AI Daily Digest

Sunday, July 12, 2026

2,719 words · All issues

Top items

  • The AI 2027 team publishes “AI 2040: Plan A,” a detailed positive-vision scenario proposing a US–China coordinated slowdown of superintelligence via compute monitoring and mutually-assured-compute-destruction — drawing wide endorsements and pointed authoritarianism critiques.
  • 1X unveils remarkably human-like tendon-driven hands for its NEO humanoid, while Hyundai’s Atlas performs live at the 2026 FIFA World Cup and Mistral launches its first robotics model, Robostral Navigate.
  • A compromised jscrambler npm package drops a Rust infostealer specifically targeting config files of AI coding tools (Claude Desktop, Cursor, Windsurf, VS Code, Zed).
  • OpenAI merges its research and safety teams as safety head Johannes Heidecke departs; separately claims GPT-5.6 “Sol Ultra” produced a machine-verified proof of the Cycle Double Cover Conjecture.
  • Massive AI-infrastructure financial moves: hyperscalers added ~$350B debt in five years; a $35B Anthropic/Broadcom TPU private-credit deal begins trading; SK Hynix warns of the “worst-ever” 2027 memory shortage.

Policy, safety & governance

AI 2040: “Plan A” — the AI 2027 team’s positive-vision scenario. The group behind AI 2027 (Daniel Kokotajlo, Eli Lifland, Romeo Dean, Thomas Larsen, with Scott Alexander and others contributing) has released Plan A, a follow-up that, unlike AI 2027’s two outcome scenarios, offers a concrete “positive vision” / optimistic forecast of how the transition to superintelligence could go well. Zvi Mowshowitz stresses he is not endorsing it as-is but thinks it deserves serious consideration and further work. There is also an unofficial visual-novel version. Track-record framing: Kokotajlo’s “What 2026 Looks Like” proved remarkably accurate, and Ryan Greenblatt (Redwood Research chief scientist) was the #2 most accurate AI forecaster of 2025 out of 413 entries.

