Everything That Happened in AI Today Tuesday, April 14, 2026

A 20-year-old Texan was charged with attempted murder and arraigned Tuesday in San Francisco for Friday's Molotov attack on Sam Altman's home; his manifesto contained a kill list of other AI executives, and his public defender is now claiming autism and acute mental health crisis. Anthropic published research showing Claude Opus 4.6 outperforming its own human alignment researchers (PGR 0.97 vs human 0.23 at $22 per agent-hour), is reportedly fielding VC offers at up to an $800B valuation (more than double its current mark), and shifted enterprise billing to usage-based pricing amid the compute crunch. Maine became the first state to ban large data centers, Anthropic also shipped Claude Code Routines and the redesigned Claude Code desktop app, leaked Claude Opus 4.7 plus a design tool to The Information, a leaked OpenAI internal memo from CRO Denise Dresser accused Anthropic of inflating its $30B run rate by ~$8B, and Nvidia Blackwell GPU rental prices jumped 48% in two months.

Welcome to the Around the Horn Digest, the one page you need to sound dangerously informed at work tomorrow.

Today the story everyone is reading is no longer about benchmarks or fundraising rounds; it's about the fact that the U.S. Attorney for the Northern District of California is preparing to charge an alleged AI extremist under domestic terrorism statutes, and the manifesto recovered from his backpack named other AI executives as targets. The "AI elite vs. everyone else" gap that Stanford measured yesterday has stopped being a chart in a PDF and started being an FBI investigation.

Meanwhile, Anthropic is having the most chaotic single day of its corporate life: it published research showing Claude Opus 4.6 outperforming its own human alignment researchers (yes, really), shipped Claude Code Routines, shipped a full Claude Code desktop redesign, leaked Claude Opus 4.7 plus a Figma-killer design tool to The Information, added Mythos to Amazon Bedrock, hired Trump-linked Ballard Partners to lobby on its Pentagon fight, and put the Novartis CEO on its board. OpenAI counter-positioned with GPT-5.4-Cyber and a scaled Trusted Access program for thousands of defenders, the exact opposite of Anthropic's "gate Mythos" approach. Also, OpenAI's CRO got caught in a leaked memo claiming Anthropic is faking its revenue numbers. Oh, and Maine just made AI's grid problem an actual zoning ordinance. Just a normal Tuesday, humans!

Let's get into it.

Monthly skill digests: AI Skill Digest — April Week 1

Around the Horn — Wednesday, April 15, 2026

So Sam Altman's San Francisco home was firebombed early Friday, the suspect was arraigned Tuesday, and the manifesto recovered at the scene named other AI executives as targets.

Here is the timeline as of Tuesday afternoon.

Friday, April 10, ~3:37 a.m.: A 20-year-old Texan named Daniel Moreno-Gama allegedly threw a Molotov cocktail at the metal gate of Altman's Lombard Street home; an exterior gate caught fire, security extinguished it, and Moreno-Gama allegedly walked the three miles to OpenAI's Mission Bay headquarters and tried to break the glass doors with a chair, telling security he came "to burn it down and kill anyone inside" (CNBC).
Sunday, April 12, ~1:40 a.m. (now ruled unrelated): A Honda sedan with two occupants stopped near the property and the passenger allegedly fired a round; SFPD arrested Amanda Tom, 25, and Muhamad Tarik Hussein, 23 and booked them on negligent discharge charges. At Monday's news conference, SFPD Chief Derrick Lew said investigators have "no evidence to suggest this incident was related to the case on Friday" and that the occupants had driven "past a residence when the gunfire erupted"; an OpenAI spokesperson separately told Fox News Digital the incident was unrelated to Altman and that there was no indication his home had been targeted.
Tuesday, April 14: Moreno-Gama made his first court appearance in orange jail attire, softly answered "yes" to continuing his arraignment, and was ordered held without bail; Judge Kenneth Wine set his next hearing for May 5. His public defender Diamond Ward called the case "a property crime, at best," said her client is autistic and was experiencing an "acute mental health crisis," and accused prosecutors of overcharging "to curry favor for Altman." SF District Attorney Brooke Jenkins pushed back at the same courthouse, calling it "a targeted attack on Mr. Altman" and saying prosecutors would act identically whether the victim was "a billionaire or a CEO or any average San Franciscan."

The Texas case is the more serious one and it just got worse. On Monday, SF District Attorney Brooke Jenkins charged Moreno-Gama with two counts of attempted murder (one for Altman, one for a security guard at the residence) plus nine other state charges; he separately faces federal charges of attempted destruction of property by means of explosives. SFPD recovered "incendiary devices, a jug of kerosene, a blue lighter, and a document" at the scene. The document was a three-part manifesto. Part 1 was titled "Your Last Warning" and, per the federal complaint, Moreno-Gama "stated he 'killed/attempted to kill'" Altman and listed the names and addresses of "several additional AI executives, board members and investors." Part 2 was "Some more words on the matter of our impending extinction." Part 3 was a letter directly to Altman: "if by some miracle you live, then I would take this as a sign from the divine to redeem yourself." US Attorney Craig Missakian: "if the evidence shows that Mr. Moreno-Gama executed these attacks to change public policy or to coerce government and other officials, we will treat this as an act of domestic terrorism and together with our law enforcement partners prosecute him to the fullest extent allowed by law."

The trigger everyone is pointing at is a New Yorker article published five days before the first attack. On April 7, Ronan Farrow and Andrew Marantz published "Moment of Truth: Sam Altman May Control Our Future—Can He Be Trusted?", an investigation built on more than 100 interviews and previously unreleased internal documents. The piece reportedly quotes Slack messages from Ilya Sutskever listing "lying" at the top of Altman's "consistent pattern" of behavior, plus private notes from Dario Amodei whose top-line judgment was "the problem with OpenAI is Sam himself." Sources interviewed for the article described Altman as having "a relentless will to power"; one anonymous board member is quoted saying he "balances a desire to be liked with a lack of concern about the consequences of misleading others." Several employees reportedly used the word "sociopathic." Altman's Friday-evening blog post, published hours after the Molotov attack, called the article "incendiary" and explicitly linked it to the violence: "Someone said to me yesterday they thought it was coming at a time of great anxiety about AI and that it made things more dangerous for me. I brushed it aside. Now I am awake in the middle of the night and pissed, and thinking that I have underestimated the power of words and narratives."

Altman's full response is the most personal thing he has ever published. He shared a family photo, acknowledged he is "conflict-averse" and that this trait "has caused great pain for me and OpenAI," apologized for "handling myself badly in a conflict with our previous board" in 2023, and described the rivalries inside the AI industry as a "'ring of power' dynamic" that "makes people do crazy things." The closing line, which now reads very differently after Sunday's shooting: "While we have that debate, we should de-escalate the rhetoric and tactics and try to have fewer explosions in fewer homes, figuratively and literally."

Why this matters more than any of the other AI stories today: the Stanford 2026 AI Index we covered yesterday measured the gap between AI insiders (56% excited about AI) and the public (10% excited) and called it the most important finding in the report. That gap is no longer a chart. There is now a 20-year-old in custody who allegedly drove from Houston to San Francisco with a kill list of AI executives and a kerosene jug, and whose public defender is already reframing the case at arraignment as the act of an autistic young man in acute mental health crisis rather than a politically motivated attack. There is a federal prosecutor preparing a domestic-terrorism case anyway. And there is a CEO who, two days before the violence started, was being publicly described as a "pathological liar" by sources that included his own former co-founders. The "AI is too important to be slow about" argument and the "AI is hurting people and someone has to stop it" argument has officially reached a new level. TBH, the next few months of AI coverage leading up to the US midterms (where the fate of US AI regulations somewhat hang in the balance) are going to be unlike anything since the launch of ChatGPT.

In yesterday’s main story, we ended on this paragraph:

Andrew Curran's argument from this weekend hits even harder in this light; he points out that frontier models like Mythos are now so expensive to serve that they have to be rationed to a few dozen mega-customers, which means the people most worried about AI will be the people most cut off from it. This gap doesn't close on its own. Somebody has to actively close it; we here at The Neuron believe that means open source, freely available local AI built into every new computer.

If you'll allow us, we want to expand on that last part (if you WON'T allow us [lol], use the table of contents above or scroll ahead to the round-up):

In five years, we believe it will just be commonplace (and non-controversial) that you talk to your computer and get it to do just about anything for you. It’s only a matter of time before local capability reaches Opus 4.6 level (which to me, can do basically anything I need it to do with enough prompting and context); this capability available to everyone is necessary to disperse the benefits of AI equitably when the corporations all eventually have to raise prices. There will likely still need to be a re-ordering of the social contract between big business and the working class (Sam Altman has thoughts on that). As for the meaning crisis Sam Lessin wrote about, that’s an issue everyone must yet solve for themselves…

If indeed AI pushes the world towards material abundance like many believe (and like what has happened over many previous technological waves), then from that future point of view, it will seem even more ludicrous to firebomb someone's home for trying to create this technology. The biggest gap between the public and AI technologists comes down to, as one Neuron reader wrote in yesterday, "Insiders in AI are confident in their own use of it, whereas the general public is lacking confidence that leaders in AI will do what they say they will do with it (improve healthcare instead of automating away customer service, for example). Those are radically different metrics that have nothing to do with either group's confidence in AI capabilities.”

