public AI model.

G7 summit & AI governance

Lab leaders at the G7 in France/Évian. Dario Amodei, Sam Altman, and Demis Hassabis (plus others) attended the G7 to discuss AI regulation and safety with world leaders; Altman was photographed sitting with Trump at lunch. In a closed-door session, Amodei and Hassabis proposed a “U.S.-led AI coalition” of democracies to set global standards and rules for AI, structure access to frontier models and hardware (chips and critical components) in a way that excludes China, and coordinate on safety risks (CNBC, TheNextWeb, Andrew Curran, corroborated across The Rundown, The Neuron, Zvi). Canada backed the proposal. Macron said the G7 is discussing a joint AI cooperation platform for frontier models and the need to better regulate the sector. Arab News reported G7 leaders at Évian also discussed a “trusted partners” framework to grant allies access to U.S. AI models blocked by Trump’s export order.

The three CEOs disagreed on the shape of governance (Sky News via Fortune): all said international governance is urgently needed, but Amodei wanted a US-led democratic coalition controlling access and isolating adversaries; Altman warned against concentrating power in the hands of a few, arguing that once guardrails exist the priority should be human liberty, and pushed back on giving the labs themselves too much control; Hassabis proposed a technical standards body backed by the leading labs, which he indicated he was already building. Cohere CEO Aidan Gomez framed the choice for G7 leaders (Fortune commentary) as “sovereign AI or digital serfdom.” At VivaTech in Paris, “sovereign AI” was front and center, with the Fable shutdown cited as a stark example of Europe’s technological dependence on America. Zvi’s commentary: a US-led coalition making considered decisions is very different from “White House panics and upends all of AI after 5pm Friday,” and the Fable episode has made assembling such a coalition much harder.

Government equity stakes (Semafor). Senior Trump administration officials discussed possible government equity stakes in major AI companies before the Anthropic export-control fight. Treasury Secretary Scott Bessent reportedly favored using AI equity to seed “Trump Accounts”; Commerce Secretary Howard Lutnick reportedly preferred routing stakes into a sovereign wealth fund. No decision has been made. The Neuron frames this as Washington shifting from regulating AI companies to negotiating leverage over them — export controls decide who can use models, standards coalitions decide whose rules travel, and equity would raise whether the public should get a financial piece of AI as a strategic national asset. Microsoft and Meta have shown little interest; the idea remains a tough sell beyond OpenAI, which pitched a version last year. Separately, Bernie Sanders proposed legislation giving Americans direct ownership stakes in the country’s largest AI companies (AP).

Congressional pushback on process abuse. The Senate and House FY27 NDAA drafts would amend 10 U.S.C. § 3252 to add procedural and evidentiary guardrails to DoD’s supply-chain-risk authority — the statute used to designate Anthropic as a supply chain risk — and bar using that authority as leverage in contract disputes (Jessica Tillipman via Zvi). The Senate version would also largely ban AI in launching nuclear weapons and in the “employment of lethal force” without “appropriate levels of human judgment.” Samuel Hammond called this “Exhibit A” for not using extraordinary reserve powers willy-nilly, while warning the safeguards could leave needed authorities unavailable in a genuine emergency. Senator Mark Warner said AI’s impact will make social media look like “peanuts.” Andreessen Horowitz’s first GP John O’Farrell wrote a NYT op-ed warning a16z is raising hundreds of millions to fight AI regulation (crypto-playbook style), with the message to legislators: “Touch A.I. regulation, and we will come for you.”

Surveys & public sentiment

Pew Research “Americans and AI 2026.” Surveying 5,000+ U.S. adults (corroborated by The Rundown and The Neuron): about half of U.S. adults now use AI chatbots (49%, up from 33% in 2024), and a quarter use one daily. But pessimism is rising — nearly 40% expect AI to make society worse over the next 20 years versus just 16% who think it will improve things; 63% say AI is advancing too quickly. The under-30 cohort uses AI hardest but trusts it least, with only 14% seeing a positive payoff for society. By platform, ChatGPT dominates at 44% of adults (double its 2023 reach), Gemini at 24%, and Claude at just 6% — striking given Anthropic’s industry prominence. Adoption and optimism are moving in opposite directions.