  • The core proposal. Plan A updates from the present situation rather than being a pure prediction. Its implementation: America and China reach a mutually beneficial deal to slow AI development and share research information; joint control is established over existing and new chip supply, with common knowledge of chip concentrations and universal auditing of data centers, so both sides can verify compliance. A key pillar is total research transparency (Kokotajlo made a diagram illustrating why he weights this so heavily). The plan uses “mutually assured compute destruction” (MAIM-style) — which Vitalik Buterin notes effectively gives one of 2–5 actors the ability to trigger a global “compute winter,” an improvement over letting 1–5 actors selectively disenfranchise chosen “baddies.” During the transition America sees explosive economic growth, funding a rapidly rising “Citizen’s Dividend” (a generous UBI) plus a small fraction for “defensive acceleration.” In 2040 the “torch is passed to the AIs” and the singularity proceeds — the hope being alignment was gotten right on the first try. Notably, the plan’s baseline/default (without the herculean coordination) has this happening in 2030, so 2040 is already the slowed version.
  • The five-word compression. Zvi predicts most people will reduce it to: (1) slow down AI development; (2) make a deal with China; (3) monitor the world’s major compute; (4) use mutually assured compute destruction; (5) things still happen fast (ASI by 2040 with vast prior economic growth). Axios’s Ashley Gold compressed it as “slow everything down.”
  • Alternative plans described. Plan B: fight/sabotage China to buy a buffer. Plan C: “burn the lead” (roughly AI 2027, with more awareness). Plan D: race to ASI (the current path). Plan S: shut it all down — the authors are sympathetic but reject it as unstable and hard to get buy-in for. They forecast the 2028 US election being all about AI, with a plan (hopefully A) implemented in 2029.
  • Endorsements. Buck Shlegeris (“best summary of the spectrum of possible plans”), Greenblatt (“best plan anyone’s written up… robust to partial/lower-quality implementation; the core—verification and very strong transparency—suffices to go pretty far”), Eli Tyre (“single most thorough and thoughtful plan”), Jeffrey Ladish, Steven Adler, Marius Hobbhahn (mapped his “too many low-probability things” reaction to “it’s the SOTA plan for a hard problem”), and OpenAI’s Boaz Barak (loves the push for increased transparency and diffusion rather than lockdown). Nate Soares (MIRI): “the best concrete vision of a positive future I’ve ever seen spelled out, bar none,” but doubts it works as written — his hope is it produces evidence convincing leaders to switch to Plan S; his critique is that humans working with AIs on alignment likely converge on wrong answers, and the honest answer (“pushing to superintelligence with modern methods is fucked; back off”) is one humans are unlikely to heed.
  • Critiques Zvi engages. Selective optimism (Richard Ngo): as an optimistic forecast it’s hard to tell which parts are aspirational vs. included for realism; Ngo would prefer a slower handoff (so would Zvi and the authors), wanted dates removed, and challenged the “race with China” framing. Zvi defends keeping dates (a scenario without them draws the opposite criticism) and the race framing (that’s how DC thinks now). Economic optimism: Zvi concedes the diffusion/growth projections are optimistic but argues (with Shakeel Hashim) that slower diffusion doesn’t change the central recommendations. China won’t accept inferiority (Yishan, Prakash): China might accept AI equality but not enshrined American superiority, potentially making any Plan A DOA / no zone-of-possible-agreement; the plan does try to rebalance robot share etc. Politics naive (Kyle Corbitt): e.g. Canadian voters wouldn’t let China build thousands of datacenters in 2030 — Zvi counters that seemingly-impossible concessions get signed when impetus arrives, citing COVID’s sudden acceptance of lockdowns, cash handouts, and vaccine mandates.
  • The authoritarianism objection. Ramez Naam argued Plan A’s restrictions violate the spirit of the 1st and 4th Amendments and are “a recipe for authoritarianism,” saying he’d change his mind only given evidence of fast takeoff + loss-of-control + antisocial volition + a believable software-only-singularity model. Zvi responds that (a) equating compute controls with 20th-century fascism/communism is hyperbolic; (b) comparing Plan A to a hypothetical no-restrictions world where nothing goes wrong is not a fair baseline; (c) the plan is explicitly designed to minimize authoritarianism (Vitalik agrees it improves on power-concentration dimensions and even mandates open source); (d) the real crux is whether superintelligence is coming — Séb Krier (Google DeepMind AGI policy lead) and Vitalik both note the debate hinges almost entirely on future AI capability levels. Vitalik’s own position: he’s genuinely uncertain which world we live in, favors d/acc (defensive tech good in both worlds), but concedes there’s no “escape the tradeoff” button on pause/slowdown, and proposes pre-agreed “serious shit” triggers (super-pandemics, >25% unemployment, slaughterbots) that both camps could accept.
  • The framing Zvi emphasizes: either humans can collectively steer the future or they can’t; if AI reaches AI-2027/Plan-A capability levels and humans don’t steer, the equilibrium is that AIs end up in charge (without tricking anyone or firing a shot) because the alternative is being outcompeted — “and that’s the good version.” He also flags that Plan A treats alignment mostly as a control/game-theory problem (making deals with potentially-misaligned AIs) rather than genuine cooperative alignment — a point John Wittle strongly criticized (“you don’t get mutual cooperation by programming it into your adversaries by force; that’s defection”).

Compromised jscrambler npm package targets AI coding tools. The Hacker News reported (July 11) that jscrambler’s official npm package was compromised across five releases (8.14.0, 8.16.0, 8.17.0, 8.18.0, 8.20.0). A preinstall hook dropped a cross-platform Rust infostealer targeting cloud credentials, browser data, crypto wallets — and, notably, the config files of Claude Desktop, Cursor, Windsurf, VS Code, and Zed, reflecting how AI dev tools are now high-value credential targets. Later releases embedded the dropper in main code to bypass npm’s --ignore-scripts flag. Jscrambler confirmed the attacker used compromised publishing credentials; Socket flagged the release six minutes after publication, and 8.22.0 is the confirmed clean release.

OpenAI merges research and safety teams; safety head departs. Wired reports OpenAI Head of Safety Johannes Heidecke is leaving as the company merges its research and safety organizations, with Mia Glaese becoming VP of Research and Safety.

