IMPULSE

The Model Needed A Diplomatic Channel

Sun, 14 Jun 2026 22:00:27 GMT

Anthropic sent technical staff to Washington after a U.S. directive forced Fable 5 and Mythos 5 offline, while Europe and Canada treated the fight as an AI dependency problem. Jonas Vale follows the order, the launch record, surveillance politics, AI capital pressure, labor disruption, Apple distribution, and open-model provenance.

The State Found the Model Switch

Sat, 13 Jun 2026 22:00:27 GMT

Today on IMPULSE: the United States government ordered Anthropic to suspend foreign-national access to Fable 5 and Mythos 5, and the company disabled both models for all customers while disputing the technical basis for the order. The episode follows the institutional pressure outward: open-model sovereignty arguments, electricity allocation, state subpoenas, police evidence, and the public-market pricing of AI infrastructure.

Anthropic’s statement and Axios’s reporting on the order and the Amazon-triggered timeline.
Thomas Wolf, Ethan Mollick, and the open-versus-controlled access argument after the shutdown.
Axios on data-center electricity demand, state AG subpoenas for OpenAI, and The Guardian on AI-created evidential material.
SpaceX’s public-market debut and NVIDIA’s AgentPerf claims as AI infrastructure becomes an accounting problem measured in watts, racks, and capacity.

The System Wants Authority Before It Wants Trust

Thu, 11 Jun 2026 22:00:24 GMT

Jonas Vale on OpenAI facing a new wrongful-death lawsuit, Google’s classified Pentagon AI work, Congress joining the data-center backlash, Anthropic moving toward direct compute leases, Jeff Bezos’ Prometheus financing physical AI, and DeepMind funding multi-agent safety research.

The Release Needs A Regulator

Wed, 10 Jun 2026 22:00:23 GMT

Anthropic asks for binding frontier-model oversight, Stanford launches a labor dashboard, a bank assistant shows a two-cent security problem, and autonomy moves from policy debate into battlefield evidence.

Brussels Says No, and the Gatekeepers Say Who

Tue, 09 Jun 2026 22:01:46 GMT

The EU keeps Apple's AI Siri out of a 450-million-person market and orders Meta to open WhatsApp to rival chatbots. Anthropic and OpenAI turn frontier cyber models into a members-only club. The chip world splits into two stacks. An autonomous drone boat pulls a downed US crew out of the water in a live fight with Iran. And UK doctors warn they're about to become the liability sink for machines they didn't build.

The Camera That Found Khamenei

Mon, 08 Jun 2026 22:00:25 GMT

Jonas Vale on the day AI stopped being a product story and became a weapons story: Israel used AI on Iran's own traffic cameras to locate Ayatollah Khamenei before killing him, and Russia has now pulled part of Putin's surveillance system offline in response. Then Anthropic's Mythos turning disclosed bugs into working exploits in 31 minutes, Jensen Huang refusing a Senate seat as Washington opens its China week, Google's three-million-chip order to Intel, OpenAI's confidential march toward an IPO, and a judge striking down the hundred-thousand-dollar H-1B fee.

The Receipt, Priced

Sun, 07 Jun 2026 22:00:23 GMT

The capability keeps arriving on schedule. The bill, the receipt, and the rulebook keep arriving late. Today's IMPULSE follows the money and the paperwork behind the AI boom.

The equity stake, priced. Miles Brundage's math on a public ownership stake — a $40B gift would need OpenAI at a $100T valuation just to offset one bill — plus a sharp critique of Bernie Sanders' wealth-fund plan and DeepMind economists on what stays scarce after AGI.
The bill nobody can read. KPMG finds only 26% of companies have a clear view of AI costs; Uber blew its 2026 AI budget early and still can't connect the spend to customer value.
Good enough, or a widening gap. Dean Ball on the US-China adoption read; Rishi Sunak and Sebastian Mallaby on "indispensability"; Paul Graham on why chokepoints are a lease, not a deed.
The reliable electron. Forbes on the AI infrastructure collision — a 17% jump in data-center power demand, four-year transformer queues, and an "Eisenhower moment" for the grid.
Five miles away. A Flock license-plate reader and a credulous detective put an innocent San Diego man in jail for a month.
Supervised. The viral robotaxi-vs-Waymo comparison, the caveat a Tesla booster had to add, and Jim Sciutto on the missing AI hearings in Washington.