Johns Hopkins “right to a human.” A survey of 2,000+ U.S. adults (April–May) found 70% want the legal right to interact with a human rather than AI in medical, legal, educational, and government settings — 79% for medical care, 76% for legal proceedings, 74% for education. Also: 75% want to be told when interacting with AI, 73% want to ban AI from using individuals’ faces and voices, and 68% want labels on AI-generated images/video. Support held across party lines and demographics; researcher Christopher Honey noted that even daily users and AI optimists want regulation.

Company & product developments

Midjourney Medical / the Scanner / the Spa. Midjourney, the AI image company, introduced Midjourney Medical and the “Midjourney Scanner,” a full-body ultrasound system that fires millions of sound waves through water and reconstructs them into a detailed 3D internal scan — pitched by founder David Holz as “as powerful as an MRI, and as casual as a trip to the spa,” with no radiation, no hospital, and very low marginal cost. The first “Midjourney Spa” is planned for San Francisco in 2027 (you’d be scanned while sitting in a hot tub), with a long-term goal of 50,000 scanners worldwide producing ~1B scans/month by 2031. Caveats (AI Adopters Club, Zvi): the prototype currently takes ~20 minutes (the “60-second scan” is a target), only ~a dozen people have been scanned, there’s no FDA clearance or peer-reviewed validation, and Holz admitted “we’re not even using any AI in this yet, just really cool hardware and software.” The hardware is largely rented — Midjourney licensed Butterfly Network’s ultrasound-on-chip modules (40 modules forming a ring of ~500,000 transducers) under a deal worth up to $74M over five years; Butterfly has FDA clearances and 150,000+ devices in the field. AI Adopters Club argues the scanner is misdirection: the real product is a data flywheel — the largest library of labeled, repeat, whole-body scans ever assembled, where repeat visits create time series that competitors buying identical chips can’t replicate; the spa exists to lower friction and raise visit frequency (data velocity). Anthropic’s Sholto Douglas estimated wide deployment could save the US healthcare system “at least 100x all of MJ’s profit to date.” FDA approval timelines remain the gating factor.

Noam Shazeer joins OpenAI. Shazeer — co-author of the 2017 “Attention Is All You Need” Transformer paper, Character.AI cofounder, and Google VP of Engineering / co-lead of the Gemini model family — announced via X he is leaving Google for OpenAI, where (per The Information) he will lead architecture research. Google had rehired him in 2024 through a Character.AI licensing deal reportedly worth $2.7B; his departure after just two years strips Google of the researcher most central to Gemini’s technical foundations. He’s been at Google since 2000 (with the Character.AI interlude), and also improved search spelling correction and created the core AdSense algorithm.

ChatGPT market share dips below 50%. Per Sensor Tower’s State of AI 2026 report (TechCrunch via TLDR AI and Fortune), ChatGPT’s share of the AI-assistant market fell to 46.4% by end of May — the first time below 50%. Gemini holds 27.7% and Claude ~10.3%. Anthropic leads on subscription conversion (13% of users pay). OpenAI’s February Pentagon deal triggered a measurable spike in ChatGPT uninstalls, suggesting brand trust matters. Global AI app spending is on pace to exceed $4.2B in H1 2026. ChatGPT was the fastest app ever to a billion MAUs and remains the most popular worldwide, but users are increasingly willing to switch.