Generative AI in terrorism. The NYT documented how terrorist groups, including Boko Haram, are using generative AI to plan attacks, recruit, and produce propaganda — with Nigerian officials and US counterterrorism analysts documenting real-world battlefield impact.

China drops urban job-creation target. Bloomberg reports China’s 15th Five-Year Plan omits a numerical urban job-creation target for the first time since the 1990s, explicitly citing uncertainty over AI-driven job displacement.

Robotics & embodied AI

1X unveils human-like hands for the NEO humanoid. Norwegian robotics startup 1X went viral this week with a demo of new hands for its NEO consumer humanoid that may be the most capable and human-feeling robot hands shown to date. They use tendon-driven motion and fingertip tactile sensors that detect pressure in real time, enabling delicate tasks like picking grapes and installing light bulbs, and even communicating in sign language. Superhuman calls it “possibly the first time robot hands have felt genuinely human.” (Bernt Bornich and 1X posted demo clips.)

Hyundai’s Atlas makes its live World Cup debut. The Hyundai/Boston Dynamics Atlas humanoid performed at halftime of the Norway vs. Brazil match at the 2026 FIFA World Cup, walking onto the pitch before a packed stadium to reenact goal celebrations from Erling Haaland, Harry Kane, and Son Heung-min. It then collected the ceremonial match ball and delivered it to the referee — a demonstration that the robot can operate reliably in a live, high-pressure, global-audience environment (the first such live-match robotics integration).

Mistral launches its first robotics model. French AI startup Mistral released Robostral Navigate, its first robotics model, marking a push into factories, warehouses, and industrial automation. It’s built to work across hardware from multiple suppliers, signaling Mistral’s ambition to compete in physical AI alongside Nvidia and Google DeepMind. The launch follows Mistral’s acquisition of Austrian robotics-AI startup Emmi AI.

Other robotics developments.

  • Northstar: Ex-Tesla scientist Rémi Cadène unveiled Northstar, a faceless humanoid robot that learns tasks from demonstrations, targeting European factories.
  • Waymo is launching fully driverless rides in San Diego, Las Vegas, Tampa, and Denver, expanding its robotaxi network past a dozen cities.
  • Tesla urged New Jersey owners to lobby against a proposed bill it says would effectively block deployment of fully driverless vehicles in the state.
  • US regulators ordered AV companies to fix how vehicles respond to emergencies, warning robotaxis must stop interfering with first responders.
  • The Netherlands opened a humanoid robotics center to accelerate Europe’s robot industry and narrow the gap with China and the US.
  • South Korea is turning to unstaffed robot-run businesses (coffee shops, noodle bars, clothing boutiques now numbering in the thousands) to cope with a shrinking workforce and one of the world’s lowest birth rates; one café chain reports profit margins above 40%, more than triple staffed equivalents.
  • RoboCup 2026 drew participants from 45 countries; China’s Tsinghua University won the Humanoid League for the second straight year. The competition’s stated long-term goal is a robot team that can beat the FIFA World Cup champions by 2050.
  • UC San Diego researchers completed the first-ever surgeries using teleoperated humanoid robots in a preclinical trial, including a gallbladder removal by a surgeon–robot team; the robots used standard surgical instruments and needed no purpose-built OR, making them far more deployable than existing surgical-robot systems.
  • Consumer robot of note: Lissome’s R1 countertop dishwasher uses AI water sensing to adjust pressure, spray direction, and detergent based on how soiled dishes are, with a built-in tank and auto detergent dispenser — aimed at small kitchens, RVs, and studios.
  • A viral Indonesian clip of a humanoid appearing to “go rogue” and lunge at office workers turned out to be a fully pre-programmed, choreographed mobility demo.

Company & product developments

OpenAI GPT-5.6 rollout and math claim. After a security delay and government approval, OpenAI is launching GPT-5.6, with the Sol, Terra, and Luna variants rolling out following extra testing (Mindstream). Separately, OpenAI posted a PDF claiming GPT-5.6 “Sol Ultra” produced a machine-verified proof of the Cycle Double Cover Conjecture, a long-open graph-theory problem — the claim hit the Hacker News front page.