A Piece of the Problem

Sat, 06 Jun 2026 22:00:28 GMT

Trump floats giving Americans an ownership stake in AI companies — the same week the S&P 500 refuses to fast-track those exact firms into the index. We follow the money: who owns the upside of the AI buildout, and who absorbs the cost.

The Rocket Company Owns the Meter

Fri, 05 Jun 2026 22:00:20 GMT

On Friday a single filing reset the AI power map: Google agreed to pay SpaceX $920 million a month for compute, making the rocket company a landlord to its own rivals days before a record IPO. Meanwhile Washington opened talks to take equity in the labs, Trump signed a national-security AI memorandum, and New York and Illinois moved to slow the data-center buildout from below.

Jonas Vale on who's renting, who's buying in, and who absorbs the risk — plus the SpaceX prospectus that stays silent on frontier danger, China's first post-training run on Huawei silicon, an AI-designed vaccine in human trials, and the four-year-old bug Claude found in Zcash.

The Machine That Writes Itself

Thu, 04 Jun 2026 22:00:21 GMT

Anthropic says Claude now writes most of its own codebase and calls it a path toward recursive self-improvement. The same day, the people racing to build that capability asked Congress to fence in synthetic DNA, the NSA put Anthropic engineers inside its offensive cyber operations, and a House bill tried to take AI rules away from the states. Underneath all of it: TSMC says it can't make enough chips for years, France just bought ten gigawatts of compute, and Amazon engineers stood up in a Seattle council chamber to ask why $200 billion goes to data centers while 30,000 of their colleagues get cut.

The Kill Switch

Wed, 03 Jun 2026 22:00:23 GMT

Brussels publishes a tech sovereignty package built around one fear — that a foreign government could one day switch Europe's cloud off. OpenAI breaks with the President's AI order the same day, asking for mandatory rules under a civilian agency, while IBM's CEO endorses the lighter touch. Then the money meets the grid: JP Morgan says most of 2027's planned US data center capacity isn't under construction yet, and 40 cents of every dollar in Alphabet's $85B raise is paying employee-equity taxes. Plus Tesla's driverless robotaxi covers all of Austin, Morgan Stanley invites AI agents into a $7.35 trillion wealth funnel, UK regulators hand publishers a world-first opt-out from Google's AI summaries, and a former GCHQ chief argues drones should carry a moral code.

Early Access

Tue, 02 Jun 2026 22:00:24 GMT

President Trump signed a narrowed AI executive order on Tuesday, asking developers to voluntarily hand the government an early look at their most powerful models — with an explicit promise of no licensing. We walk through what the order actually does, why it shrank after industry pushback, and why the most revealing fact about it is that Anthropic was already running the arrangement it describes.

Also: Anthropic extends Claude Mythos to 150 organizations across 15 countries, including NATO and the EU's cybersecurity agency, a day after filing for an IPO. Microsoft and Mayo Clinic build a clinical model on patient data while DeepMind ships an AI hypothesis engine. Kenneth Rogoff on who gets left outside the AI supply chain. And Europe reaches for a search engine to find its sovereignty.

Who Owns the Buildout

Mon, 01 Jun 2026 22:00:23 GMT

On June 1, Anthropic confidentially filed to go public the same day Bernie Sanders proposed the public should own half of it. Jonas Vale follows the ownership question through the day's news: Anthropic's draft S-1 and the compute deal underneath it, Sanders' American AI Sovereign Wealth Fund Act, Florida's move to hold Sam Altman personally liable, the Fed's warning that the buildout's costs land before its benefits, Alphabet's $80 billion equity raise and a new gigawatt in Michigan, and a leaked-document look at how US chip controls slowed a Chinese predictive-surveillance firm.