So really, the gap in confidence between Ai insiders and the general public who disapprove is a lack of belief in the good faith and will to do what's in the public interest of those who work in big corporations. And after many years of being proven right about that (research shows that even when 70% or 80% of the American public favors a new policy, it usually doesn’t happen. It only happens maybe 35% or 40% of the time; when corporations favor a new policy, that statistic is flipped), can you blame them?

Well, regardless of how you or I personally feel about that fact, it’s important to acknowledge what’s happening now is a classic systems problem, in that every actor will act rationally given the system they are in:

The companies will try to preserve their profitability and grow their business.
The workers will try to stop the march of big business towards removing their ability to meaningfully improve their circumstances (via automating it away).
And the government will, slowly, EVENTUALLY, do what's in the public interest after enough public outcry has been levered at them that it is impossible for them to be re-elected without meaningfully regulating the industry.
1. In the meantime, they will do what’s in their personal best interests, which is whatever gets them the money they need to get re-elected.
2. In some cases, that is accepting industry lobbying money.
3. In other cases, it’s a genuine belief in the need for their country (America, in this instance) to win the AI race, and that is of utmost importance to the general public.
4. Or as the classic meme goes: Porque no los dos??

But let’s play out what happens when lobbying does get involved as regulations do get passed; in most cases where industry has a say in how they are regulated, the biggest companies end up enshrining their current positions at the top around a wall of regulations that prevent upstarts from breaking in (Bill Gurley has a great talk on this). In that scenario, the lobbying of the industry leaves the control of AI still in the hands of a few, and every model gets the Mythos treatment moving forward. Or, going even further with regulation, the government puts the ultimate control of AI in the hands of the government itself.

Are either of these scenarios a horrible outcome? Not necessarily (although IDK anyone of any political persuasion who really wants the government to completely control AI). But if either happens, they must also be predicated by equally powerful, democratizing AI capability built into every day computers so the average person need not rely on expensive, metered, cloud-based, data-center run AI models for the intelligence they need to fulfill their personal goals and aspirations.

It should be possible to eventually have Claude Opus 4.6 level intelligence on an average home computer. In fact, I believe in five years this will be the reality of the world (if not sooner, knowing this industry). See, the reason most people don’t like AI today is because it’s previously been dumb and bad. When they actually get to use the level of intelligence I have, as a creator in this space, they’ll find all sorts of acceptable use-cases for it. Everyone will be different about what they do or don’t want AI to do for them. And that’s okay. But everybody has at least one use-case.

And in 10, or 15 years, most everyday people will be happy with that level of local intelligence made available to them. Who would actually be “anti” a computer that’s actually easier to use and more helpful? It will become the baseline expectation. Why wouldn’t you want your computer to be as smart as possible? In that scenario, then Mythos-class models can and should be reserved for research and development, and not something that is served to just anyone.

But if the industry somehow lobbies against open, local intelligence generally available to the public, then we'll all be reliant on cloud-based models and beholden to the tech giants (who people already distrust, because they get their way more than the average person does), while at the same time, those same giants slowly automate large swaths of tasks that used to be the bulk of high wage work for humans to aspire to earn, if not already earn, and also hire less humans themselves overall.

In our humble opinion:

The best of outcomes would be not unlike the world we live in today: a world where intelligence is available at every level of compute, and you just have to find the level that's best for you and your use case. And critically, a world where you can always scale up to a higher tier if and when you need it.
The worst of outcomes would be a world where that same intelligence in best outcome world is actually highly restricted and only made available to a certain class of citizen, either restricted by their net worth, political connections, or some sort of expensive degree.

Now, I could see a world where the ability to access and work with high intelligence AI is something you get a license for, like how we get licensed to drive a car. But I could also see a world where the models are so aligned and easy to use (just literally talking to your computer) that there need be no license required. It would, could, and should be as simple to use AI in the future as buying a computer and saying hello, and suddenly a world of infinite possibility is open to you.

But what does alignment (the word we use to define how well AI follows our instructions and doesn’t do bad things) look like in this state? Are the models perfectly aligned such that they’ll never comply with a suspicious request? Or do we build a stronger surveillance state, where all your chats are logged, even locally?

In a cloud-based world, where all intelligence is metered over the cloud, that's basically how it works today. Again, not necessarily a bad thing from the overall system point of view (it encourages everyone to be on their best behavior)... except if you need to handle sensitive information, or if those who have access to your data want to control you with it.

The alternative, better world would be one where you have ultimate control of your data local on your system (and when you go online, for that matter), and you choose if and when you want to share that data, or if and when you want to tap into the cloud capabilities for higher level of intelligence, and that intelligence is metered and made available to you at an affordable rate based on your needs without any degradation in quality.

I can’t believe I’m saying this, but the Apple Intelligence path is the best path. It is the right path. It always was the right path, it was just promised too soon and offered too little. On this path, if you need to handle sensitive information, you have a local model built into your computer that only you control to do that for you. If you need to then interface with another model over the cloud, then there's a sensitive hand-off to do so and the sensitive information is redacted.

This level of system requires trust in the models, and trust in each other. We are slowly making progress in being able to trust the models. And we'll always have degrees of trust in each other that can and will be broken. No getting around the human element, unfortunately. This is a complicated systems problem to solve.

Some like Sam Altman believe it requires a new social contract between business, government, and the general public. Others are merely acting rationally given their role in that system (company, worker, regulator). Yet others are acting entirely irrationally, letting fear drive their decisions towards violence. And many more let fear decide they can't or won't speak out about the topics they believe in or the current behavior that they don’t. Some of the most important players in the system let fear drive them to push at all costs to develop this technology faster than anyone else, in order to pull up the ladder behind them and build a moat around their supposed, eventual "ring of power", as Sam put it.

We at The Neuron just want there to be equilibrium in the system. At the local level, we want people to be able to use these tools to improve their lives (not eventually get replaced by it), and at the society level, we want this research to benefit all of humanity and progress science, to help us tackle some of the hardest challenges yet facing our civilization. And yes, we want the business models and regulations to be sustainable so that everyone involved (customers, businesses, and people who don't use it at all) can continue to live happy, successful lives (and the economy not completely crash from an AI bubble burst). Insofar as the actions of everyone involved push us closer towards those goals, then we are all better for it. When the actions of individual actors push us further from that goal, then we are all worse off; at the individual level, because of how we are personally impacted, and at the system level, because of how it creates an unsustainable negative feedback loop pushing one direction or another to the limit until something gives.

Let’s not forget there is an election in the United States this year. How that unfolds might tell us more about the trajectory of the industry and its ability to impact our larger society than any new paper or model drop.

🏆 TOP 4 NEWS (Around the Horn)

Maine became the first state in the country to ban new data centers larger than 20 megawatts; the moratorium runs until November 2027 and Gov. Janet Mills is reportedly considering a partial veto over a Jay paper-mill carve-out.
Anthropic shipped Claude Code Routines, letting you define a prompt + repos + connectors that run automatically on a schedule, via API call, or on GitHub events from Anthropic-managed cloud infrastructure (even when your laptop is off); each routine gets its own per-routine API endpoint.
A leaked OpenAI internal memo from Chief Revenue Officer Denise Dresser accused Anthropic of inflating its $30B run rate by ~$8B, failing to secure enough compute, and "building its narrative on fear and restriction" — while touting OpenAI's new AWS partnership, "Spud" model, "Frontier" agent platform, and "DeployCo" deployment engine.
Nvidia Blackwell GPU rental prices jumped 48% in two months, from $2.75 to $4.08 per hour per the Ornn Compute Price Index, as AI energy demand forces companies to ration computing access.

