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.

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 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.

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.

• Maryland's $2B grid-upgrade complaint at FERC (Tom's Hardware)
• DeSantis signs SB 484 on data-center cost allocation (CBS12)
• David Sacks on gigawatt payback math
• David Sacks on the AI Action Plan and pre-market regulation
• crepesupreme on procurement-as-vetting
• Emil Michael on Mythos as a national-security moment (CNBC)
• Pentagon's seven-firm GenAI.mil contracts (SiliconANGLE)
• Tom English on the May 1 timing
• Alper Ferudun on the sovereign AI stack
• CodeWall's writeup of the Lilli breach
• Jeff Ladish on drones and onboard pilots
• Andreas Kirsch on DeepMind unionization and the 2014 ethics board
• Ethan Mollick on the geography of AI use

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.

• Palisade Research
• Tim Gowers's blog
• The New York Times
• TechCrunch
• FCC
• Hacker News
• Reddit

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.

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.

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

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.

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.

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.