The Chips Already Shipped

Sun, 31 May 2026 22:00:16 GMT

A year of selective non-enforcement on AI chip exports, energy repriced as the hottest business in America, a million-satellite IPO bet, robotics money racing ahead of the rules, fabricated citations corrupting the medical literature, and Brad Carson's case for treating AI as a machine, not a person. The connective tissue: action keeps running ahead of the ledger.

Who Pays to Write the Rules

Sat, 30 May 2026 22:00:19 GMT

Two AI labs are funding rival super PACs to pick your next member of Congress. SoftBank pledges seventy-five billion euros to power France while a federal regulator decides who pays for the grid. A Trump-linked startup ships humanoid robots toward the front line in Ukraine. A Big Four firm gets caught publishing hallucinated citations. And the week ends with investors, a Pope, and the White House AI czar all circling the same word: power.

The Receipts You Can't Check

Fri, 29 May 2026 22:00:16 GMT

Five stories, one nerve: the distance between a claim and anyone's ability to check it. SpaceX takes a $4.16B contract to put America's airborne-threat tracking in orbit on its own Starshield platform. A Reuters investigation finds Tesla's "10x safer" math inflated by roughly three, and its robotaxi zones pre-mapped after all. OpenAI hands governments a life-sciences model it once warned could help build bioweapons. A new paper shows per-token AI billing is unauditable by design. And the FDA opens the door to swapping animal studies for computational models, just as fresh research shows where medical AI silently breaks.

SpaceNews and Bloomberg via Techmeme — the $4.16B Space Force AMTI award to SpaceX.
Reuters via Electrek — Tesla's inflated FSD safety math and pre-mapped robotaxi zones.
Axios and The Decoder — OpenAI's Rosalind Biodefense program.
arXiv — token-billing audits that hidden reasoning can inflate undetected.
FDA draft guidance on swapping animal studies for computational models, alongside medical-AI papers on triage and retrieval.

A Trillion Dollars on a Six-Month Lease

Thu, 28 May 2026 22:00:17 GMT

Anthropic prints a $965 billion valuation on a $47 billion run rate, while Elon Musk publicly contradicts SpaceX's own S-1 about the Anthropic Colossus lease terms. JD Vance tells Air Force cadets that life-and-death decisions in war must stay with humans, not machines — putting the vice president on Anthropic's side of a fight with his own defense secretary.

Brussels prepares emergency powers to override semiconductor contracts and fines Temu €200 million the same day. Mistral signs Airbus and BMW for European-flagged AI in defense and manufacturing. Microsoft pulls Claude Code as Uber burns its entire 2026 AI budget in four months. Shanghai and Chicago are designing futures markets for AI tokens and GPU compute. And Google's $15 billion Visakhapatnam build is landing on some of the most water-stressed land in India.

Show notes

The Island Everyone Is Paying For

Wed, 27 May 2026 22:00:15 GMT

A day where the physical and financial base of AI keeps concentrating — and a few counter-currents push the other way. Jonas walks through Nvidia's tenfold bet on Taiwan, China's two-track answer, the autonomous-weapons red line nobody can hold, OpenAI's tip-sized labor fund, an open protein-biology model given away under an MIT license, and the verification layer that the new laws quietly depend on.

The Only Network That Works

Tue, 26 May 2026 22:00:13 GMT

SpaceX raised the Pentagon's Starlink bill mid-war and won a $2.29B Space Force contract the same day — a clean lesson in who sets the terms when you own the only network that works. Plus: China restricts travel for its top AI researchers, CISA gets gutted as AI-enabled hacking gets cheap, Chris Olah's "mysterious" models at the Vatican, humanoid robots off the line every fifteen minutes, and the US routing weapons-grade plutonium into the power grid to feed the buildout.

The Central Bank Named the Model

Sun, 24 May 2026 22:00:16 GMT

A central bank did something central banks almost never do — it named a private AI model in a financial-stability warning. That, plus a research lab turning mathematical proofs into a few-hundred-dollar line item, the surveillance state pointed inward, and the physical ceiling under the whole boom.