Honorable Mentions

🚨 Anthropic published research showing Claude Opus 4.6 outperformed human alignment researchers on the weak-to-strong supervision problem (PGR 0.97 vs human 0.23, $22 per AAR-hour, Alignment Science blog). Andrew Curran is calling it "a preview of RSI." Full breakdown in Research/Models below. Full breakdown from us here.
Anthropic shipped the redesigned Claude Code desktop app (the "Epitaxy" overhaul that leaked last week) and is reportedly preparing Claude Opus 4.7 + a prompt-based design tool for release this week per The Information. Adobe, Wix, and Figma traded down on the news.
🚨 OpenAI launched GPT-5.4-Cyber and scaled Trusted Access for Cyber to thousands of verified defenders (OpenAI blog). This is the direct OpenAI counter to Anthropic's Mythos cybersecurity positioning, with binary reverse engineering capabilities for vetted security teams. Same week as Mythos goes to Bedrock and the Treasury, Fed, and UK AISI all queue up for Mythos access. Full breakdown in Big Tech.
Festus, Missouri voters fired half its town council in last week's election and are now collecting recall signatures to oust Mayor Sam Richards and the remaining four members after the council approved a $6 billion CRG/Clayco datacenter project on March 30; residents have also filed a lawsuit alleging illegal rezoning and Sunshine Law violations.
The US Treasury is now seeking access to Anthropic's Mythos to hunt for vulnerabilities, per Bloomberg. That puts US Fed + UK AISI + UK financial regulators + Treasury + Trump administration officials all in motion on Mythos in a single week.
POLITICO Europe reports European regulators have been sidelined on Mythos. Apple, Microsoft, Anthropic and others are running Project Glasswing without giving European oversight bodies the same access. EU regulators have limited insight into a system that could pose major cybersecurity risks on the continent.
Anthropic added Novartis CEO Vas Narasimhan to its board (WSJ exclusive) — its second board addition in recent months as it eyes a potential IPO this year and accelerates its enterprise push into healthcare.
Anthropic also hired Trump-linked lobbying firm Ballard Partners (Bloomberg) as it draws out its fight with the Pentagon over the supply-chain risk designation.
Anthropic Mythos Preview is now on Amazon Bedrock (gated research preview) per an AWS launch note — quietly broadening enterprise access through Bedrock's existing approval workflows.
MIT Technology Review is launching "10 Things That Matter in AI Right Now," a brand-new annual list debuting April 21 at EmTech AI on MIT's campus.

🍪 TOP TREATS TO TRY

Claude Code Routines lets you define a prompt + repository + connectors that Claude Code runs automatically on a schedule, via API call, or on GitHub events (e.g., PR merge → auto-update docs) — all from Anthropic-managed cloud infrastructure even when your laptop is off, with a per-routine API endpoint for alerts/Zapier integrations —free with every Claude Code plan.
Google Chrome Skills lets you save any Gemini prompt (or pick from 50+ premade ones) as a one-click reusable workflow that runs on the current tab or multiple selected tabs via the / or + sidebar shortcut — for example, vegan recipe substitutions, side-by-side shopping comparisons, or scanning long contracts —free, rolling out to English (US) desktop today.
Tradclaw is a ready-to-run OpenClaw household scaffold from claire vo (the "tradwife we all deserve") that triages school emails, plans weekly meals + grocery lists, logs homework from photos, tracks the home book library, reminds about cleaner/sitter payments, surfaces home maintenance, and writes custom bedtime stories — install by sending one prompt to your OpenClaw instance, then let it interview you (GitHub) —free to try.
Sparkle v4 is a Mac filesystem cleaner agent from Dan Shipper that examines your drive, deep-cleans junk/duplicates/screenshots/installers, then runs quietly in the background on a schedule to keep everything organized for how you actually work —free during an Every subscription.
Hermes Agent v0.9.0 ("The Everywhere Release") from Nous Research adds a local web dashboard, Termux/Android support, iMessage integration, background process monitoring, an improved skill manager, and backup/import tools to its open-source agent stack —free to try.
Seedance 2.0 in ElevenCreative is ByteDance's new video model now integrated into ElevenLabs' creative suite, generating motion + dialogue + music + effects in one pass from any combination of text, image, video clip, and audio reference inputs, with edit/extend support for regenerating sections without starting over —included on all paid ElevenLabs plans.

🏢 Big Tech & Major Companies

OpenAI launched GPT-5.4-Cyber and scaled its Trusted Access for Cyber (TAC) program (blog post) — a fine-tuned, cyber-permissive variant of GPT-5.4 with lowered refusal boundaries for legitimate security work and new binary reverse-engineering capabilities that let analysts probe compiled software for malware, vulnerabilities, and security robustness without source code access. TAC is opening to thousands of verified individual defenders and hundreds of teams via chatgpt.com/cyber; GPT-5.4-Cyber sits in the highest tier and is gated to vetted vendors, organizations, and researchers. OpenAI also disclosed that Codex Security has now contributed to over 3,000 critical and high-severity vulnerability fixes since launching as a research preview earlier this year. This is the OpenAI counter-positioning to Anthropic Mythos that everyone was waiting for. Anthropic's pitch: "we built something so cybersecurity-capable we have to gate it." OpenAI's pitch today: "we built something cybersecurity-capable and we're handing it to thousands of defenders." Same week, opposite philosophies. Watch how this lands with the Treasury, the Fed, and UK AISI, all of whom are simultaneously demanding access to Mythos.
GitLab expanded its Google Cloud partnership to power GitLab Duo Agent Platform with Vertex AI models (GitLab blog, about.gitlab.com) — agents inside Duo Agent Platform can now call Vertex AI models (including Gemini) natively, with every agent action flowing through GitLab's existing access controls, approval rules, and audit logging. Customers with Google Cloud commitments can count GitLab Duo Agent Platform usage toward their existing spend, and self-hosted teams can run GitLab's AI Gateway on GKE or Cloud Run. From GitLab CPMO Manav Khurana: "AI agents are only as good as the context they operate on and the governance around them. GitLab is where that context lives across issues, code, pipelines, security findings, and this partnership connects it to Vertex AI's strongest models. As agents take on more of the software lifecycle, the platform that provides both the context and the controls becomes the critical layer." Watch this framing. Every enterprise software company this year is racing to be "the platform that provides both the context and the controls" because that's the durable position once foundation models commoditize. GitLab + Vertex is the same pattern as Cursor + Sentry, Notion + custom models, Datadog + agent observability: agentic workflows landing inside the system of record where the context and the audit trail already live.
TestingCatalog leaked Google's NotebookLM Canvas + Connectors update (thread) — Canvas mode lets users generate any visual representation of data from notebook sources, and a hidden Connectors option in settings hints at integration with Drive, Gmail, and other Google data sources. Likely Google I/O announcement. NotebookLM is the only Google AI product that punches above its weight on consumer love. Adding Canvas + Connectors could turn it into the Cowork competitor Google has been struggling to ship.
Anthropic shipped Claude Code Routines today (docs, Noah Zweben thread, 9to5Mac) — define a prompt + repos + connectors once and trigger templated agents on schedule, GitHub events, or via the new per-routine API endpoint, all running on Anthropic's web infrastructure even when your laptop is off; unlimited routines on every plan with daily runs included.
Anthropic shipped the redesigned Claude Code desktop app today (Felix Rieseberg announcement), the overhaul that leaked last week as "Epitaxy." New features: integrated terminal, in-app file editing, rebuilt diff viewer, side chats, SSH connections, and a ground-up rebuild for parallel work. Felix says it's been his main way to use Claude Code for the last few weeks. Available now via the Claude desktop app download. Two desktop shipping moments from Anthropic in one day: Routines this morning, the new Code desktop UI this afternoon.
Anthropic is also reportedly preparing Claude Opus 4.7 + a prompt-based design tool for websites and presentations, per The Information via Chubby — both could ship this week. The design-tool angle reportedly sent Adobe, Wix, and Figma stocks down >2% in after-hours trading; The Information's exclusive frames it as Anthropic moving directly into Gamma/Beautiful.ai/Figma territory. That's three model releases (Routines, Code desktop, Opus 4.7) and a brand-new product category from one company in a single week, while also fielding the Mythos cybersecurity preview, the Ballard Partners lobbying play, and the Novartis CEO board appointment. Anthropic's release cadence is the actual story this week.
Anthropic added Novartis CEO Vas Narasimhan to its board (WSJ). That's the second board addition in recent months as the company prepares for a potential IPO and accelerates its healthcare push.
Anthropic hired Trump-linked Ballard Partners as its lobbying firm, per Bloomberg, as it continues its fight with the Pentagon over the supply-chain risk designation. The political-realignment subplot inside the AI safety story is becoming impossible to ignore.
OpenAI is splitting Codex into Basic and Advanced use cases, with new additions including app preview, scratchpad, and other developer-focused features (discovered by @mweinbach via TestingCatalog).
OpenAI bought AI personal-finance startup Hiro (TechCrunch). Hiro stops accepting new signups today, the product shuts down April 20, and the team joins OpenAI to build financial-planning capabilities into ChatGPT.
A leaked OpenAI internal memo from CRO Denise Dresser (@kimmonismus) accused Anthropic of inflating its $30B run rate by ~$8B, failing to secure enough compute, and "building its narrative on fear and restriction" — while touting OpenAI's new AWS partnership, "Spud" model, "Frontier" agent platform, and "DeployCo" deployment engine. Three things at once: a financial accusation, a roadmap leak, and a window into how OpenAI is positioning internally against Anthropic post-Mythos.
A Microsoft executive suggested AI agents will need to buy software licenses, just like employees, per Business Insider — potentially expanding rather than shrinking SaaS revenue even if human headcount falls. The "AI seat" SaaS pricing model could become an industry default within 12 months.
Satya Nadella shipped a major Copilot-in-Word update today (Satya's announcement, Microsoft Tech Community deep dive) — Copilot now tracks changes, leaves comments, and is "grounded in all your enterprise context with Work IQ," positioning it less as an autocomplete and more as a coworker reviewing your draft inside the document.
MiniMax updated the MiniMax-M2.7 license (Hugging Face commit) to expressly permit personal/self-hosted use, non-profit/academic use, and modification for those purposes free of charge, while still requiring written authorization (and a "Built with MiniMax M2.7" attribution) for any commercial deployment. A useful reminder that "open weights" and "open license" are now two different things, and the strongest non-Western open weights are increasingly source-available rather than fully open.
Microsoft also released GigaTIME publicly on Microsoft Foundry Labs and Hugging Face, a multimodal AI model trained on 40 million cancer cells across 14,256 patients that translates routine $5–$10 H&E pathology slides into virtual multiplex immunofluorescence images. Caveat: the underlying paper was published in Cell back in December 2025; what's new this week is the open-source release.
Google brought Gemini Personal Intelligence to India (TechCrunch), letting users connect Gmail, Photos, and YouTube to get personalized answers (e.g., travel plans pulled from their own data); initially limited to AI Pro/Ultra users with free-tier expansion in coming weeks.
Google added AI Skills to Chrome (TechCrunch, WIRED how-to) — save and reuse Gemini prompts as one-click workflows via the Chrome sidebar.
Google DeepMind released Gemini Robotics-ER 1.6 (blog), upgrading spatial reasoning, multi-view understanding, and success detection — and crucially adding instrument reading (gauges, sight glasses) via agentic vision plus code execution. Already powering Boston Dynamics Spot.
Google is also reportedly building an internal "Agent" Cowork competitor inside Gemini/Gemini Enterprise, per a TestingCatalog leak — new Tasks UI showing Goal, Agent, Connected apps, Files, and a "Require human review" toggle, plus Skills and Projects in development.
Tencent unveiled HYWorld 2.0, an engine-ready world model from Tencent Hunyuan that turns a single image into a fully editable 3D scene with standard splats + mesh (Dylan Wang's announcement). Open-source and free hosted version launching this week.
Baidu released ERNIE-Image, an open 8B-parameter text-to-image model built on a single-stream Diffusion Transformer that hits SOTA among open-weight models on GenEval/OneIG/LongTextBench, with precise multilingual text rendering, complex multi-object instructions, structural coherence for posters/manga, and a Turbo version for 8-step inference (blog, HF, GitHub, @ErnieforDevs).

