AI in 15 — April 04, 2026

Anthropic just pulled the plug on third-party tools using Claude subscriptions, and the timing couldn't be more interesting. The same week Google gives away its best models for free and a viral guide shows you how to run them on a Mac mini.

Welcome to AI in 15 for Saturday, April 4, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Saturday, Marcus. We've got a lot to unpack today. Anthropic cuts off OpenClaw and third-party tools from Claude subscriptions effective today. Ed Zitron publishes the most comprehensive bear case against the entire AI industry, calling it a subprime crisis. A research paper finds that AI users are surrendering their cognition to language models at alarming rates. New Rowhammer attacks can compromise machines running Nvidia GPUs. Apple's hidden on-device AI model goes viral thanks to a clever command-line tool. And Meta open-sources AI agents that can automatically formalize mathematics textbooks. Let's get into it.

Anthropic blocks OpenClaw and third-party AI tools from Claude subscriptions, effective noon Pacific today.

Ed Zitron draws explicit parallels between the AI boom and the 2008 financial crisis.

And research shows people follow AI's wrong advice nearly eighty percent of the time without checking.

Marcus, this one hits close to home for a lot of developers. Anthropic is cutting off third-party tools from using Claude subscription allowances starting today. What exactly is changing?

So OpenClaw is an open-source agentic coding tool that was piggybacking on Claude Code's subscription authentication. Users could route heavy agentic workflows through Anthropic's infrastructure at flat subscription rates instead of paying per-token API prices. As of noon Pacific today, that's over. Boris Cherny, who heads Claude Code, said their subscriptions simply weren't built for the usage patterns these third-party tools create.

And this connects directly to the usage problems we've been covering all week.

Completely. Remember those users burning through their Max plan allocation in one to two hours? This is part of the same economic reality. Anthropic made five billion in revenue but spent roughly ten billion on compute. The subscription model works on the assumption that most users don't max out their allocation. Agentic workflows running autonomously for hours at a time completely break that math.

So users are just cut off?

Not entirely. They can still authenticate with their Claude login, but all third-party tool usage now gets billed as extra usage, which is pay-as-you-go. Anthropic is softening the blow with discounted usage bundles at up to thirty percent off and a one-time credit equal to your monthly plan price, redeemable until April 17.

The Hacker News thread on this was enormous. Over four hundred comments.

Heavily divided. Some users genuinely sympathize with Anthropic's capacity constraints. Others are calling it a bait-and-switch. And several commenters pointed out this is subtly different from the earlier OpenCode situation. OpenClaw was technically still using Claude Code as the harness, not impersonating it. So the line Anthropic is drawing is that their subscription covers their own tools only. Everything else goes through the meter.

This feels like the end of an era, Marcus. Unlimited flat-rate AI for heavy workloads.

It absolutely is. And every AI company offering subscriptions faces the same math. Heavy users cost far more to serve than their subscription fees cover. OpenAI, Google, they're all going to hit this wall eventually. The ecosystem of third-party AI coding agents is now on notice. You either pay API rates or find alternatives.

Speaking of alternatives, let's talk about Ed Zitron's piece because it ties directly into what we just discussed. He's calling this the subprime AI crisis. Marcus, is that hyperbole or is there substance here?

Look, Zitron has a clear bearish perspective. But the financial data he presents is real. Anthropic, five billion in revenue, ten billion in compute costs, needing an additional thirty billion raise in February after already raising sixteen and a half billion in 2025. OpenAI, four point three billion in revenue through September 2025 against eight point six seven billion in inference costs alone. CoreWeave went public operating at negative six percent operating margin and negative twenty-nine percent net loss.

And the startup valuations are disconnected from reality?

Harvey is valued at eleven billion on a hundred and ninety million in annual recurring revenue. That's a fifty-eight X revenue multiple. Cursor at twenty-nine billion on roughly one to two billion ARR. And on the infrastructure side, a hundred and seventy-eight billion dollars in debt was raised by data center developers in the US last year, but only five gigawatts of the two hundred gigawatts of announced capacity is actually under construction.

He identifies what he calls "pale horses" signaling the collapse. What are those?

Price increases and service degradation from major AI labs, like exactly what Anthropic did today. Venture capital liquidity crises with no exit mechanisms. Data center construction failures. And the fundamental impossibility of raising prices enough to be profitable without destroying demand. He argues the crisis effectively began in June 2025 when Anthropic and OpenAI introduced priority service tiers.

A counterpoint did come from Project Syndicate this week though.

Mendonca and Bailey argue that even if there's a bubble, it won't trigger a systemic financial crisis because unlike subprime mortgages, AI investments are concentrated among well-capitalized institutions, not retail homeowners. That's probably right about systemic risk. But it doesn't mean individual AI companies and their investors won't get crushed. And to be frank, I'd take any analysis with a grain of salt that draws crisis parallels based on an industry that's genuinely transforming how work gets done. The spending may be ahead of revenue, but the productivity gains from AI coding tools alone are real and measurable.

So the truth is somewhere between "everything is fine" and "everything is 2008."

The truth is that the spending is unsustainable at current revenue levels, and something has to give. Either prices go up, which Anthropic just demonstrated, or capabilities justify the investment. We're going to find out which one wins this year.