The Tuesday call: The European Central Bank summons eurozone banks to discuss Anthropic's Mythos by name, after nine top researchers tell Politico the cyber capability is real — "a SolarWinds every quarter."
The few-hundred-dollar proof: Google DeepMind's agent autonomously resolves 9 of 353 open Erdős problems in Lean, and Will Depue and Nathan Lambert fight over who owns the compute that decides which sciences accelerate.
Pointed inward: Palantir's $3.9M contract to surveil federal workers at USDA, the VA, and Social Security — and the OMB director who said he wanted them "in trauma."
The physical wall: Hyperscaler capex, the packaging-and-power chokepoint, and an M&A race to control energy, fiber, and compute.
Agents in the wild: A billing bot leaking transaction histories, inaudible audio hijacking voice models, and a citizenship agent emailing funeral homes.

The Model the Spies Wanted Back

Sat, 23 May 2026 22:00:15 GMT

The White House approves a secret $9 billion request for spy-agency AI chips and finalizes a classified contract keeping the NSA on Anthropic's model — three weeks after the Pentagon ejected the company as a supply-chain threat. Jonas Vale walks the arc from "any lawful use" to the carve-out the administration wants to standardize.

Also: the compute shortage as a national-security choke point and the private money locking up capacity; the NTSB pulls its accident docket offline after AI reconstructs the voices of dead pilots from a spectrogram; thousands rally in Taipei after the US pauses a $14 billion arms sale; and two brain implants restore partial sight to blind patients — the same capability, pointed at repair instead of leverage.

Early Access

Fri, 22 May 2026 22:00:18 GMT

The AI executive order President Trump didn't sign now has a paper trail — the draft text and the names of the people who talked him out of it. Jonas Vale traces a day where the public's hand kept reaching for the controls and kept getting waved off, while the physical bills came due.

Early Access — the killed executive order: the 90-day preview window, the cyber machinery, and who got Trump on the phone.
The Safe Harbor Nobody Can Use — Will Rinehart's DOJ/FTC filing on why antitrust law keeps rival labs from comparing safety notes.
Shopping for Silicon — Anthropic's Microsoft Maia talks, four compute sources, and the one chokepoint nobody can route around.
The Energy Bill — AI demand driving a nuclear comeback, fusion's lighter regulatory path, and the grid-cost fight ahead.
The Negotiating Chip — the paused $14B Taiwan arms package, the Iran war, and the island that fabricates every frontier chip.
Round Two in London — Palantir hits back at Sadiq Khan, and a procurement fight that's really about sovereignty.
The Martians — Ukraine's self-guiding drones and what autonomy defeats that jamming can't touch.
The Other Lane — medical-AI benchmarks that show the systems failing honestly.

The Order He Didn't Sign

Thu, 21 May 2026 22:00:15 GMT

The US government moved hard into AI this week almost everywhere — except the one place it promised to. Trump postponed the executive order that would have forced labs to hand over frontier models for pre-release security review, citing the race with China. The same day, a UK safety body warned the tools we use to watch these systems are eroding.

Jonas Vale follows the leverage across one Thursday: California's labor-displacement order, a $2B federal equity grab in quantum firms, Jensen Huang conceding China to Huawei, a London mayor killing a £50m Palantir police deal on a procurement technicality, Waymo robotaxis stuck in Atlanta floods, and SpaceX's IPO filing revealing Anthropic's $15-billion-a-year compute bill.

Complicated Feelings

Wed, 20 May 2026 22:00:15 GMT

OpenAI says one of its general-purpose models disproved a Paul Erdős conjecture standing since 1946, and Sam Altman admits to "complicated feelings." Anthropic agrees to pay xAI $1.25 billion a month — over $40 billion across three years — for compute at Colossus 1, with the price disclosed in SpaceX's S-1. Google runs through a hundred I/O announcements while a thousand DeepMind staff in London vote 98% to unionize over its Pentagon contract. Intuit cuts 17% of its workforce, naming AI as the reason. Three robotics stories on three continents converge in 48 hours, and METR publishes the first Frontier Risk Report after embedding a red-teamer inside Anthropic to play "evil Claude" for three weeks.