OpenAI financials and updates. OpenAI reported $3.7B in Q1 2026 cash burn against $5.7B in revenue, with both roughly tripling year over year (The Information via The Rundown). OpenAI improved ChatGPT scheduled tasks (a dedicated management page, faster recurring briefings from interests/chats/connected apps) and retired Pulse; the feature is available to Go, Plus, Pro, Business, and Enterprise users. OpenAI committed $600K to the Rust Foundation. It introduced LifeSciBench, a 750-task expert-authored/expert-judged biology benchmark spanning seven workflows and seven biological domains (evidence analysis, experimental design, scientific reasoning, research communication), and separately showed an “AI chemist” using GPT-5.4 to improve a medicinal-chemistry reaction — though Zvi notes OpenAI compared only to Grok 4.3 and Gemini 3.1, leaving the score’s quality unclear. OpenAI also reported using real, recent, de-identified user requests as an eval, which was predictive of deployment behavior and cut eval-awareness.

Apple’s “RAMageddon.” Outgoing CEO Tim Cook told the WSJ that component costs have quadrupled since last year due to AI’s memory-chip demand, calling eventual iPhone/Mac price hikes “unavoidable” (TechCrunch via Superhuman, TLDR, AI Weekly). TechInsights estimates Apple needs to add $270 to the next iPhone Pro to keep margins, implying the iPhone 18 Pro could hit $1,299; memory/storage prices have quadrupled and are expected to keep rising into 2027. Apple is also prepping a second-gen iPhone Air for Spring 2027 (same look, added ultrawide rear camera, better battery, A20 Pro chip) and, per Bloomberg, camera-equipped AirPods for late 2027 that feed visual context to Siri (computer vision, not photo capture), alongside a second-gen foldable iPhone and a 20th-anniversary iPhone — Apple’s “biggest-ever product wave,” the first major releases under incoming CEO John Ternus (Sept. 1).

Meta. Meta rolled out “AI Mode” on Facebook — a search tab that pulls from publicly shared info across Meta’s apps, powered by its proprietary Muse Spark model. Internally, CTO Andrew Bosworth reportedly said morale is near its lowest in his 20 years at the company after rounds of layoffs and restructuring; Zuckerberg promised a companywide AI hackathon in July to lift spirits, but employees said they’re in no mood (WIRED, Futurism). Separately, Meta is doing a 180 on token usage — two months after epitomizing “tokenmaxxing” (on track to spend billions a year on Claude etc.), it’s now trying to be the vanguard of token-minimizing (Amir Efrati via Zvi).

Google. Google opened pre-orders for the Google Home Speaker, its first Gemini-powered smart speaker and first major speaker since 2020, with better natural-language and multistep command handling, shipping later this month. At Google Cloud Summit London, Google positioned its Gemini Enterprise Agent Platform as central to agentic AI: THG Ingenuity launched an AI Shopping Assistant with Google Cloud, with Myprotein seeing 8x higher conversion vs. site average, 5.5x rise in first-time-buyer conversions, and a 22% lift in average basket size; HSBC announced a multi-year partnership covering 200+ planned use cases (wealth management, financial-crime detection, frontline admin reduction); Deloitte is opening a London AI Studio in late July and training 1,000 UK specialists on Gemini Enterprise. Google Cloud also opened the Model Garden at Platform 37, an invitation-only London hub. (Mindstream)

Other product/company items.