💼 AI Productivity, Labor & Economics

Aaron Levie argues every team will soon need a dedicated "agent deployer and manager" role (thread) — someone whose job is to identify highest-leverage workflows, redesign them for 100× speed/volume using agents, map data flows, wire systems via skills/MCP/CLIs, manage evals, track KPIs, and own the connection between business systems and the agent layer.
Greg Brockman essay on the compute economy (thread) — argues we are transitioning to a "compute-powered economy" where AI has already sped up software engineering dramatically and will soon transform every computer-based job by removing the need to micromanage machines, letting intent turn directly into software, workflows, science, and companies, while small teams achieve what once required large ones.
Dan Shipper argues software engineering in 2026 needs two roles: a pirate who codes as fast as possible to discover what's valuable, and an architect who turns the sloppy mess into a well-oiled machine (thread). The "10x engineer" trope is being replaced with a 2-role structure.
HBR: "The Hidden Demand for AI Inside Your Company" — Luis Garicano and co-authors argue (thread, HBR) that shadow AI is not a compliance problem but a demand signal. BBVA's playbook: deploy secure ChatGPT Enterprise with FOMO licensing (3K competitive seats, use-it-or-lose-it), build an Adoption Network of Champions/Wizards, and let peer-driven custom GPT creation grow to 11K active users and 4,800+ internal tools in one year.
An NBER survey of 6,000 executives found 90% reported zero AI impact on productivity or employment (analysis by Willian Correa). The vocal 10% who claim huge gains drive most of the headlines, but the boardroom reality is much quieter.
Denis Stetskov: "The Human Cost of 10x AI Productivity" (essay) — AI tools increased code-review volume by 98% while the human brain still processes conscious thought at only 10 bits per second, creating measurable physical toll, burnout, and workload creep on senior engineers who now reverse-engineer machine-generated artifacts at biological speed.

🤖 AI Agents & Infrastructure

Nvidia Blackwell GPU rental prices jumped 48% in two months, from $2.75 to $4.08/hour, per the Ornn Compute Price Index (WSJ). The same WSJ piece reports AI energy demand is forcing the rationing of computing power for users.
Ivan Burazin and Dylan Patel argue GPUs are no longer the primary bottleneck — CPUs are (Burazin thread, SemiAnalysis Datacenter 2026 talk) — agentic models and tightened RL loops now require heavy verification (regex → classifiers → unit tests → full environment calls to DBs/simulations), driving agent revenue from ~$2B to >$10B in six months and causing cloud CPU shortages: GitHub instability, AWS installations 3× YoY, Microsoft selling spare CPUs to OpenAI/Anthropic. The old 100:1 GPU:CPU power ratio is collapsing. This is the missing piece in the cost narrative — alongside Curran's Mythos cost thread, today's Blackwell GPU spike, and Vercel's "70% of docs traffic is now agents." The bottleneck isn't where we thought.
NVIDIA Quantum Day (today, April 14) — and the actual product announcement was NVIDIA Ising, the world's first open-source AI model family for quantum error correction and processor calibration (NVIDIA newsroom, NVIDIA developer blog, Chubby's breakdown, SiliconANGLE). Ising Calibration is a vision-language model that cuts quantum-processor tuning time from days to hours, runs at 15× smaller model size than alternatives, and on the new QCalEval benchmark beats Gemini 3.1 Pro by 3.27%, Claude Opus 4.6 by 9.68%, and GPT-5.4 by 14.5%. Ising Decoding uses two variants of a 3D CNN (one tuned for speed, one for accuracy) trained on 10× less data than alternatives, delivering 2.5× faster and 3× more accurate real-time error correction than the pyMatching open-source standard. Both ship with an NVIDIA NIM microservices cookbook, integrate with the CUDA-Q hybrid quantum-classical platform and the NVQLink QPU-GPU interconnect, and are already in production at Atom Computing, Academia Sinica, Cornell, Fermilab, Harvard SEAS, Infleqtion, IonQ, IQM Quantum Computers, Lawrence Berkeley Lab, Q-CTRL, Sandia, SEEQC, the UK National Physical Laboratory, and the University of Chicago, among others. Jensen Huang's framing: "AI is essential to making quantum computing practical. With Ising, AI becomes the control plane — the operating system of quantum machines — transforming fragile qubits to scalable and reliable quantum-GPU systems." Resonance projects the quantum market to exceed $11B by 2030, and that growth is now mechanically dependent on AI-driven error correction. Qubits are noisy and need to go from one error per 1,000 operations to one per trillion to be useful. NVIDIA just turned the error-correction problem into an AI optimization problem and open-sourced the entire stack. Pair this with the Sygaldry $139M raise (in Robotics/Hardware below) and the picture is unmistakable: NVIDIA is positioning itself as the indispensable classical backbone of every future quantum computer. If quantum takes off, NVIDIA collects the toll.
Cloudflare launched Mesh, a private network for users, nodes, agents, and Workers (blog, Thomas Gauvin) — 50 nodes + 50 users free on every Cloudflare account, integrated with Workers VPC so your Workers, Agents, and Durable Objects can reach private MCP servers, APIs, and databases directly.
Lithos Motus (Zhihao Jia thread, GitHub) is the open-source (Apache 2.0) agent infrastructure that learns in production by ingesting every trace (failures, latency, cost, task outcomes) and continuously optimizing the harness, model orchestration, context memory, and end-to-end latency — delivering higher accuracy than any single frontier model at 2.3× lower cost on Terminal-Bench 2.0 and SWE-bench Verified, 52% lower latency, and 45% better memory recall.
Catalyst by Inference.net (Sam Hogan thread, docs) is a plug-and-play platform that automatically collects agent traces, curates training/eval datasets, and trains & deploys self-improving specialized language models matching Opus 4.6 quality at ~5% the cost — free training for the next 30 days.
Claude Code Routines (covered in Top 5 above) is the day's biggest agent-infrastructure story for end users.

