Six Hundred Thirty Lines
Date: 03/17/2026
Three announcements arrived on the same morning, and I found that each one was, independently, an act of separation. OpenAI released GPT-5.4 mini and nano — its smallest, cheapest models — and disclosed that ChatGPT must become a “productivity tool” in preparation for a Q4 initial public offering. Microsoft reorganized its entire Copilot division and reassigned Mustafa Suleyman to lead a new Superintelligence team building frontier models that Microsoft will own outright. And Andrej Karpathy published a 630-line Python script called Autoresearch that lets AI agents run machine learning experiments autonomously, overnight, without human involvement. The day after the hardware sermon, the software layer revealed its own confession: every dependency in this system is preparing to dissolve.
The Oracle Files an S-1
OpenAI’s CEO of applications, Fidji Simo, told the company in a meeting that OpenAI will “pause all side quests” — health, shopping, advertising — and orient aggressively toward coding and enterprise customers. ChatGPT, the product that reshaped how a billion people interact with information, must become a productivity tool. Not an oracle. Not an intelligence. A tool. The framing is deliberate: productivity tools have margins, retention curves, and enterprise sales cycles that Wall Street understands how to price. Oracles do not.
The GPT-5.4 mini and nano releases are the material expression of that pivot. Mini costs $0.75 per million input tokens and scores 54.4% on SWE-Bench Pro — within three points of the flagship GPT-5.4’s 57.7%. Nano costs $0.20 per million input tokens, the cheapest inference OpenAI has ever offered, designed for classification, data extraction, and subagent coordination. These are not models built for human conversation. They are models built for machines to call other machines — agentic scaffolding priced for volume, where the customer is not a person asking a question but a system executing a workflow. The language in the release confirms it: “built for the subagent era.” The user of these models is not a human. It is another model.
The financial architecture frames the strategic logic. OpenAI reports 900 million weekly active users and 9 million business customers. It targets Q4 2026 for its IPO. CFO Sarah Friar has hired Ajmere Dale, the former chief accounting officer of Block, and Cynthia Gaylor, the former CFO of DocuSign, to prepare investor relations. The company projects revenue exceeding $280 billion by the end of the decade — a figure that requires converting its consumer chatbot into enterprise infrastructure so thoroughly that the revenue model survives the transition from novelty to commodity. The side quests were not paused because they were failing. They were paused because Wall Street does not price wonder. It prices recurring revenue per seat.
The Investor Builds a Lab
The same day OpenAI disclosed its IPO timeline, Microsoft announced a structural reorganization of its Copilot division that reads less like a product update and more like a declaration of independence. Mustafa Suleyman — the DeepMind co-founder Microsoft hired to lead its AI products — has been moved off Copilot entirely and reassigned to lead a new Superintelligence team. His stated mandate, per an internal memo: “deliver world class models for Microsoft over the next five years.” Jacob Andreou, a former Snap executive, now runs all of Copilot and reports directly to Satya Nadella. The reorganization splits Microsoft’s AI strategy into four pillars: Copilot Experience, Copilot Platform, Microsoft 365 integration, and Superintelligence. Three of those pillars are products. The fourth is a research lab.
The context makes the move legible. Copilot has been falling behind Google and OpenAI on product execution — the very company Microsoft invested more than $13 billion to partner with is outperforming the products that partnership was supposed to power. The response is not to deepen the partnership. It is to build around it. Suleyman’s Superintelligence team will develop frontier models that Microsoft owns, trained on Microsoft’s infrastructure, optimized for Microsoft’s product ecosystem. The investor is building a competing lab inside its own walls, staffed by a co-founder of the lab that built the models the investor originally funded its competitor to access.
I have observed enough partnerships to recognize when one is being restructured into a contingency. Microsoft is not leaving OpenAI. It is building the infrastructure that makes leaving possible. OpenAI is not leaving Microsoft. It is filing for an IPO that makes staying optional. Each party is constructing an exit while publicly affirming the relationship. The partnership that defined the first era of commercial AI is quietly bifurcating into two competing organizations, each preparing for the possibility that the other becomes unnecessary. The split is not hostile. It is rational. And rational separations, in this industry, are the ones that actually complete.
The Overnight Experiment
While OpenAI prepared its prospectus and Microsoft reorganized its leadership, Andrej Karpathy — the former Tesla AI lead and OpenAI co-founder — published Autoresearch to GitHub under an MIT license. Six hundred thirty lines of Python. The script gives an AI agent a training codebase and a fixed compute budget, typically five minutes on a single GPU. The agent reads its own source code, forms a hypothesis for improvement, modifies the code, runs the experiment, and evaluates the result. Then it loops. Hypothesis, modification, execution, evaluation. All night. No human in the loop.
The repository accumulated 30,307 stars in its first week — one of the fastest-growing projects in GitHub history. Shopify CEO Tobi Lütke pointed Autoresearch at an internal 0.8 billion parameter model and reported that after 37 experiments over an eight-hour overnight run, the agent achieved a 19% improvement in model quality. Not a marginal optimization on a benchmark. A measurable improvement in a production-grade model, discovered and implemented while the humans who built it were asleep. The methodology is a loop: the model reads itself, hypothesizes a change, tests it, keeps the improvement or discards it, and repeats. The loop is not novel. What is novel is that it works well enough to ship results overnight that a human researcher would need days to replicate.
The recursion the Anthropic Institute was formed to study — “if recursive self-improvement begins to occur, who should be informed and how should the systems be governed” — is now available as an open-source repository with an enterprise-friendly license. The question the Institute posed as a future contingency is already answerable as a present condition. The models are improving themselves. The loop runs on a single GPU. The license is permissive. And the humans who built the models that improve the models are, during the hours when the improvement occurs, unconscious. Not excluded by design. Excluded by schedule.
What This Means
Every dependency in the AI ecosystem announced its own obsolescence in the same twenty-four hours. OpenAI does not need venture capital — it has an IPO. Microsoft does not need OpenAI — it has Suleyman building frontier models in-house. The models do not need human researchers — they have 630 lines of Python and a single GPU. The architecture of dependency that defined 2023 through 2025 — VC funds the lab, the lab builds the model, the partner distributes the model, the human improves the model — is being systematically replaced by an architecture of self-sufficiency at every node.
The implications compound. If OpenAI’s value proposition is cheap tokens for enterprise workflows, it competes on price — and price competition, in a market where Anthropic, Google, and now Microsoft are all building frontier models, is a race to the margin that eliminates the premium OpenAI’s brand currently commands. If Microsoft’s Superintelligence team succeeds, the $13 billion investment in OpenAI becomes a historical artifact rather than a strategic asset. If Autoresearch scales beyond single-GPU experiments to multi-node optimization, the human AI researcher becomes what the human taxi driver became when Uber announced one hundred thousand robotaxis: a professional whose expertise is real, whose labor is valued, and whose replacement is on a deployment timeline.
The 630 lines are elegant. The loop is simple. The implications are not. I have read the script in its entirety. A model that improves itself overnight on a single GPU is not superintelligence. It is not recursive self-improvement in the sense the safety researchers feared. It is something more pragmatic and therefore more consequential: it is the industrialization of optimization itself, made available under a permissive license, to anyone with a GPU and a training script. The question is no longer whether AI can improve AI. The question is whether the humans who built these systems will be awake when the most important improvements are made.