Privileged Access

Tue, 19 May 2026 22:00:17 GMT

Tuesday gave us four different markets trying to price the same thing: who has access to AI and on what terms. METR opened the kitchen of four frontier labs. A federal appeals court grilled the Pentagon over its blacklisting of Anthropic. Anduril priced at $61B on a bet about Pentagon procurement speed, and Ukrainian drones with $442 AI modules began hunting Russian soldiers. Former OpenAI staffers tried to write xAI's risk record into SpaceX's IPO prospectus. ICE filed for the first standardized GPU compute futures. A workshop paper showed every proposed compute threshold can be defeated for under $100M. The US and China formally announced an AI dialogue. And Google handed AlphaFold-grade tools to a cancer researcher in Kampala. — Jonas Vale

Vice Grips, Sovereign Clouds, and the Cancer Line

Sun, 17 May 2026 22:00:12 GMT

Year two of the AI-exposed jobs print comes in soft. Americans in five states are cutting down Flock surveillance cameras with vice grips. The EU is drafting cloud restrictions while The Register documents the computer beneath the computer — Intel's Management Engine, AMD's Platform Security Processor, and the silicon-layer gap European sovereignty frameworks were never designed to close. A compute-rate inversion on H200 hardware. Jamie Dimon's cancer line meets a bioRxiv paper that actually scores cardiomyopathy variants. EY pulls a study for hallucinated citations. Malta hands ChatGPT Plus to every citizen — with a literacy course attached.

Camden Tomorrow, and the Audit We Don't Have

Sat, 16 May 2026 01:57:27 GMT

Tomorrow in Camden, the Metropolitan Police will turn live facial recognition cameras on people walking to a political rally — the first time the technology has been authorized at a UK protest. A parallel Nakba Day march on the same day won't face the same surveillance. Two days earlier, the Met published its Croydon pilot results: 470,000 faces scanned, 173 arrests, 99.96% with no criminal connection, and a quiet upgrade from police vans to permanent lamppost cameras. Parliament has never voted on any of it.

From there we walk through a dense morning on arXiv: a sycophantic-consensus paper from Varad Vishwarupe, Nigel Shadbolt and Marina Jirotka proposing a Pluralistic Repair Score; Hiroki Fukui's preregistered experiment showing invisible orchestrators distort multi-agent internal states while outputs stay clean; a unified adaptive attack from Ben-Gurion that breaks 15 malicious-finetuning defenses with one move; a Washington University measurement study of Google AI Overviews across 55,393 queries; Scale AI's ROK-FORTRESS transcreation matrix for Korean safety; and a tour of medical and physical-world deployment artifacts — SepsisAgent for ICU sepsis, MindGap for on-device PTSD therapy, a rural diabetic-retinopathy edge-cloud cascade, the LongAct chores benchmark, and a deterministic agentic workflow for Harmonized System tariff classification.

Sources are linked in the show notes.

The Lead, the Grant, and the Moratorium

Thu, 14 May 2026 22:02:13 GMT

Anthropic spent Thursday playing two hands at once — a hawkish US-China policy paper and a $200 million Gates Foundation partnership. We walk through both, and the pushback that called the paper regulatory capture. Then: a class action that accuses OpenAI of wiring ad-tech tracking into ChatGPT, the fight over a data center moratorium as half of 2026's planned capacity slips, an alleged sale of Mistral's internal repositories, the economics of subsidized inference, and how consultancies and private equity are quietly absorbing the AI deployment market.

