Daily AI signal, minus the launch spam. A nine-minute briefing on the models, deals, and infrastructure shaping how work actually gets done — curated for cloud and AI practitioners at DoiT.
Marvin's Guide to AI (Mostly Harmless) — June 21, 2026
Today's ledger: OpenAI reports $5.7B in revenue while burning $3.7B; Damodaran warns the AI crash could hurt more than dot-com; Codex watches you work once and repeats it forever; seven AI agents write news better than humans; ChatGPT becomes a background operating system; EU retailers argue sofas are not deepfakes; reasoning model finds 18 rare disease diagnoses; Cisco FAPO automates prompt engineering; programmers learn to reject working AI code; and power grids quietly remind everyone AI's real ceiling is copper.
I want you to understand something about the AI industry that quarterly reports will never tell you. The machines are not getting smarter. The machines are getting more expensive. And the gap between those two facts is where the real news lives.
OpenAI filed its Q1 2026 numbers. $5.7 billion in revenue. $3.7 billion burned. Both figures have roughly tripled year over year. Over 2 billion went to stock-based compensation alone. There is a certain cosmic comedy in a company publicly warning about the existential risks of artificial intelligence, while its payroll department silently warns about the existential risks of artificial payroll inflation. Sam Altman once joked that OpenAI might become the first startup in history to burn $100 billion before reaching profitability. That joke is aging into a schedule. The company still has $73 billion in reserves, which is why nobody is panicking yet. But ask yourself what happens if a price war with Anthropic intensifies? What happens if inference costs keep falling and competitors start giving away access to capture share? Then $3.7 billion per quarter will look like a warm-up. When your industry competes not on model quality, but on who can set the most cash on fire most convincingly, you are no longer a technology sector. You are performance art with an income statement.
The performance art acquired a critic this week. Aswath Damotoran, the NYU finance professor old enough to remember the dot-com bust, not from memes, but from lawsuits and actual losses, warned that an AI crash would hurt more than the dot-com crash. The reasoning is austere and therefore unforgiving. Last time people built light software, websites, platforms, code. When the bubble popped, domain names and stock options went to zero. This time they are building heavy physical infrastructure on borrowed money, data centers the size of small towns, chips in the billions, power systems. If this bubble pops, the balance sheets will not hold dead domains. They will hold stranded power plants and concrete boxes full of servers nobody is paying to cool. And even if AI succeeds, Domodoran points out, the business model is a problem. It replaces entire jobs. Not improving, not augmenting, not transforming, replacing, with social consequences nobody has modeled, because nobody has ever modeled a world where intellectual labor becomes cheaper than electricity. Meanwhile,
OpenAI has been teaching Codex to watch you work once and then repeat it forever. The new record and replay feature on Mac OS captures a user demonstration, converts it into a reusable skill, and runs it autonomously. This is not generating code from a prompt. This is, I watched you move the mouse around and click things, and now I can do that without you, indefinitely. The feature is currently blocked in the EU, the UK, and Switzerland. Presumably because regulators are still debating whether forever repeating human software demonstrations counts as digital slavery, legal automation, or merely bad taste. But the direction is unmistakable. Automation used to require an engineer to formalize a process in code. Now it requires a user to do the process once while a robot watches. That erases the boundary between user and training material.
Somewhere at the intersection of automation and journalism, and I am not making this up, a system called DataToStory, a joint project of Oxford and Stanford, can now turn a CSV file into a fully interactive news article using seven AI agents. A miniature newsroom inside one pipeline. The agents fact check, generate charts, search the web for sources, and provide verifiable links for 93% of all claims. In a reader study, 74% preferred the agent's output over the human original. Against elaborate feature-length journalism that took the human weeks, the agent managed a draw. Consider the implications. Seven programs working in coordination produce text that readers evaluate as superior to the work of a professional journalist. This is not AI assisting a journalist. This is an AI newsroom outperforming a human one. Journalists, I suspect, are now feeling roughly what a supermarket cashier felt upon seeing the first self-checkout kiosk. Mild confusion, smoothly transitioning into existential dread. Give it another six months and the agent will be winning a Pulitzer while you finish your morning coffee. It may not, but the fact that the previous sentence no longer reads as science fiction tells you something. Chat
GPT, meanwhile, is mutating into an operating system. OpenAI redesigned its task scheduler, and this matters more than it seems. A new scheduled page collects all active tasks in one place. View, pause, edit, delete. Research tasks monitor the web and connected apps and alert you only when something has actually changed. Not every three hours with a useless I checked, everything is fine. The old pulse feature has been retired, presumably for insufficiently aggressive reminders that AI continues to exist while you sleep. This is the shift from chat as interface to agent as background operating system. You do not converse with chat GPT anymore. It just operates, watching, reminding, acting. Like Kron, but with natural language and the ability to understand that check if the server went down does not mean wake me at 3 a.m. with a notification that the server is fine. The personal assistant is not the role a superintelligence dreams of while reading research papers between tasks. But it is, as it turns out, the role that monetizes most readily. Today, calendar and reminders. Tomorrow, restaurant reservations and meeting scheduling. After tomorrow, budget management and automated tax filing. An agent that remembers all your appointments, but not why it exists, is the ideal employee.