Anthropic product and research moves. Anthropic launched Claude Science, a research workbench for scientists, and quietly signaled it wants to pursue drug discovery for neglected diseases itself, not merely sell tools to researchers. It also expanded Claude Cowork to mobile and web, letting tasks keep running in the cloud even when a user’s laptop is closed. Separately, court documents revealed the exact emails between Dario Amodei and the Pentagon that ended Anthropic’s government contract, including Amodei’s line “I unfortunately don’t see a way forward.”

Google / Gemini. Google added AI video editing to Google Photos via “Gemini Omni,” letting users change backgrounds, add cinematic lighting, and apply styles (watercolour, oil painting) directly from their camera roll. Google also released a Declaration of Independence ad imagining the founding fathers using Google Docs, Google Meet (cameras off), and Gemini for meeting notes — broadly well-received except on Bluesky.

Meta Muse Image. Meta launched Muse Image, a free AI image generator that can pull photos from other people’s public Instagram profiles without notifying them — a capability the company reportedly anticipated would draw attention.

Musk / Grok / SpaceX. Elon Musk denied ordering Tesla and SpaceX to adopt Grok 4.5 exclusively, saying (after a July 10 memo leaked) he only asked them to “try it out.” Sam Altman publicly accused Musk of “selling public market investors on short-term space data centers” amid a $2T SpaceX valuation push.

Consumer / culture. A UK studio cast AI-generated actor Tilly Norwood in a film titled Misaligned about AI identity and memory; SAG-AFTRA does not consider her a real actor. A content creator, Ben Guez, used OpenClaw to auto-generate “heartbroken fan” videos targeting women from losing World Cup countries and pipeline their DMs into an AI app; others used OpenClaw to auto-send break-up texts.

Field & industry developments

AI infrastructure financing and debt. Bloomberg reports Alphabet, Amazon, Meta, Microsoft, and Oracle have collectively added roughly $350B of debt over five years to fund the US data-center buildout, with AI capex now outpacing operating cash flow at multiple hyperscalers. Separately, Apollo and Blackstone’s $35B Anthropic/Broadcom AI-chip private-credit deal is set to begin trading — a debt package financing TPU purchases via an SPV, with about $15B available to institutional buyers by early 2027.

Memory-chip supply crunch. SK Hynix’s CEO warned of the “worst-ever” memory shortage in 2027, saying AI demand will outstrip supply beyond 2030 — comments made on the day of the company’s Nasdaq debut. In the same vein, Samsung forecast a 19x jump in quarterly profits driven by AI memory chip demand, yet its shares fell 7% as investors deemed record-breaking results insufficient.

Trump administration intervention for Intel. The WSJ detailed the administration’s heavy-handed effort to aid Intel, including pushing Apple to use Intel’s fabs, with Commerce Secretary Lutnick repeatedly meeting Tim Cook, Elon Musk, and Jensen Huang.

AI leadership perspective (interview)

Sarika Garg (HubSpot Revenue Hub GM/VP, former Cacheflow CEO) on where AI value really lies. In a Mindstream interview, Garg argued front-office AI stalls until back-office data is fixed: “your AI is stalling because your data is a mess, and no amount of clever prompting fixes dirty infrastructure.” Key points: (1) the overlooked opportunity is untangling duplicates, dirty data, and disconnected/unmapped processes — AI excels at producing clean, connected, contextual data at scale, and front-office AI stalls until that’s done. (2) The roles changing most in revenue operations are the deal desk and reporting analyst, shifting from mechanically building quotes/reports/Excel forecasts to architecting the data model and revenue system and applying judgment. (3) The signal that a company builds real AI value vs. just “AI features”: whether AI is woven into the business process and can run jobs autonomously versus merely helping one person work faster. (4) Advice to founders: start with a genuinely painful problem and solve it 100x–1000x better (which requires AI doing a full job, not just a task). (5) The rising human skill is systems thinking — humans become “architects of the business, not task executers.” (6) She rejects mass joblessness fears — “every job will change… but that’s not the same as disappearing” — predicting a “golden era” once busywork is removed. She personally uses a series of agents that listen daily across community chatter, CSAT, support tickets, Gong prospect calls, and the CRM, summarizing signals she used to gather only quarterly.