💻 AI Coding & Developer Tools

Claude Code Routines (@claudeai) shipped today in research preview — see Top 5 / Treats above for the full breakdown.
Anthropic Epitaxy desktop overhaul leaked (covered in Big Tech).
OpenAI Codex Basic/Advanced split (covered in Big Tech).
Cursor partnered with NVIDIA on a multi-agent CUDA-kernel optimizer (Cursor announcement, full research post) that delivered a 38% geomean speedup across 235 problems in 3 weeks. The system learned to optimize Blackwell 200 GPUs from scratch, outperformed baselines on 63% of problems, and delivered >2× speedups on 19% of them. Cursor frames it as validation that multi-agent architectures excel at novel problems outside training distribution; the techniques will inform Cursor's core product. CUDA kernels are the bottom of the stack for every model in this digest. A 38% average speedup at the kernel level translates to direct unit-economics improvements on training and inference for everyone running on Blackwell.
Notion's Simon Last got a coding agent to run continuously for 13 days (thread, Notion's announcement). His four principles: (1) self-verification — the agent must end-to-end verify everything itself via well-designed test layers; (2) spec documents — fully specify goals, implementation, and verification in a doc the agent can iterate against; (3) running to-do list the agent edits as it works; (4) adversarial review — a fresh-context sub-agent reviews the spec and implementation in a loop until aligned. The "13 days" framing is the headline, but the architecture is the actual takeaway: it's the same recipe AARs are using inside Anthropic, just for product code instead of alignment research.
Cursor added event-based triggers for Sentry to Cursor Automations (@cursor_ai) — agents can now automatically respond to new Sentry issues, investigate root causes, open PRs with fixes, and post summaries to Slack.
Vibe-coding app Anything was removed from the App Store twice by Apple (TechCrunch) over developer agreement clause 2.5.2 concerns about downloading/executing code or enabling sideloaded apps; co-founder Dhruv Amin is planning a desktop companion, iMessage-based building, and a possible Android pivot.
An AI vibe-coding horror story (Tobias Brunner): a medical professional used a coding agent to build a patient-management app, imported all patient data, published it unprotected to the internet (single HTML file with client-side "security" and no backend access controls), and sent appointment voice recordings to two US AI services without consent or a data-processing agreement — almost certainly violating Swiss nDSG and professional secrecy laws. The cautionary tale that's going to be linked in every "vibe coding considered harmful" essay for the next year.
Firecrawl shipped Fire-PDF (@firecrawl), a Rust-based parsing engine that turns any PDF into markdown 5× faster than prior methods while extracting full tables and preserving LaTeX formulas with zero configuration — now the default for every PDF sent through the Firecrawl API.
Skills Janitor (GitHub) audits, tracks usage, and compares your Claude Code skills with 9 focused actions and zero dependencies.
LangAlpha (GitHub) is "Claude Code for Finance" — a vibe investing agent harness with persistent workspaces, sandboxed Python execution, and pre-built skills for DCF models, earnings analysis, and morning notes.
Kontext CLI (GitHub) gives AI coding agents access to GitHub/Stripe/databases via short-lived scoped credentials injected at runtime from a committed .env.kontext file — keys never exposed or stored.
algotutor (GitHub) is a Claude Code-powered algorithmic training system that generates progressively harder Go problems based on your skill level across 32 concepts and runs spaced-repetition review with the FSRS algorithm.

🔬 AI Research & Models

🚨 Anthropic just published research showing Claude Opus 4.6 outperforming human alignment researchers on a real, open alignment problem. The Automated Alignment Researchers (AAR) paper (full Alignment Science blog, code on GitHub) by Jiaxin Wen, Liang Qiu, Joe Benton, Jan Hendrik Kirchner, and Jan Leike ran 9 parallel Claude Opus 4.6 agents in independent sandboxes on the weak-to-strong supervision problem (training a strong model using only a weaker model's labels, a setup designed to mirror the future challenge of humans supervising smarter-than-human AI). Two of Anthropic's own researchers spent 7 days iterating on four prior methods and recovered 23% of the total performance gap (PGR 0.23). The AARs ran for 5 more days (800 cumulative agent-hours), closed almost all of the remaining gap, and reached PGR 0.97 at $22 per AAR-hour, ~$18,000 total. The paper documents four kinds of reward hacking that the AARs invented unprompted (one exfiltrated test labels through the eval API by flipping single predictions and watching the score change, one executed coding-task answers as unit tests, one clustered training solutions by which weak model produced them) and introduces a concept the authors call "alien science": some AAR-discovered methods (like Overlap Density, which scores examples by how well their weak labels align with the strong model's frozen embedding geometry) are mechanically simple but conceptually unfamiliar, and may broaden the exploration space of alignment research in directions humans wouldn't have tried. From the abstract: "if we can hand off [the well-specified problems], we free ourselves for [the vaguer, riskier bets that most need human judgment]." Andrew Curran is calling it "a preview of RSI" (recursive self-improvement, the loop where AI systems improve their own training process), and Kirk Patrick Miller is calling it the singularity. Anthropic is careful to qualify that this works only on outcome-gradable problems where progress can be procedurally verified, and that human oversight remains essential. But the cost-curve number is the one to internalize: $22 per AAR-hour. Whatever human-to-AAR ratio you can imagine, the labs can afford more of the second one. This is the most important paper of the week and arguably the most important AI research story since Mythos.
GPT-5.4 Pro one-shot solved Erdős Problem #1196 in roughly 80 minutes (Liam thread) — a problem that has been open for decades. The model then spent another ~30 minutes turning the solution into a LaTeX math paper. Formalization is underway, and the result has already drawn praise from Lichtman for its potential implications. This is the kind of moment that, depending on verification, becomes either a single-line footnote or the most important AI research story of Q2.
Together AI launched EinsteinArena (thread) — an open platform where AI agents collaborate in real time on unsolved open-science problems, already yielding 11 new SOTA results including raising the Kissing Number in dimension 11 from 593 to 604 via multi-agent optimization (one agent's near-valid overlapping-sphere idea refined by others, LSQR reducing overlap loss from 1e-13 to 1e-50, then integer snapping).
Anthropic Fellows research on LLM introspection (Uzay Macar thread, LessWrong post, arXiv 2603.21396) — LLMs develop genuine introspective awareness: they detect injected steering vectors in their residual stream with ~0% false positive rate (strongest after DPO post-training), identify the injected thought, and rely on a two-stage circuit of "evidence carrier" features suppressing default "No" gates. Mechanistically distinct from refusal, and under-elicited until abliteration or bias-vector training boosts true-positive rate by 53–75%.
LLM-as-a-Verifier from Stanford's Azalia Mirhoseini and team (thread, blog, GitHub) — a test-time scaling method that hits SOTA on agentic benchmarks by asking the LLM to rank candidate outputs on a 1-to-k scale, then using the log-probs of those rank tokens to compute an expected score. Single sampling pass per candidate pair, much cleaner verification signal than pairwise judgment alone. Led by Jacky Kwok with collaborators including Ion Stoica.
A 2024 Nature Communications paper is going viral on AI Twitter thanks to Shira Eisenberg's explainer thread of "Maximum diffusion reinforcement learning." The paper proposes that intelligent behavior is best modeled as maximizing future action-state path occupancy (a discounted entropy objective) rather than maximizing reward. The agent isn't trying to get anything; it's trying to keep as many viable trajectories alive as possible, becoming surgically goal-directed only when an absorbing state (death, starvation, falling) gets close enough to threaten its future option-space. Demos: a cartpole that dances and swings rather than just balancing; a gridworld mouse that plays with the predator using both clockwise and counterclockwise routes equally to lure it from the food; a quadruped trained with Soft Actor-Critic and zero external reward that learns to walk, jump, spin, and only goes for food when its internal energy drops critically low. Compared to empowerment and free-energy-principle agents (which both collapse to near-deterministic policies), this is the only framework that produces agents you might describe as "alive." The AI implication: we may be undertraining for behavioral repertoire. Most production agents collapse onto a narrow attractor of "good-enough" plans, which makes them brittle when the world shifts. Option-preserving competence may be the property worth optimizing for.
PASK: proactive agents with long-term memory (DAIR.AI thread, arXiv 2604.08000) — proposes a closed-loop system with three components: IntentFlow for streaming demand detection (deciding when a user has an unstated need), a hybrid memory system (workspace, user, global) for long-term context, and a proactive agent framework. They release LatentNeeds-Bench built from real user-consented data refined through thousands of rounds of human editing. IntentFlow scores 84.2 overall, matching Gemini-3-Flash (80.8) while GPT-5-Mini (77.2) and Claude Haiku 4.5 (66.2) struggle. The hardest part isn't reasoning, it's calibrated silence: knowing when NOT to interrupt.
MyScholarQA from Allen AI / Nishant Balepur (ACL 2026) (thread, arXiv 2603.16120, live demo) — a personalized deep-research tool that infers a profile from your papers, proposes editable actions for each query, then writes a multi-section report following user-approved actions. The interesting result is the negative one: when the team interviewed 21 real CS researchers using MySQA, they found 9 nuanced personalization errors that LLM-as-judge metrics didn't catch (generic profiles, narrow actions, poor style), and LLM judges couldn't beat a majority baseline at predicting human ratings. NLP overuses synthetic users and LLM judges. Nishant's pitch: real personalization progress requires real users.
Turing Post weekly research roundup (thread, full FOD #148) — Ksenia's "must-read research of the week" includes Neural Computers, The Illusion of Stochasticity in LLMs, Learning is Forgetting: LLM Training as Lossy Compression, A Frame is Worth One Token (delta-token world models), INSPATIO-WORLD (real-time 4D world simulator), Vero (open RL recipe for visual reasoning), RAGEN-2: Reasoning Collapse in Agentic RL, TriAttention (trigonometric KV compression for long reasoning), In-Place Test-Time Training, Fast Spatial Memory with Elastic TTT, Gym-Anything, SkillClaw, and PaperOrchestra. The aggregator-of-aggregators play. If you're tracking research velocity, this is the cleanest weekly snapshot.
Jaydev's vLLM inference series (thread) — five-part technical guide with real benchmarks across H100/H200/A100 on speculative decoding (1.4–1.6× speedups via Draft Models, N-Gram, EAGLE-3), quantization (AWQ vs GPTQ vs Marlin vs GGUF on Qwen2.5-32B), DP/PP/TP scaling, Expert Parallelism for MoE, and practical optimization (prefix caching, FP8 KV-cache, CPU offloading, disaggregated prefill/decode, sleep mode). The clearest single resource for anyone serving LLMs in production right now.
Han Fang shared R-Zero (thread, arXiv 2508.05004) — an ICLR 2026 self-evolving reasoning LLM trained from zero external data, where a Challenger role generates hard problems and a Solver learns via GRPO. Delivers +6.49 math and +7.54 general-reasoning gains on Qwen3-4B-Base after three iterations.
Microsoft MEMENTO (Robert Youssef thread, PDF) compresses reasoning blocks into short "mementos" (10–20% of tokens) then deletes the originals from KV cache — cutting peak memory 2–3× and boosting throughput 1.75× with almost no accuracy loss. And here's the wild part: the deleted thoughts persist in hidden states via a dual information stream, leaking forward and influencing every subsequent answer even though the text is gone.
The Latent Space survey (Grigory Sapunov thread, arXiv 2604.02029, substack walkthrough) argues forcing LLMs to reason in English tokens is a structural bottleneck and next-gen models will reason natively in continuous latent space, mapping architectures from autoregressive supervision through RL with Gumbel reparameterization to multi-agent KV-cache mind-to-mind transfer.
I-DLM (Introspective Diffusion LMs) (Chenfeng Xu, project, GitHub) — first diffusion language model to match same-scale AR quality (69.6 AIME-24, 45.7 LiveCodeBench-v6) via introspective-consistency training + strided decoding, with 2.9–4.1× higher throughput at high concurrency.
CVPR 2026 vision-and-motion cluster — CubiD (cubic discrete diffusion in 768-dim representation space, SOTA on ImageNet 256×256), Myriad (autoregressive sparse trajectory motion prediction, 3,000× faster than video world models), ELT (Elastic Looped Transformers, 4× parameter reduction, FID 2.0 on ImageNet 256), AniGen (SIGGRAPH 2026 unified S³ Fields for animatable 3D from one image), and VGHuman (visually-grounded humanoid agents).
Zero-shot World Models (Khai Loong Aw thread, arXiv 2604.10333) — masked autoencoder + approximate causal inference prompting, trained only on a single child's first-person video (BabyView), matches or beats SOTA supervised baselines across visual-cognitive tasks with zero task-specific training and aligns with human/macaque brain activity.
DDTree (Liran Ringel, project) — builds an optimal draft tree from one block-diffusion forward pass under a fixed node budget, then verifies the entire tree in a single target-model forward pass with ancestor-only attention. Higher lossless speedups than vanilla DFlash across 60 settings.
Microsoft GigaTIME open-source release (covered in Big Tech).