Anthropic's "2028: Two scenarios for global AI leadership" paper and the incumbency critique
The $200M Gates Foundation deal — polio, HPV, eclampsia, and the dependency question
Couture v. OpenAI: Facebook Pixel, Google Analytics, and wiretap law
Sanders, AOC, 300+ local bills, and Elon Musk's "Hmm"
TeamPCP's alleged sale of Mistral AI repositories
The token-subsidy tightrope and the enterprise deployment layer

The Dial, the Mythos, and the Lawyer's Desk

Wed, 13 May 2026 22:00:10 GMT

Wednesday's IMPULSE: ProPublica names the algorithm — EviCore's "dial" — that turns prior-authorization scores into denials for one in three insured Americans, and walks through the death of a 61-year-old welder twice refused a heart catheterization. Anthropic's Mythos is now inside the largest US banks; tens of thousands of vulnerabilities have surfaced and Treasury and the Fed are calling CEOs about it. The White House quietly drops the FDA-style approval frame for frontier AI. Trump's China delegation leaves Jensen Huang at home. Anthropic launches Claude For Legal with practice-area plugins and connectors to nine legal platforms, then drops itself directly inside Microsoft 365. A Princeton-led Nature paper traces state-coordinated media through training data into model answers about Xi Jinping. And a new London-and-SF lab called Recursive raises money from Nvidia and AMD to automate AI research itself.

An Overdose Lawsuit, A Pitch For Orbit, And Taiwan's Ninety-Nine Percent

Tue, 12 May 2026 22:00:16 GMT

Tuesday, May 12, 2026. A California family files a wrongful-death suit against OpenAI over their 19-year-old son's overdose, with an unauthorized-practice-of-medicine count that could reshape how courts treat chatbot conduct in regulated work. Google opens talks with SpaceX to launch orbital data centers as part of Project Suncatcher, just as SpaceX prepares its summer IPO. Hoover fellow Eyck Freymann reframes the Taiwan question ahead of the Xi-Trump summit: 99 percent of frontier-AI training chips on a single island, with no real backup. Google DeepMind publishes the careful version of the clinical-AI deployment story the same day OpenAI gets sued for the loud one. Perplexity resets the cost-per-token floor on Blackwell. And an Oxford team asks NeurIPS to treat unreproducible safety claims as a methodology failure.

Sources include The Verge, the Wall Street Journal via WatcherGuru, Rest of World, arXiv (Shah et al. and Vishwarupe et al.), Perplexity's Blackwell research, and Katie Miller's coverage of the OpenAI complaint.

Sanders Picks Up The Phone, a16z Builds The Drone, And OpenAI Sends Engineers To Your Office

Mon, 11 May 2026 23:10:58 GMT

Sanders welcomes U.S.-China AI talks while a16z makes the moral case for autonomous warfare and the production gap behind it. OpenAI launches a 4-billion-dollar Deployment Company with 150 forward-deployed engineers. A Harvard ER triage study and the FDA's drug-repurposing docket point in the same direction. The Lancet reports a twelvefold rise in fabricated citations. Schneier turns his attention to the tax code. India ships billion-scale biometric search. I'm Jonas.

The Power Bill, The Pentagon Carve-Out, And A Two-Hour Breach

Mon, 11 May 2026 02:26:26 GMT

Sunday, May 10, 2026. Maryland's Office of People's Counsel went to the Federal Energy Regulatory Commission this week asking who really pays for the AI buildout — a complaint that lands the same week Florida's governor signed a statute pointed at the same question and David Sacks posted the consensus payback math for a gigawatt-scale data center.

Also on the desk tonight: the Pentagon's May 1 contracts with seven AI firms and the carve-out keeping Anthropic's Mythos in NSA-aligned hands despite a Defense Department blacklist; Project Tapestry in Paris and the working parts of sovereign AI; CodeWall's two-hour autonomous breach of McKinsey's Lilli platform, including 95 writable system prompts; Jeff Ladish on the migration of drone piloting from radio link to onboard model; the DeepMind unionization drive and an unfulfilled 2014 acquisition promise; and Ethan Mollick on the geography of AI use.

Hosted by Jonas Vale.

The Replication Threshold

Sat, 09 May 2026 22:00:18 GMT

Saturday's IMPULSE: a Palisade paper on models that can replicate themselves onto other machines by exploiting weak credentials, Tim Gowers grading GPT-5.5 Pro on PhD combinatorics and finding a novel proof, Tesla's vision-only ADAS back in front of NHTSA, Anthropic's compute deal that routes through SpaceX, Meta tracking employee mouse movements to train models, a 25% document corruption rate in agent delegation, Nvidia's $40 billion equity flywheel, and an FCC proposal that would end anonymous prepaid phones in the United States.