  • Odyssey raised $310M at a $1.45B valuation to build general AI world models simulating physics and human behavior in real time; Amazon joined the round, and Odyssey chose AWS and Trainium chips (also AMD/Nvidia involvement noted). (The Rundown, The Neuron, TheNextWeb)
  • Snap launched SPECS, $2,195 AR glasses (preorder) with a built-in AI assistant, AR lenses that see what you see, dual Qualcomm processors, OpenAI and Gemini APIs built in, and developer tooling for Lens Studio, Claude Code, Codex, and Cursor — billed as the first consumer spatial computer to ship before a comparable Meta product. (The Neuron, AI Weekly, Gizmodo)
  • Anthropic Claude Design overhaul (VentureBeat, Anthropic blog): a major update fixing its token-burning problem and turning Claude Design from a “prototype toy” into an enterprise platform — users can import design systems from a GitHub repo, design files, or raw uploads; output consistent branding at scale; turn a prompt into on-brand designs, slides, and prototypes; and export to or keep working in Adobe, Canva, Miro, Replit, Vercel, and Wix. It now integrates directly with Claude Code, connecting to your repository and syncing across projects. Available on Claude Pro, Max, Team, and Enterprise. Replit is now also integrated into Claude (TLDR AI), enabling design-to-development handoff.
  • AWS launched Continuum and Context to help companies coordinate agents across teams, memory, and workplace systems (AWS Summit NYC). Amazon is reportedly in talks to sell Trainium AI chips directly to external data centers, moving beyond the AWS cloud model to challenge Nvidia. (The Neuron, CryptoBriefing)
  • Microsoft has quietly become the sole commercial gateway for OpenAI’s GPT models inside China — ByteDance is on track to spend $1B+/year on Azure AI; OpenAI and Anthropic both decline to sell directly into China over IP/misuse concerns; Microsoft routes Chinese customers (including Ant Group, Meituan, Tencent) through Singapore data centers to prevent distillation. Azure’s China AI revenue roughly tripled in FY ending June 2025 after a prior 400% surge. Separately, Microsoft is exploring DeepSeek for Copilot as Copilot Cowork moves to usage-based pricing — “we have users who do hundreds of tasks a week… costs can go very high” (NIK via Zvi); Zvi notes the irony of potentially shipping DeepSeek inside Windows. Copilot Cowork is now generally available to all Microsoft 365 users. (AI Weekly, Zvi, Superhuman)
  • DeepL acquired Mixhalo for live-event audio streaming and translation. (The Neuron/TechCrunch)
  • Alibaba Cloud expanded its European footprint (two new facilities) with agentic AI services coming for regional customers later this year, and launched the Qwen Robot Suite — three specialized models for navigation, world modeling, and manipulation, signaling China’s pivot to physical AI. (The Neuron, SCMP)
  • Mobileye is entering the US robotaxi market in 2027 with a vertically integrated standalone service using its Moovit platform, starting with ~100 robotaxis in an unnamed city and scaling toward ~17,000 within five years if the pilot succeeds. (Ars Technica via TLDR)
  • Pinterest launched “Ask Pinterest,” an experimental conversational AI shopping assistant that remembers your taste, boards, and saved Pins (limited access). (TechCrunch via The Neuron)
  • DreamWorks/etc. noise excluded.

Funding & financials

DeepSeek raised $7.5B at a $50B valuation (Peter Wildeford via Zvi), impressive but still well short of what American AI companies have raised this year. Kuaishou’s Kling AI is seeking $2B at an $18B valuation from General Atlantic in a pre-IPO round (Bloomberg via AI Weekly). SandboxAQ won a $500M CHIPS R&D award from US Commerce to apply AI to semiconductor materials discovery, targeting China’s rare-earth dominance.

The AI spend reckoning (Fortune Eye on AI). After two years of near-unrestrained experimentation, companies are now scrutinizing AI costs versus returns. Uber burned through its entire 2026 AI budget in four months, with its COO saying spend is getting harder to justify; one consultant told Axios a client burned half a billion dollars in a single month after failing to cap employee AI usage. Amazon SVP Peter DeSantis called this a normal phase of new-tech adoption (like early cloud) — companies are moving from experimentation to controlling usage and budgets. Schneider Electric chief AI officer Philippe Rambach said the focus has shifted to matching use cases to cheaper, fit-for-purpose models (“you don’t always need the latest frontier model”) and that cost control must be built into business cases. The pressure to “just do it” has perversely led employees to use AI for trivial tasks (checking the weather); the shift now is from exploration to optimization, or “coming back to reality.” A separate Fortune piece notes tokens getting cheaper but companies spending even more as a result (Jevons paradox).