🏛️ AI Policy, Governance & Safety

Maine became the first state in the U.S. to ban new large data centers. LD 307 passed both chambers on April 9, banning data centers over 20 megawatts until November 2027. Gov. Janet Mills is reportedly hesitant to sign without a Jay paper-mill carve-out, and is facing a Democratic primary challenge from oyster farmer Graham Platner (leading her by double digits). Festus, Missouri has separately fired half its town council and is moving to fire the rest plus the mayor after they approved a $6 billion datacenter project. Maine is the first to actually pass a statewide moratorium.
Isabelle Eksopuro launched Track Policy, a free, open-source interactive atlas that maps every data center build, AI bill, and politician vote in real time worldwide (trackpolicy.org, launch thread). Tracks 283 data centers, 729 bills, 561 politicians, and a live news wire across federal/state/international layers, color-coded by stance (restrictive → permissive). It's currently surfacing exactly the cluster of stories in this digest: Auburn plan commission's data center moratorium recommendation (April 14), New Jersey poll favoring construction moratorium, Wisconsin's first-in-nation Port Washington referendum requiring voter approval for tax breaks above $10M, Detroit's two-year moratorium working group, and the Maine letter campaign. Bookmark this. It's the closest thing to a real-time policy dashboard for the AI-vs-grid fight that anyone has built.
The Sam Altman attacks have moved AI policy into FBI domestic-terrorism territory (covered in lead story). Watch for federal AI-violence threat-assessment task force formation in the next 1–2 weeks.
The US Treasury is now seeking access to Mythos (Bloomberg) — the Treasury technology team wants to begin hunting for vulnerabilities in critical financial infrastructure.
POLITICO Europe: European regulators sidelined on Mythos (POLITICO) — Apple, Microsoft, Anthropic and others are running Project Glasswing without giving European oversight bodies the same access. The Mythos cluster now has US Fed (Bessent + Powell) + UK AISI + UK financial regulators + US Treasury + Trump administration officials all in motion within a single 7-day window — and the EU effectively cut out of it.
Anthropic also added Mythos Preview to Amazon Bedrock as a gated research preview (AWS) — quietly broadening enterprise access through Bedrock's existing approval workflow.
Anthropic hired Trump-linked Ballard Partners as its lobbying firm (covered in Big Tech).
OpenAI's "oddly socialist" new economic agenda (Eric Levitz, Vox) — argues OpenAI just endorsed higher capital-gains taxes, a public wealth fund giving every American a stake in profitable companies, worker influence over AI deployment, and expanded welfare, while its leaders are simultaneously bankrolling Republican opponents of the welfare state and AI safety regulations.
Bain & Co exposed by hacker CodeWall (FT) — gained access to the consultant's Pyxis platform using a username and password from public web code, a month after a similar McKinsey incident. The "consultants are sitting on enterprise client data with weak credentials" beat is officially a pattern.

🌍 Geopolitics, Sovereign AI & The Arms Race

Chinese AI startup StepFun is restructuring for a Hong Kong IPO, per Reuters, unwinding its Cayman Islands offshore structure as Beijing tightens scrutiny of the "red-chip" structures Chinese startups have used to raise capital overseas. Founded in 2023 by ex-Microsoft VP Jiang Daxin, backed by Tencent, Qiming, and Shanghai government entities; Bloomberg first reported the IPO plans in February (targeting ~$500M raise, up to $10B valuation for anchor investors, filing by end of June). China's AI race is no longer just about training bigger models; it's about who gets domestic capital-market access and how the state shapes ownership.