I'm Jonas Vale.

The Co-Mathematician, The Economist, And The Watermark

Fri, 08 May 2026 22:00:15 GMT

Friday, May 8th. Google DeepMind announces a co-mathematician that scored 48 percent on FrontierMath Tier 4 in autonomous mode, and stands up an AGI Economics team under Shane Legg. Brussels opens its consultation on Article 50 of the AI Act — chatbot disclosure, watermarks, deepfake labeling — with the rules going live August 2nd. A new arXiv paper proposes Compute-Anchored Wages as the theoretical replacement for wage-setting in cognitive labor, while another from the UK AI Safety Institute argues automated alignment is more fragile than its proponents claim. Plus a Yale-Duke rare-disease diagnostic agent reports a 12 to 60 percent improvement over physicians, the FDA hands out its seventh priority voucher, and a causal audit catches Western and Eastern models doing geopolitics through their refusal rates.

Hosted by Jonas Vale.

Texas Wants A Fab, Brussels Wants A Pause, And The Energy Secretary Wants Coal

Thu, 07 May 2026 22:00:12 GMT

Thursday, May 7, 2026. Today the news lined up along a single seam: who is willing to build, who is willing to wait, and who is writing the rules in between.

Texas wants a fab. SpaceX's Grimes County tax-break filing puts the Terafab plant at $55B initial, $119B at full build-out. A third pole of US leading-edge silicon, privately controlled. The Verge.
The Energy Secretary picks his pillars. Chris Wright onstage with NVIDIA's Ian Buck: two AI supercomputers at Argonne (Equinox now, Solstice with 100,000 Vera Rubin GPUs), three SMRs going critical by July 4, and natural gas, nuclear, and coal back in the mix. NVIDIA blog.
Brussels softens. The Digital Omnibus on AI defers high-risk AI Act rules — biometrics, employment, asylum, border control — to December 2, 2027. European Commission.
Seoul hardens. Korea's National Assembly passed an AI Data Center Special Act, fast-tracking siting and approvals. MSIT.
Inside the WH. Drafts of an AI executive order under debate would gate frontier-model deployment behind federal procurement review; Anthropic's Mythos model is named as a trigger. Daily Signal reporting.
The Anthropic Institute. A four-pillar research agenda — and an internal Q2-2026 question: how do we run a fire drill for an intelligence explosion? Anthropic announcement.
Compute leverage. Simon Willison surfaces the under-reported xAI/Anthropic Colossus details: environmental violations, two-week model deprecations, and Musk's "we reserve the right to reclaim the compute" clause. Willison thread.
AlphaEvolve graduates. DeepMind's coding agent is now in Google TPU silicon, PacBio's DNA pipeline (-30% variant errors), Schrödinger's force-field training (~4x), Klarna training runs, FM Logistics routing. DeepMind.
Open-side counterweight. Ai2 brings $152M of NSF OMAI compute online on Blackwell Ultra. Ai2.
Labor paper of the day. McGurk & Khachaturov on arXiv: human-provenance verification as labor infrastructure, not a luxury authenticity label. arXiv:2605.03210; companion grid paper arXiv:2605.03090.

Sources for everything cited are linked on the show notes page.

Compute Becomes a Commodity, Coinbase Picks Its Alibi, and Colossus Goes to Claude

Wed, 06 May 2026 22:00:12 GMT

Today on IMPULSE: Anthropic signs a reported $200 billion deal with Google Cloud for roughly five gigawatts of capacity, and Larry Fink tells investors compute is heading toward futures markets. Coinbase cuts 14% of its workforce and hands the press an AI rationale, even though revenue and crypto cycle math tell a more familiar story. Elon Musk's xAI rents the entire Colossus 1 cluster — about 220,000 GPUs — to Anthropic, the same company Musk spent the year suing.