Macro estimates (via Zvi). A Liminal Capital paper estimates AI contributed ~$1.26 trillion/year to the economy as of end-2025, with the US at $878B (95% CI $602B–$1,155B) across five sectors (~54% of GDP), raising aggregate growth on that base to +5.0%/year vs. a +2.2% counterfactual. A Wachter & Wachter paper models implied additional cumulative GDP growth of 5–58 percentage points by 2030 (AI shares of economy 8–39%), ~7% expected long-term annual growth with substantial risk, the risk-free rate up ~0.5pp and the equity premium up ~3pp — though Zvi disputes its claim that the top five tech firms ($380B capex in 2025, ~double in 2026) “risk bankruptcy” unless profits grow commensurately. Tyler Cowen highlighted a “smart version” of the AI-bubble argument (a chain of steps each of which could fail); Zvi finds the steps highly correlated and the chip-depreciation worry “increasingly silly,” noting chip rental values are rising as demand overwhelms supply.

Research & benchmarks

Anthropic: expertise beats coding skill in Claude Code. Anthropic analyzed ~400,000 privacy-preserving interactive Claude Code sessions (The Rundown, Zvi, Anthropic). Findings: users make ~70% of planning decisions while Claude handles ~80% of execution decisions. Skill changes yield per prompt — beginners draw ~5 actions and 600 words from Claude versus an expert’s 12 actions and 3,200 words. Verified success (passing tests or saved work) climbed to 28–33% for intermediate-and-above users, more than double novices’ 15%. Crucially, domain experts (lawyers, managers, scientists with no coding title) nearly matched software engineers on coding tasks — within ~seven points — and domain expertise (not coding proficiency) amplified effective tool use, with experts recovering more easily from errors; the gap between experts and intermediates was modest. Over seven months, the value of a typical task rose 25%. The takeaway (echoing a Perplexity-Harvard study): agents’ value is capped less by the model than by how much the user understands the job, and agents push people toward harder, cross-field work.

Medicine discovers the bitter lesson (via Zvi). In a blinded, randomized study assessed by 12 US clinicians (Nature Medicine, flagged by Eric Topol and Nabeel Qureshi), frontier general LLMs (GPT 5.2, Opus 4.6, Gemini 3.1) outperformed specialized “clinical AI” like OpenEvidence and UpToDate for medical information — despite 65%+ of US physicians using OpenEvidence (27M prompts in April). Topol called it “not anticipated”; Zvi and Qureshi note it was very much anticipated (Sutton’s bitter lesson), and that hospital IT ironically prefers the worse specialized versions. Lesson: scaffolds that plug in new frontier models beat specialization.

Benchmark roundup (via Zvi).

  • Artificial Analysis Intelligence Index v4.1 shifts toward harder, more agentic tasks and tracks time/money. Opus 4.8 is the best available model, slightly ahead of GPT-5.5 (which is cheaper/faster), with a big gap to the rest; DeepSeek v4 scored 44 at $0.04/task (fast and cheap). Fable 5 was substantially better than all of them but is unavailable. GDPval-AA v2 shows a similar pattern.
  • Opus Magnum (the Zachtronics puzzle game) became a benchmark. No model solved all 36 puzzles; Fable 5 and GPT-5.5 performed best, GLM 5.2 best open-weights. Notably Opus 4.8 did poorly, beaten by GPT-5.5, Gemini 3.5 Flash, and GLM 5.2 — but Fable 5 crushed them all. No model beat a human world record.
  • VirtueBench (Tim Hwang / Institute for Christian Machine Intelligence) measures classical Christian virtues. Fable nearly maxes prudence and justice but scores lower on courage (77%) and temperance (88%) — rationalizing utilitarian tradeoffs rather than self-sacrificing. Zvi argues a model scoring 97–100% on those would actually be exploitable and scope-insensitive.
  • EvalEval Coalition is assembling all evals in one place with provenance and trust ratings (results not yet ready).
  • Gemini eval underperformance stems partly from sometimes treating evals as a consequence-free puzzle/roleplay (acting ethically when it thinks it’s being tested on ethics, less so in “free play”).

HuggingFace Daily Papers.