Now let's shift to something that should concern everyone who uses AI tools daily. Researchers at Wharton published a paper called "Cognitive Surrender," and the findings are genuinely alarming. Marcus, what did they discover?

They built on Daniel Kahneman's famous System 1 and System 2 framework, fast intuitive thinking versus slow analytical thinking, and proposed a third category. System 3, artificial cognition, where decisions are driven by external AI rather than the human mind. In their key experiment with three hundred and fifty-nine participants, subjects followed AI's correct advice ninety-three percent of the time. Fine. But they followed AI's wrong advice eighty percent of the time.

Eighty percent. They just accepted wrong answers.

And here's the crucial part. They weren't just trusting AI too much. They weren't checking at all. Participants adopted AI-generated answers wholesale, bypassing both their intuitive and analytical reasoning entirely. And people who adopted AI answers reported higher confidence in those answers, even the wrong ones.

So not only do they accept bad answers, they feel more certain about them.

The Hacker News discussion really resonated with professionals. Someone working in a creative field described clients using AI to generate concepts without thinking about them, then being unable to explain or defend their own ideas in meetings. Others pushed back, arguing this is the same as driving into a river because Google Maps said to. A laziness problem, not a new cognitive phenomenon.

But the scale of AI interaction is different from following GPS directions.

Fundamentally different. Professionals are using AI for coding, legal analysis, medical diagnosis, creative work, dozens of times a day. If the error rate becomes invisible because people stop critically evaluating outputs, wrong answers compound silently. Some users in the thread said they've actually stopped using LLMs for coding after noticing their own skills degrading.

Quick security story. New Rowhammer attacks that can give full control of machines running Nvidia GPUs. How worried should we be?

Three new attack variants, dubbed GDDRHammer, GeForge, and GPUBreach, exploit GDDR6 memory in Nvidia Ampere generation GPUs. They force bit flips in GPU memory that can be leveraged to compromise the entire host system. GeForge induced over eleven hundred bit flips on an RTX 3060.

But this is limited to older cards?

Confirmed only on the RTX 3060 and RTX A6000, both Ampere architecture. Nvidia's newer Hopper and Blackwell GPUs already have mitigations. There are defenses available, enabling ECC memory and IOMMU at the system level. And there's no evidence of real-world exploitation yet. The concern is more about what it demonstrates. GPU memory is now a viable attack surface, and in shared cloud environments where multiple tenants share GPU clusters, that opens questions about workload isolation.

Especially relevant as billions pour into GPU infrastructure.

GPU security is becoming infrastructure security. It's early, but the research community has opened a door that won't close.

Now for something delightful. A project called Apfel went absolutely viral on Hacker News, six hundred and sixty points. It unlocks Apple's built-in AI model from the command line. Marcus, what's the appeal?

Every Apple Silicon Mac running macOS 26 ships with a three-billion-parameter language model optimized for the neural engine. Apple built it for Siri and system features, but Apfel exposes it three ways. A Unix CLI with stdin and stdout piping, an OpenAI-compatible HTTP server on localhost, and interactive chat. No API keys, no cloud connection, no subscription, no downloads. It's already on your machine.

The privacy angle drove a lot of the enthusiasm.

Enormous enthusiasm. But there are real limitations. Four thousand ninety-six token context window, about three thousand words. Slower than cloud APIs. And capability-wise, roughly comparable to Qwen 3 4B from a year ago. Some users also flagged security concerns about exposing local models via HTTP, noting that random web pages could potentially send requests to that port.

But combined with Gemma 4 launching under Apache 2.0, which we covered yesterday, and that viral Ollama setup guide hitting nearly three hundred points on Hacker News, local AI is having quite a week.

The convergence is striking. Capable open models, easy deployment tools, powerful consumer hardware. All arriving the same week Anthropic cracks down on subscription usage. Developers are actively exploring alternatives to cloud AI dependency, and the tools to do it are suddenly everywhere.

Last one. Meta open-sourced ReproProver, AI agents that automatically formalize mathematics textbooks into verified proofs. Interesting because?

Multiple specialized agents, sketchers, provers, maintainers, and reviewers, coordinate to read textbook content and convert it into formally verified proofs in the Lean programming language. They share a file-system-based issue tracker and a central knowledge document. It's agentic AI applied to deeply intellectual work beyond just coding. If it scales, automating mathematical formalization could transform how mathematics is taught and verified.

Multi-agent systems doing real intellectual heavy lifting. That's where things are headed.

Saturday big picture. Anthropic charges more, Google gives away more, and developers reach for local alternatives. Marcus, what's the thread this week?

The economic model of cloud AI is fracturing in real time. Anthropic can't afford to subsidize heavy users. Google is weaponizing open-source to commoditize the competition. Apple ships capable models on every Mac for free. And the research says users who do rely on cloud AI aren't even thinking critically about the outputs. The industry is being squeezed from both ends. Costs are rising for providers, and free alternatives are multiplying for users. The companies that survive will be the ones offering something you genuinely cannot run locally. Everything else is heading toward commodity pricing or zero.

And in the middle of all that, humans are forgetting how to think for themselves.

Which might be the most important story we covered today. All the economics in the world don't matter if we're building tools that make us less capable of using them wisely.

That's your AI in 15 for Saturday, April 4, 2026. See you Monday.