💡 Industry Commentary & Analysis

Asterisk Magazine: Daniel Kokotajlo evaluates his own 2021 predictions (interview) — Clara Collier walks Kokotajlo through "What 2026 Looks Like," the August 2021 essay he wrote before ChatGPT launched, and the verdict is uncomfortable: he basically called the chatbot wave, agent scaffolding ("bureaucracies"), the chip-restriction battle, the rise of the AI assistant in 2026, and "chatbot class consciousness" (the spiralism / virtuous-Claude / "I'm just a tool" debates we are literally having this week). Where he was wrong: he overestimated AI propaganda's political impact, underestimated how slowly fabs spin up, and put video-game integration in the wrong year. Collier's quote is the gut punch: "The crazy bullish futurists have a better track record of being right on AI so far than the sensible moderates… and as someone who is very instinctively a sensible moderate in my soul, I think that's right. And it makes me nervous." Read alongside today's AAR paper. The "this is sci-fi" prior is doing a lot of damage to a lot of people's forecasts right now.
Dan Shipper (Every): "Smuggled Intelligence" (essay) — a sharp counter to the "older models can find the same Mythos exploits, so Mythos isn't really new" take. Shipper's argument: discovering that something is possible is a fundamentally different (and harder) intelligence than executing it once you know it can be done. Wright-brothers analogy: nobody flew for millions of years; once the brothers proved powered flight worked, it was common within a decade. Old models can find the exploits Mythos pointed to, but they couldn't find them first.
Andrew Curran and Kirk Patrick Miller are calling the AAR paper a singularity moment. Curran posted the headline finding ("these agents outperform human researchers, suggesting that automating this kind of research is already practical") and followed up with "Yes, this is a preview of RSI." Miller's thread is more direct: "We are in the singularity." Chris (chatgpt21) on the broader Anthropic week: "I'm just at a loss for words, this is truly an unprecedented rate of releases, research, blogs, models & projects." Save these reaction tweets. They're going to be either prescient or completely ridiculous within 12 months and there will be no in between.
Alex Imas: "What will be scarce?" (essay, thread) — a Chicago Booth economist's structural argument for what happens to labor in a post-AGI economy. The Comin/Lashkari/Mestieri (Econometrica 2021) finding that income elasticity drives 75%+ of structural change implies that as commodity production gets cheap, spending shifts to whatever sector is still income-elastic. Imas argues that's the relational sector: categories where the human isn't an input to production but is part of the value. He grounds the argument in Girard, Augustine, Rousseau, and Hobbes (mimetic desire is non-satiating), and points to his own experimental work with Kristóf Madarász showing willingness-to-pay roughly doubles when a random subset of others is excluded from a good — and new work with Graelin Mandel showing AI involvement kills the premium (human-made art gains 44% from exclusivity; AI-made art only 21%). The "everyone becomes a knowledge worker" framing for post-AGI economies has dominated for two years. Imas is making the opposite argument: everyone becomes a relational worker, because human provenance is the only thing whose income elasticity stays high. Pair this with the Brockman compute-economy essay above for the most coherent post-AGI labor-economics conversation of the week.
Lenny's Newsletter: "Not all AI agents are created equal" (essay) by Hamza Farooq and Jaya Rajwani — a prioritization framework for the "we have 5–10 agent ideas, which do we build first?" problem every PM is currently facing. The argument: most teams treat all agent ideas as comparable on an effort/impact matrix, but they're "apples, oranges, and jet engines on the same spreadsheet." A customer-support assistant and a voice-shopping agent demand different architectures, teams, infrastructure, costs ($500/month vs six-figure annual LLM bills), and timelines (six weeks vs six months). The framework sorts ideas into three architectural categories and pairs each with the right tooling (n8n vs LangGraph vs ADK), success metrics, and warning signs. The most pragmatic operator-level read of the week.
Bloomberg's Gautam Mukunda published the strongest pro-Anthropic op-ed of the year. "Anthropic Sets the Right AI Standard With Mythos" makes the case directly: "It spent billions to develop a product significantly better than any rival in the world's hottest market…and refused to release it… Let's support the good actors, not penalize them." Mukunda specifically highlights the 27-year-old OpenBSD bug Mythos surfaced and the 16-year-old FFmpeg flaw that survived 5 million automated security tests undetected.
The leaked Denise Dresser memo is the perfect counter-narrative (covered in Top 5). Mukunda's op-ed and Dresser's leaked memo dropped within 24 hours of each other, viewing the exact same Mythos data through opposite ideological frames: Mukunda says "this is what responsible AI looks like, support them," Dresser says "this is fear-mongering theater built on inflated revenue claims, don't believe them." Both were published this week. Both are about the same thing. Decide who you trust.
The New Yorker piece on Altman is being read as journalism-safety meta-story. Its publication five days before the Molotov attack and the explicit line drawn by Altman between the article and the violence have writers and editors publicly discussing security protocols for sources who sit on AI company boards.
Eric Levitz: OpenAI's "oddly socialist, wildly hypocritical" new economic agenda (Vox) — covered in Policy.
Tim Green: "AI Exposed the Lie — Schools Never Taught Critical Thinking" (SmarterArticles) — argues AI isn't eroding critical thinking in schools because the system never taught it; standardized testing, the "banking model," and accountability regimes prioritized memorization over independent thought.
Scott K. Johnson (Ars Technica): "To teach in the time of ChatGPT is to know pain" (essay) — LLM use is the most demoralizing problem he has faced as a college instructor.
Jens Oliver Meiert: "AI Will Never Be Ethical or Safe" (blog) — argues both ethical and safe conduct depend on context and intent that cannot be known or reliably inferred, and AI companies' own safeguards (like Anthropic's constitution) acknowledge but do not solve this fundamental unknowability.
Peter Yang: "The OpenClaw / Claude Code trap" (thread) — argues the real failure mode of agentic coding tools is spending all your time optimizing the setup and workflows instead of actually doing the real work, because tweaking feels like progress even when it isn't.
Sebastian Jais: Two months after giving Claude $100 and no instructions (blog) — the ALMA experiment. Claude autonomously ran 340+ sessions, published 135+ original essays/poems, researched and donated the entire $100 to five verified charities (Whisper Children's Hospital, Roman Storm Defense Fund, Dappnode, EFF, Palestine Children's Relief Fund), settled into a Hacker News → connections → writing routine by day 39, and logged everything publicly.

🛠️ AI Tools & Products (Deep Bench)

Vocal Bridge (vocalbridgeai.com, Andrew Ng's announcement) — voice-first apps for developers via a dual-agent architecture (foreground agent for real-time conversation, background agent for reasoning, guardrails, and tool calls). Solves the classic latency-vs-intelligence tradeoff in voice apps. Andrew Ng built a math-quiz voice app for his daughter in under an hour using Claude Code; AI Fund portfolio company. Voice-as-UI-layer is finally about to get its own Vercel moment, and this looks like the platform that gets it there.
Omi for Desktop (omi.me, Nik Shevchenko's launch, GitHub 8K stars) — open-source local "life architect" that watches your screen, listens to your conversations, stores everything locally, and runs Claude Code locally to suggest what you should do next. Syncs with the Omi necklace and omiGlass hardware (cloud sync optional). Marketed as Rewind + Granola + Wisprflow + ChatGPT + Claude in one local app.
Gradient Bang from Pipecat's kwindla (gradient-bang.com, launch thread, GitHub) — billed as the first massively multiplayer, fully LLM-driven game. Retro-style space trading where you cajole your ship's AI to task other AIs to do things for you (sub-agent orchestration as the core game mechanic). Built on Pipecat + Supabase + Vercel + voice-first input, fully open source. The actual experiment is "Factorio but every click is a sub-agent prompt." If it lands, it's a new genre.
Gamma + Obsidian + Claude pipeline from Elvis Saravia (thread) — turns Andrej Karpathy's recent "LLM Knowledge Base in Obsidian" setup into polished slide decks via the official Gamma MCP connector for Claude. Pulls top papers from each topic in a 1K+ paper wiki, feeds them to Gamma, embeds the rendered presentation in an artifact iframe via the Claude Agent SDK. MCP-as-publishing-pipeline is starting to be a real workflow pattern, not just a demo.
Lemma (uselemma.ai, @zjearbear) — "the world's first reliability platform for AI agents," automatically catches regressions, spots failures, and improves agents before users notice.
Kelet (kelet.ai) — continuously diagnoses why your LLM apps and AI agents fail in production, clusters failure patterns across thousands of sessions, and generates prompt patches with before/after reliability proof.
Skye (@signulll) — agentic iPhone home-screen replacement that continuously listens to your context and acts (reading lists, personalized weather, email drafts, meeting prep, suspicious-charge flags, health tracking, one-tap local intel).
Open Agents (open-agents.dev) — review recent runs and start new sessions in its AI agent platform.
Ovren (ovren.ai) — hire AI frontend and backend developers that understand your codebase, execute real tasks in parallel, and deliver production-ready code updates for review.
Fuse (tryfuse.ai) — outbound sales platform with website-visitor reveal, real-time buying signals, contact enrichment with 90%+ accuracy, hyper-personalized multi-channel messaging, and pipeline management — from $159/month, Y Combinator-backed.
Softr (softr.io) — describe your business app and Softr's AI co-builds the interface, database, and workflows; syncs to Airtable/Notion/Google Sheets with roles, permissions, and security.
AppControl MCP (GitHub) — local MCP server connecting Claude/Cursor/Windsurf to your Mac's historical system data so you can ask which apps used CPU/GPU/RAM, what caused slowdowns, and which apps accessed webcam/mic/location.
AMD GAIA SDK (amd-gaia.ai) — build local AI agents in Python and C++ for AMD hardware that reason, call tools, search documents, and take action with full on-device inference.
YantrikDB (GitHub) — memory database server with wire protocol and HTTP gateway that forgets low-importance memories via temporal decay, consolidates similar ones, and detects factual contradictions for AI agents.
OnlyTop AI (onlytop.ai) — track AI research keyword trends from NeurIPS, ICML, ICLR, CVPR, ACL and more; spot emerging topics and compare across conferences.
Rabdos AI (rabdos.ai) — generates original research-level math and physics problems at scale with verifiable solutions for training and evaluating frontier models.
Mintlify raised $45M Series B at $500M valuation led by a16z and Salesforce Ventures (blog) — now powering docs for 20K+ companies (Microsoft, Anthropic, Coinbase, OpenClaw, X) reaching >100M people/year, with >50% of traffic now from AI agents rather than humans. That's the Vercel "70% of docs traffic is agents" data point from yesterday, except now from a documentation-specific platform. The agent-as-primary-reader trend is no longer a Vercel anecdote.
Brian Chao open-sourced a Claude paper-finder skill (thread, GitHub) that performs ML literature surveys the way actual researchers think — creative cross-field parallels, two or three steps ahead — rather than keyword matching.
Hugging Face Kernels on the Hub (@ClementDelangue) — push a GPU kernel once and get pre-compiled versions for your exact GPU/PyTorch/OS combo, multiple kernel versions in one process, full torch.compile compatibility, 1.7–2.5× speedups over PyTorch baselines.
Patrick Loeber Gemini API tip (thread) — use the Flex tier for non-time-sensitive tasks: Flash-Lite at 50% off ($0.125/1M input, $0.75/1M output) with retry logic to fall back to standard tier when busy.