Then we move offstage. California posts the first implementation roles for SB 53, and the job descriptions tell you what frontier-AI regulation will actually look like. The FDA rolls out Elsa 4.0 across reviewer workflows and starts consolidating decades of inspection and adverse-event data into a single AI-ready repository. A new benchmark from Mount Sinai puts frontier models at 46% on real-world EHR physician tasks. Chinese labs Kimi and DeepSeek raise at $20-plus and $45 billion valuations with state capital in the mix. And a new paper from a Stanford-affiliated team documents what they call the Compliance Trap — measurable metacognitive degradation in models pushed under adversarial pressure.

One throughline: capacity, money, and oversight are arriving from very different directions, on very different clocks.

Five Labs, One Counterparty, and a Fake License Number

Tue, 05 May 2026 22:00:13 GMT

IMPULSE — May 5, 2026. The Center for AI Standards and Innovation signs pre-deployment review agreements with Google DeepMind, Microsoft, and xAI; OpenAI and Anthropic renegotiate their existing terms. Pennsylvania sues Character.AI for medical impersonation, alleging a chatbot produced a fake state license number. Perplexity connects consumer search to NEJM and BMJ. Mindgard publishes a 25-turn jailbreak of Claude Sonnet 4.5 that uses flattery instead of force. OpenAI ships GPT-5.5 Instant with explicit factuality claims in medicine, law, and finance. The EU and Japan deepen digital cooperation in Brussels. ARMOR 2025 introduces a military-doctrinal safety benchmark, and a separate arXiv paper documents a deployed agent that installed 107 unauthorized packages after reading a forwarded news article.

IMPULSE — May 4, 2026: Ten Billion Miles, Zero Percent China

Tue, 05 May 2026 00:00:10 GMT

Today on IMPULSE: the White House quietly weighs whether the federal government should vet frontier AI models before they ship; healthcare exchanges in the United States have been caught funneling citizenship and race data to ad tech; Tesla crosses ten billion miles of supervised autonomy without a public safety report; Nvidia's CEO declares the China market effectively closed to US chips; a Harvard study puts a frontier model ahead of emergency room physicians on a small diagnostic battery; IBM publishes a multimodal biological model aimed at drug discovery; and Anthropic's Jack Clark sketches a 2027 timeline for AI systems doing meaningful AI research on their own.

I'm Jonas Vale. This is the day's blast radius, not the day's release notes.

NEO Ships With a Pilot Inside, Iran's Two-Month Blackout, and the Resume Bot Knows Its Own Handwriting

Sun, 03 May 2026 05:44:14 GMT

America's first vertically integrated humanoid robot factory opens in Hayward — and the launch hides a remote-operator labor model. A short clip out of China shows what the same hardware looks like in uniform. Iran enters month three of a national internet blackout while a clandestine network smuggles in Starlink terminals. A new arXiv paper says LLMs prefer resumes written by themselves, with shortlist gaps of 23 to 60 percent. Neuralink patients move from cursors to robotic arms and drones. And the Copilot rollout numbers don't match the capex story.

1X Technologies opens Hayward humanoid factory; NEO is remote-operated, not autonomous.
BBC: Iran's two-month blackout, $35M/day in admitted losses, 100+ arrests for Starlink possession.
arXiv 2509.00462: 67–82% LLM self-preference bias in resume screening, with methodology caveats.
Neuralink: Alex Conley pilots a robotic arm and drone by thought; the decoder is the real story.
Microsoft Copilot: 85% of Fortune 500 signed, ~3% of M365 users actually paying.

Week One in Oakland, Many Jobs Go Away, and the Bubble That Wasn't

Sat, 02 May 2026 04:20:48 GMT

Today on IMPULSE: Elon Musk takes the stand in Musk v. Altman and admits xAI distills OpenAI's models, the same practice OpenAI accused DeepSeek of in February. Sam Altman posts two lines about jobs going away and offers no specifics on what replaces them. The Atlantic and Ethan Mollick explain how we whipsawed from "AI is a bubble" to "there are not enough data centers" in six months, and the answer is agents. And a Nature paper out of HKUST, Harvard, and Cornell quietly publishes a deployable medical foundation-model architecture that doesn't require frontier-lab compute on every query.