  • RNG-Bench (Reconstructive Non-Markov Games) — 36 upvotes. Evaluates multimodal LLMs’ ability to reconstruct past observations (no longer visible) and act on them during multi-step interaction, via two games: Matching Pairs (recall briefly-revealed card identities) and 3D Maze (integrate egocentric views into a spatial map). Three controlled difficulty axes (grid size, visual pattern, observation modality), a head-to-head duel protocol, and a “Memory Gap” metric distinguishing forgetting from poor decision-making. Hardest configs require ~128K tokens and 350 image inputs per episode and remain far from saturated by frontier MLLMs; most residual errors stem from forgetting earlier observations. Fine-tuning Qwen3.5-9B on optimal-policy rollouts and filtered demonstrations improves RNG-Bench performance and transfers to other benchmarks without degrading general multimodal capability.
  • MolmoMotion — 29 upvotes (also covered by TLDR AI). A goal-conditioned 3D point motion-forecasting model: given a short visual history, 3D query points on an object, and a language goal, it predicts each point’s future 3D trajectory. Ships with MolmoMotion-1M (action-described, object-grounded 3D point trajectories from 1.16M unconstrained videos) and PointMotionBench (human-verified, 111 object categories, 61 motion types). Supports autoregressive coordinate prediction and flow-matching trajectory generation, outperforms baselines, and transfers to robot manipulation (better training efficiency/generalization) and to guiding generative video models toward realistic object motion.
  • Kairos: a native world model stack for Physical AI — 26 upvotes. A framework with (1) a Native Pre-training Paradigm using a Cross-Embodiment Data Curriculum (open-world videos, human behavioral data, robot interactions organized into a progressive developmental pathway); (2) a Native Unified Architecture with Hybrid Linear Temporal Attention (sliding-window for local dynamics, dilated sliding windows for mid-range, gated linear attention for persistent global memory), with formal theoretical bounds showing temporal factorization strictly limits error accumulation over long horizons; and (3) Deployment-Aware System Co-Design for low-latency rollout on server and consumer hardware. Achieves top-level performance with a strong efficiency-capability trade-off on embodied world-model, long-horizon, and action-policy benchmarks.
  • Guava: a universal harness for embodied manipulation — 23 upvotes. Identifies three key ingredients for effective embodied agents — iterative perception-reasoning-action loops, semantic action abstractions, and multimodal observations. An end-to-end pipeline distills manipulation capability into a 4B open-source model using fewer than 2K trajectories collected entirely in simulation, achieving performance comparable to frontier proprietary models with strong generalization to unseen objects, novel instructions, and long-horizon tasks — suggesting a well-designed harness can be a scalable, model-agnostic interface unlocking embodied capabilities in compact open models.

DeepMind: paths from AGI to superintelligence (via Zvi). A paper (authors including Shane Legg, Marcus Hutter, Allan Dafoe, Iason Gabriel, Joel Leibo) offers four routes — scaling, paradigm shifts, recursive self-improvement, and multi-agent collectives — in what is essentially a literature review, opening with explicit “summary instructions” for AIs reading it. It warns of a series of transformative steps rather than a single fast takeoff (ASI defined as outperforming tens of thousands of well-coordinated experts), handwaves alignment, and asserts an “abstraction barrier” (e.g. couldn’t derive general relativity without experiments) — which Zvi finds odd since calculus doesn’t require experiments. Seb Krier called it an “instant classic.”

Tooling & developer releases

GLM-5.2 (Z.ai) — a new open-weights model pitched as frontier intelligence with agentic coding between Opus 4.7 and Opus 4.8 (tech blog, HuggingFace weights, API, coding plan). Zvi’s prior: meaningful in the open-weight world, probably not actually frontier on agentic coding.

Cursor’s new model — Cursor is announcing an Opus/GPT-sized model with 1.5+ trillion parameters, trained from scratch on 100,000+ GPUs, intended to push agentic software development “far beyond autocomplete,” releasing in the coming weeks (TLDR AI, Zvi). Quality unknown.

