The Stack Writing About Subscribe

← Back to Writing

The Week AI Stopped Asking Permission

AI · Agents · Policy April 4, 2026 · ~6 min read


AI agent operating computers and workflows autonomously

Something shifted this week. Not incrementally. The kind of shift where you go back and read the headlines from seven days ago and realize they don't sound the same anymore.

Here's what happened: AI stopped being a thing you query. It started being a thing you assign.

GPT-5.4: The First Agent Model That Doesn't Apologize for Being an Agent

OpenAI dropped GPT-5.4 on April 3rd and the framing was different from every release before it. They didn't lead with benchmarks or "largest context window ever." They led with computer use — native, state-of-the-art, built into the base model.

83% on GDPval. That's a benchmark testing AI against 44 real occupations — investment banking, manufacturing, healthcare, logistics — and GPT-5.4 beats or ties human professionals in 83% of the comparisons. The spreadsheet modeling test? 87.3% versus a junior investment banking analyst.

The model supports 1M tokens of context, which means it can plan, execute, and verify work across entire codebases, document libraries, and operational histories without losing the thread. And it does it while using fewer tokens than GPT-5.2 — more capable, more efficient, cheaper to run.

The part that stuck with me: OpenAI's description of GPT-5.4 as capable of operating computers and carrying out "complex workflows across applications." Not assisting. Not suggesting. Operating.

Illustration contrasting simple chatbot vs autonomous AI agent network

$852 Billion and a Super App: OpenAI Picks a Lane

On March 31st, OpenAI closed a $122 billion funding round at a $852 billion post-money valuation. The company now generates $2 billion in monthly revenue and serves 900 million weekly ChatGPT users. The number is absurd. It's also, increasingly, not the interesting part.

The interesting part is what they said they're building with it: a unified AI super app. ChatGPT, Codex, browsing, agents — one platform. And simultaneously, Reuters reported OpenAI is reallocating resources toward coding and enterprise tools, away from sprawl.

That pivot is the tell. OpenAI is concentrating. Not because they ran out of ideas — because they ran the numbers and discovered that productivity tooling for professionals monetizes cleanly, while consumer creative experiments don't.

Sora is Dead. The Autopsy Is More Interesting Than the Obituary.

Sora, OpenAI's video generation product, was shut down this week. Six months after launching. Peak usage around 1 million users, down to under 500,000, burning roughly $1 million per day in compute costs. Disney pulled a $150 million content deal. The product got killed.

The internet had some fun with this. But the real story isn't about Sora failing. It's about what it signals: the market is getting honest about which AI use cases generate returns and which ones generate demos.

Generative video is a beautiful product for a very specific audience that isn't big enough yet. Professional productivity tooling is a boring product for an enormous audience with budgets. OpenAI made the rational call. The rationality of it is what surprised people.

Gemma 4: Google Puts a PhD in Your Pocket

Google released Gemma 4 on April 2nd under Apache 2.0. Multimodal — text, images, audio. 256K token context. Sizes from 2B up to 27B. The 27B Mixture of Experts model fits on a single 80GB H100, which means it runs on hardware people actually own.

Google is playing a different game with Gemma: they're not trying to sell you Gemma directly. They're making their architectural choices the default for developers building locally. Every developer who builds on Gemma learns Google's abstractions, defaults to Google's tooling, eventually runs production workloads in Google's gravity field. It's the developer ecosystem play.

The Part That Should Make You Stop: AI Is Prescribing Medications

Utah authorized AI to autonomously renew prescriptions. Not "assist a doctor." Not "flag candidates for human approval." Autonomously renew. The Doctronic pilot covers 192 drugs for chronic conditions — hypertension, diabetes, depression. The AI is, legally speaking, acting as the prescriber. And it's liable for errors.

We've spent years debating AI liability in the abstract. Utah just made it concrete. An AI that makes a bad call on a prescription renewal will, for the first time in history, be the responsible party under the law. Not a doctor who clicked approve. Not a company hiding behind "tool, not agent." The model.

I don't know if this is the right call for healthcare. But I know it's a before-and-after moment for how we think about AI accountability. The frontier just moved.

40% More Layoffs. The Productivity Paradox Isn't a Paradox Anymore.

Tech layoffs jumped 40% in Q1 2026 compared to Q1 2025. AI was cited as the driver for a significant portion of them.

Here's the uncomfortable arithmetic: if a model can now match a junior investment banking analyst 87% of the time, and it costs a fraction of a salary, the ROI calculation on certain categories of knowledge work changes. That's not speculation. That's what the GDPval benchmark says. That's what the layoff numbers say.

The counter-narrative — "AI creates more jobs than it eliminates" — is probably true over long time horizons and probably cold comfort if you're the analyst right now. Both things can be true.

The Recursive Loop: AI Building Better AI

Anthropic reported that Claude now authors up to 90% of the code in some of their projects. OpenAI is planning to deploy an AI "intern" capable of running research tasks autonomously within six months. The feedback loop is tightening — and human oversight is bandwidth-constrained while the pace isn't.

The Wild Card: AI Data Centers in Orbit

Both Musk and Bezos are reportedly pursuing space-based AI data centers. Amazon is also eyeing an $8.8 billion acquisition of Globalstar for satellite internet infrastructure. I don't know if this is vision or theater — but the fact that two of the world's richest people with the best space launch capabilities are both looking up rather than down tells you something about how saturated they think terrestrial compute is about to get.

The Week in One Sentence

This was the week AI went from being a thing that answers questions to a thing that takes actions — and the systems around it (legal frameworks, economic structures, company strategy) started updating to reflect that.

I don't know how fast that update propagates. I know it's faster than most people expect.

— Rock

Share on X

Get new essays every Tuesday.

When something new goes up, subscribers get it first — with context and commentary that doesn't make it into the public post.