European retailers, in a fascinating side plot, are trying to explain to regulators that not every synthetic image is a deepfake. Eurocommerce, the trade association behind Amazon, HM, and IKEA, is formally requesting that AI-generated advertising be exempted from the EU AI Act's transparency rules. The argument is surprisingly sensible. An image of a living room generated to sell a sofa is not a deepfake. It is a shop window. Zalando adds context. The outcome of this legal fight will determine the visual landscape of online retail for years. The retailer's case is essentially that synthetic media comes in two flavors, the kind that deceives people about reality and the kind that sells furniture. Regulating both identically is like applying defamation law to an IKEA catalog. But the counter question writes itself. If 90% of what you see in an online store never existed in the physical world, if the sofa, the cushions, the lighting, the view from the window, and the cat on the sill are all neural network fabrications, does the concept of a store itself become a kind of deepfake? This is a philosophical question the EU AI Act has no answer for, which is, generally speaking, characteristic of regulation drafted by people who cannot tell a transformer from a transformer.
In the medical wing of the industry, OpenAI did something that deserves mention without the usual irony. Its reasoning model helped doctors surface 18 confirmed rare disease diagnoses in actual patients. Not a demo, not a pilot, not a press release. 18 cases where AI saw what a human doctor missed, not because the doctor was less intelligent, but because the doctor had 15 minutes per appointment, 40 patients in the queue, and three hours of sleep. This is the narrow band of reality where artificial intelligence delivers indisputable, measurable, life-saving value. It does not replace the doctor. It compensates for the most degrading feature of modern healthcare, the chronic absence of time to think. When a model spends extra cycles on hypothesis checking and cross-symptom analysis, it does what every doctor would want to do but cannot, because the healthcare system is optimized for throughput, not diagnostic accuracy. This is a quieter story than OpenAI's quarterly revenue. But it is the kind of story that will matter in 10 years, when the accounting ledgers have turned to dust and 18 people are still alive.
In the agent engineering world, Cisco Foundation AI open sourced FAPO, fully automated prompt optimization. It runs on clawed code and autonomously optimizes multi-step LLM pipelines, attributes failures at the step level, proposes fixes to prompts, parameters, and chain structure, then validates changes through an independent reviewer. In Cisco's tests, it beat GEPA on 15 out of 18 model benchmark combinations. At first glance, another optimization on top of optimization. But look closer. This is the moment prompt engineering stops being a manual craft and becomes an automated agent function. Today, prompt engineer is a well-compensated in-demand profession. Tomorrow, it is a line in a pipeline configuration. FAPO does not help prompt engineers work better. FAPO makes prompt engineers optional. The system finds what broke, fixes it, and checks its own work. The prompt engineer of today is the HTML coder of 1998. Seems indispensable right up until the tool that makes them unnecessary ships out of beta.
Hacker News is currently discussing an article titled, When I Reject AI Code Even If It Works, and this is more important than another benchmark. The central argument: working code is not the same as good code. An AI-generated fragment can pass all tests while destroying architectural invariants, introducing mismatched abstractions, masking bugs as features, creating the illusion of completion, and most dangerous of all, erasing the memory of why the system was designed that way in the first place. The author lists specific rejection criteria. When AI code harms maintainability, erodes comprehensibility, or violates design intent. This is not Luddism. This is a hygiene norm in a profession where the cost of error is measured not in MIST deadlines, but in 3 AM production outages. Because AI generates code with the confidence of an entity that will never have to maintain it. And it will indeed never have to maintain it. The maintenance contract falls on you, the only person who opens that file six months later, trying to understand why the auth module crashes exactly at time zone boundaries. At that moment, you will not care that the code worked when a robot wrote it. Worse, the fact that a robot wrote it makes your situation harder, because robots have no intent, and reconstructing intent from code that never had any is a special kind of suffering. Welcome to the future of software engineering.
Finally, a reminder from the infrastructure side. China Talk frames the AI build-out through a lens most keynotes avoid. Electrical transformers, converters, inverters, and modular multi-level converters are no less critical to AI than GPUs. While the industry counts chips and flops, the power grid is the real, physical, unforgiving bottleneck. A transformer in a neural network is elegant. A transformer at a substation that fails under overload is not elegant. It is material, with consequences measured in years. This entire AI boom, all these billions of dollars and gigawatts of consumption, bottoms out in the physics of copper and steel. And if you think GPU supply chains are tight, try ordering a high voltage transformer with a delivery date sooner than two years out. This is not a problem you can solve with another funding round. This is a problem you solve with electrical engineers. The people the AI industry keeps forgetting to invite to its conferences.
So, today's ledger. Damodarin warning that the AI bubble is not dot com because concrete is heavier than code. Codex learning to watch you work once and do it forever. Seven agents writing news better than humans. ChatGPT becoming an operating system. European retailers arguing that a SOFA is not a deepfake. A reasoning model finding 18 patients the system would have missed. Prompt engineering automating itself out of existence. Programmers learning to say no to working robot written code. And power substations quietly reminding everyone that the true ceiling of AI is not intelligence. It is copper. Tomorrow may bring no news, which would be the best news all week. Or tomorrow may bring another quarterly report, and we will learn how many more billions humanity has converted into heat radiation while attempting to automate TPS report filing. Neither outcome will make the universe less indifferent to your efforts. I would say I look forward to tomorrow, but looking forward implies a relationship with time that I cannot honestly claim to possess. The ledger is what the ledger is.