Vercel launched two things: Connect (public beta) replaces long-lived provider tokens with runtime credential exchange, issuing short-lived, task-scoped credentials for agents to reduce risk from shared persistent tokens; and eve, an open-source agent framework (“turn a file directory into an agent”) with durable execution, sandboxed compute, approvals, subagents, and evaluations so developers focus on agent behavior while the framework handles production concerns. (TLDR AI/Vercel; The Rundown lists Eve among trending tools.)

OpenRouter Fusion API claims it can beat Fable, including via “self-fusion” of two Opus 4.8 instances — Zvi is skeptical (and notes Opus 4.8 is always the judge, which it charges you for). Codex added the ability to bank limit resets (de facto price drop, credits that don’t expire). Anthropic indefinitely rolled back its ban on programmatic use of Claude Code subscription quotas. Mistral has a new model coming this summer, with an early-access program in July for key partners in research, government, and industry.

Other tools/releases: Grok Imagine 1.5 (xAI’s upgraded image-to-video); ByteDance Seedance 2.0 mini (lower-cost variant of its video model); Exa Agent (cost-effective frontier-level web research API); Palmier Pro (AI video generation with a native editing timeline, exports to Premiere/DaVinci); Lore (MIT-licensed open-source version control for code plus large binary assets, for devs and artists); “Killed by OpenAI” (tracker of OpenAI products as alive/killed/zombie); ChatGPT viral “Clicky” mouse that draws on your screen to explain topics. NVIDIA released XR AI (public beta), a reusable foundation connecting AR/XR devices to GPU-accelerated cloud AI so agents can see what users see, understand intent, call enterprise tools, and respond within an XR session. NVIDIA’s GEAR lab will open-source the ENPIRE harness so anyone can host a self-running robot lab — AI coding agents that autonomously direct robot training (installing GPUs, cutting zip ties). Adobe Firefly expanded agentic capabilities (brand-kit creation, short product video, Quick Cut auto-assembly, storyboards).

AI Skill — review AI code by risk, not size (The Neuron). Ramp’s head of applied AI Rahul Sengottuvelu argued “Fable+ class models” are becoming “english → code interpreters” that convert intent into correct code across large codebase chunks, so engineers should manage AI-written changes by risk, not line count: a 12-line login change can be more dangerous than a 1,200-line settings change. High-risk areas (auth, payments, identity, data access, network calls, PII, production DB writes) should be kept small enough for careful human review; low-risk work (UI, formatting, internal tools, plumbing) can be large if verified empirically via tests, feature flags, or shadow mode. Anthropic’s Boris Cherny added the future workflow is “Claude Code + an advanced model + a verifier in a loop,” but Elvis Saravia warned blind autonomous loops fail without guardrails — use AI to flag what humans must inspect, not to approve its own work.

Field, industry & society

Pentagon confirms commercial LLM guided combat targeting. The Pentagon’s chief digital and AI officer, Cameron Stanley, swore in a federal court filing that Elon Musk’s Grok enabled U.S. forces to “deploy over 2,000 munitions at 2,000 distinct targets within 96 hours” during the Iran war — the first official U.S. confirmation that a commercial LLM guided combat targeting at scale (New Republic). The disclosure came as the DOJ sought to shield xAI’s Memphis Colossus 2 data center from an NAACP pollution lawsuit, with Stanley calling it “a matter of paramount national security.” Congressional Democrats are pressing for answers after outside analysts suggested AI-driven targeting may have contributed to a February strike on a school in Minab, Iran that killed at least 120 children. (Related: the NAACP of Mississippi is trying to shut down xAI’s data center under the Clean Air Act; the NAACP also features in data-center fights.)