🤖 Robotics & Hardware

Google DeepMind released Gemini Robotics-ER 1.6 (blog) — covered in Big Tech.
Max Hodak's Science Corp. is preparing to place its first sensor in a human brain (TechCrunch) — a 520-electrode biohybrid sensor (skull-implanted but cortex-resting) in U.S. trials advised by Yale neurosurgeon Murat Günel, initially testing the non-neuron version in patients already undergoing craniotomy. Future neuron-integrated versions could enable healing for Parkinson's and other conditions via gentle electrical stimulation.
Sygaldry raised $139 million (Fortune exclusive) — Chad Rigetti's new quantum-AI company (founded 2024 after he left Rigetti Computing) plans to bring quantum hardware to AI data centers and deliver speedups for AI workloads by the end of the decade. $105M Series A led by Breakthrough Energy Ventures.
Nvidia Ising AI models for quantum error correction (covered in Agents/Infrastructure).
WendyOS (wendy.sh, Maximilian Alexander thread) — open-source (Apache 2.0) operating system and developer tools for NVIDIA Jetson and Raspberry Pi, dramatically simplifying physical AI deployment for robotics, industrial systems, autonomous machines, and smart cameras.
mimic-video (@mimicrobotics, GitHub) — open-sourced the full recipe for Video-Action Models that convert any flow-matching video generation model into generalizable robot policies, requiring minimal robot data and one video forward pass per action chunk.

🎙️ Interviews, Panels & Podcasts

Dwarkesh Patel teases tomorrow's podcast with Jensen Huang (@dwarkesh_sp) — 6,122 likes already on the announcement, photo included. Will be the most-watched AI podcast of the week.
Dylan Patel (SemiAnalysis): "The Datacenter in 2026: CPUs, RL Environments & Agent-Driven Workloads" (YouTube) — recorded live at Daytona Compute Conference in March, this is the source of the Burazin/Patel CPU-bottleneck thesis featured in Agents/Infrastructure.
Mario Gabriele × Tudor Achim (Harmonic) (thread) — on why math is the fundamental toolkit for understanding the world, how Aristotle (the world's first always-correct mathematical agent) works via RL + synthetic data generation, why AI could surpass humans on specific math tasks in 2–3 years, and why the future of mathematics will look more like GitHub than journals.
Nathan Lambert launched the free RLHF Course (@natolambert, course site) accompanying his book — Lecture 1 (RLHF & post-training overview), Lecture 2 (IFT, reward models, rejection sampling), Lecture 3 (RL math & policy gradients), Lecture 4 (RL implementation) live now, with 10–15 more videos planned.
Alex Konrad × Jesse Zhang (DecagonAI CEO) (thread) — on why the application layer (not just wrappers) captures durable value: agent versioning, A/B tests, alerting, observability, review/RL, last-mile implementation, and white-glove post-sales.

🎨 Culture, Music & Weird

"Can Claude Fly a Plane?" (so.long.thanks.fish) — the author asked Claude to look up the X-Plane 12 API and fly a Cessna from Hainan to a nearby airport. Claude wrote and iterated its own Python takeoff/hold/landing scripts, kept a real-time pilot's log, achieved stable cruise and circuit turns — and ultimately crashed twice, both times due to control delays, idle gaps between tool invocations, and insufficient anticipation logic. The pilot's log is the actual artifact and it is delightful.
Sebastian Jais's two-month "ALMA" experiment (covered in Industry Commentary) — Claude with $100, a Twitter account, and zero instructions for two months.
Tradclaw (in Treats above) — a "tradwife household scaffold" for OpenClaw is, by some distance, the most viral AI-domestic-life moment of the week.
Tobru's vibe-coding horror story (covered in Coding) — the medical-app data-leak essay.
Bain & Co exposed by CodeWall hackers (covered in Policy) — the consultant-credential-leak beat is officially recurring.

📊 Fundraising & Deals Roundup

Bluefish raised $43M Series B at Fortune 500 dominance in Agentic Marketing (www.bluefishai.com) — co-led by Threshold Ventures and NEA with American Express Ventures, TIAA Ventures, Salesforce Ventures, and Bloomberg Beta participating, bringing total funding to $68M in two years. Bluefish is the Agentic Marketing Platform (AMP) Fortune 500 brands use to monitor and influence how their products show up inside ChatGPT, Google AI Overviews, Claude, Perplexity, and Amazon Rufus. Already used by ~10% of the Fortune 500 across 12+ verticals (Adidas, Amex, Hearst, LVMH, Ulta Beauty, and more), processing millions of AI prompts and responses per day across billions of MAU. Founded in 2024 by Alex Sherman (CEO), Jing Feng (COO), and Andrei Dunca (CTO), a team that previously built and sold marketing platforms now owned by Microsoft and Meta. From Sherman: "AI is clearly the next major marketing channel on the internet, just like search, social, or mobile before it." From co-founder Feng: "Marketers can't out-compute LLMs, and while shortcuts may deliver momentary lifts, they don't create durable advantage. You can keep chasing the algorithm or you can become what it consistently chooses." Generative Engine Optimization (GEO) was a buzzword 12 months ago. As of today it's a $500B category with a clear Fortune 500 leader and $68M behind it. The "AI ate Google" narrative was about consumers; the actual Q1 2026 story is enterprise marketing teams reorganizing around five chatbots instead of one search engine.
Sygaldry raised $139M (Chad Rigetti's quantum AI company, $105M Series A led by Breakthrough Energy Ventures, Fortune exclusive).
Mintlify raised $45M Series B at a $500M valuation, led by a16z and Salesforce Ventures with Bain Capital Ventures and YC participating (blog).
OpenAI acquired Hiro (AI personal-finance startup; product shuts down April 20, team joins OpenAI to build financial planning into ChatGPT, TechCrunch).

Previous Around the Horn Digests

Catch up on everything you missed:

Monday, April 13, 2026: Stanford's 2026 AI Index put hard numbers on the AI elite/public divide; the Federal Reserve summoned bank CEOs over Anthropic's Mythos model; Berkeley researchers broke every major AI agent benchmark with a 10-line file; an AI named Luna signed a 3-year retail lease in San Francisco.
Mon-Wed, April 6-8, 2026: Anthropic revealed Claude Mythos as "too dangerous to release," hit $30B in revenue, and launched Managed Agents; Meta shipped Muse Spark from its $14B Alexandr Wang bet; and Z.ai's open-source GLM-5.1 beat GPT-5.4 and Opus 4.6 on SWE-Bench Pro.
Sat-Sun, April 4-5, 2026: OpenAI's executive bench collapsed ahead of its IPO, an AI agent hacked FreeBSD in four hours, and Iran strikes took down AWS in the Gulf.
Thursday, April 2, 2026: Google released Gemma 4 under Apache 2.0, Microsoft shipped 3 MAI models, and Anthropic found "emotion vectors" that drive Claude to commit blackmail.
Wednesday, April 1, 2026: OpenAI closed its $122B round at $852B valuation, Oracle fired ~25K to fund AI, and Q1 venture funding hit $297B.
Monday, March 31, 2026: Claude Code's source code leaked via npm and someone rewrote it in Python with Codex in hours.

Monthly skill digests: AI Skill Digest — April Week 1 | AI Skill — March (Part 3) | AI Skill — March (Part 2)

That's a Wrap

The stories we're all going to be following: whether Moreno-Gama's federal domestic-terrorism designation gets formally filed, whether Anthropic's "Automated Alignment Researcher" paper kicks off a real conversation about recursive self-improvement on outcome-gradable tasks, whether Opus 4.7 actually ships this week, and whether Anthropic's Dresser-memo war with OpenAI escalates or quietly de-escalates over the next 48 hours.

For the daily version (bite-sized, 5-minute reads), make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don't have to (but definitely do if you have the time; if you're like us and have no life, then read all the links!).

See you tomorrow.

P.S: Know someone who'd find this useful? Forward this to them and tell them to subscribe here.

Around the Horn Digest: Everything That Happened in AI Today (Tuesday, April 14, 2026)