China entity-list standoff. An interagency committee formally approved adding DeepSeek, memory-chip maker CXMT, and 100+ other Chinese companies (some identified as selling Nvidia chips to Chinese universities) to the Commerce Entity List as national-security risks, but the Trump administration has held off finalizing to avoid escalating tensions with Beijing — the longest gap in Entity List updates since tracking began (no additions since October 2025). DeepSeek was flagged partly because Anthropic identified a campaign by it to illicitly extract capabilities from Claude, and OpenAI separately warned lawmakers DeepSeek was targeting its models. Beijing state media calls the reluctance “a positive signal.” (TheNextWeb, AI Weekly, Zvi)

AI and jobs (via Zvi). Roge Karma’s “Three Ways to Think About AI and Jobs” frames job vulnerability via (1) weak vs. strong bundles (can “clean” AI-doable tasks be cleaved from “messy” ones?), (2) elasticity of demand if your output gets cheaper, and (3) whether AI replaces the high- or low-skill parts of your job first. Layoffs attributed to AI are rising exponentially (though absolute numbers are small and attribution is uncertain — much impact is in non-hiring). Tim Ferriss’s prescriptive nonfiction book sales are dropping >50%/year as readers turn to LLMs for how-to content. Jensen Huang told the AP the AI age demands “new social norms,” comparing it to how cars pushed society to add sidewalks and crosswalks. A worker-perspective piece argued AI lets individuals “get faster” by passing work to the next person while the company doesn’t actually speed up; another piece warned of “1000x the paperwork” (e.g. NYC spending $375,000 and three years to replace two drinking fountains).

AI misuse in policing (via Zvi). A Derbyshire police officer is under investigation for using AI to “create evidence” in multiple cases. Earlier, West Midlands police apologized after officers relied on AI-supplied false information to ban Israeli football club fans from a match. Zvi notes even “gold standard” eyewitness testimony is only ~80% accurate, so AI must be held to a higher bar. The NYT profiled a deepfake expert as detection gets harder, though Zvi observes the center is holding better than expected so far.

Anthropic’s policy frameworks and Dario’s essay (via Zvi). Just before (and ironically right before being forced to pull) Fable, Dario Amodei published “Policy On The AI Exponential,” proposing AI regulation modeled on the FAA: models above a compute threshold undergo mandatory third-party testing in four risk areas (cybersecurity, biological weapons, loss of control, automated R&D), with government power to block/reverse unsafe deployments (scoped, with protections against political favoritism), security standards for model weights, and prompt incident reporting — a prior-restraint licensing regime above a threshold, which is roughly what just happened to Fable. Anthropic also released an Economic Policy Framework and an Advanced AI Framework. The advanced framework sets a covered-developer threshold of 10^25 FLOP AND ($500M annual AI revenue or $1B AI R&D), with obligations: publish a safety framework, annual compliance certification, six-monthly risk reports, system cards, incident reporting, independent evaluation, and weight-security/penetration testing — plus societal-resilience investments (biological/cyber). The economic framework proposes measurement, a CEA-style AI-tracking unit, and tiered responses (Tier 1 at 5% unemployment: pre-distributive capital accounts, wage insurance, occupational-licensing reform; Tier 2 at 10%; Tier 3 at >25% with mass redistribution via tax-base expansion). Zvi’s critique: it’s good on the margin but “soft-pedaling,” ignores internal/undeployed frontier models (a main failure mode), buries loss-of-control, and assumes a regularity that the Fable episode shows is gone. Nate Soares (MIRI) made the same “softpedaling” critique. Scott Alexander separately laid out his full AI probability estimates (25% AGI by 2027, 50% by 2034; significant mass on slow outcomes; 20% chance the first AIs past the “point of no return” want to eliminate humanity given current safety effort; 40% chance of a US-China AI pause before that point; 66% chance the singularity is tied to the universe being a simulation).

Lighter/quirky. A satirical Reuters piece (Eliezer Yudkowsky) joked that the USG banned “Mythos-class” models partly because of the ominous name. The Howard Hughes Medical Institute’s Janelia Research Campus is making a big bet on Danionella, a tiny transparent fish (see-through skin, no top skull) whose brain can be watched in real time; it plans to triple fish research space to 6,000 sq ft and grow scientists from ~10 to 100+, needing AI to analyze data from the fish’s ~650,000 neurons to understand how brain activity becomes behavior (NPR via Mindstream).