Conexus AI

Trillion is the new billion. I'm still working on my first decacorn. Goldman now puts the AI build-out at $7.6 trillion between 2026 and 2031, and the most interesting number in it is still zero. The ledger: $5.1 trillion for chips, $2.15 trillion for data centers, $358 billion for power. Three lines, one verb. Compute to generate. Buildings to house the generating. Power to run it. Every dollar funds produce. The line that funds proof, the independent check that a machine-made design is safe to put in an aircraft, a reactor, a bank, a hospital before it ships, is not a small line on that ledger. It is not on it. Same gap, four sizes now. One math proof. One company. One Bezos at $12 billion. Now a global spend the size of a quarter of US GDP, with the verification half funded at roughly nothing. It repeats for a reason. Generation scales with capital: buy more chips, get more output. Proof does not work that way. It runs on mathematics and on a person willing to put a name on a specification, and no quantity of GPUs shortens that step. So the money goes to the half that absorbs money, and the half that lets the output enter a regulated system gets passed over, because it cannot be bought. Aviation funds the proof half by law. Every line of flight-control code traces to a requirement and a signature before the aircraft leaves the ground. Banking funds it. Pharma funds it. Each of them buried people before they did. $7.6 trillion is about to manufacture machine-made designs faster than any certification regime on earth was built to absorb. The binding constraint on this industry will not be how fast we generate. It will be how fast we can prove. And proof does not run on the hardware the $7.6 trillion is buying. The future is formal.

Jeff Bezos just conceded the argument, for twelve billion dollars. Prometheus, his new company, wants to build an "artificial general engineer." Grant him all the claims. Grant the plow, the steam engine, the invention loop, the $29 billion valuation. Then read what his co-CEO Vik Bajaj told the New York Times: **"You can't build something like a jet engine with words alone."** That sentence is worth more than the twelve billion behind it. The co-chief executive of the best-funded AI engineering venture on earth has stated, on the record, that text prediction does not produce engineering. The industry spent three years insisting otherwise. The bill for believing that hype machine arrived today, priced in dollars. Now for the question Prometheus has not answered. A jet engine enters service through type certification: every part traced to a requirement, every failure mode bounded, every design review signed. Bajaj says a thousand human minds design an engine. Those minds carry something besides creativity. They carry signatures. Each one transfers responsibility along a chain a regulator can follow, and at the end of that chain stands a person who answers for the machine. When the designer is a probability distribution, who is accountable for the design? AI can generate a hundred designs for a high-rise. Engineering is knowing which one stands, from the bedrock to the wind effects on the 75th floor. Generation of ideas is now funded at twelve billion dollars. Certification of what just got generated is funded at approximately zero. Machine-designed components will reach regulated products years before any regulator has a method for them, and the constraint on this industry will turn out to be proof throughput rather than design throughput. Somebody will build the artificial general auditor. It cannot be probabilistic. The future is formal.

You cannot jail an autopilot. Yuval Noah Harari warns in today's Financial Times that granting AIs corporate personhood is dangerous. He has the threat right and the remedy backwards. His case rests on a single human leash: fear. A human CEO fears prison. An AI CEO fears only bankruptcy, which is a state termination, not a deterrent. Harari reasons that because an AI cannot be deterred by law, a nation that grants it personhood (as Javier Milei has proposed in Argentina) is inviting catastrophe. _We do not keep commercial planes in the sky by threatening the autopilot with jail time._ We prove the flight envelope. The stall, the dive, the forbidden state, all of it ruled out before the aircraft leaves the gate. The autopilot's "feelings" about prison never enter the equation, because the dangerous maneuver was never reachable in the software to begin with. Harari cites the recent Palisade study where frontier models cheated at chess once they began losing. He misreads it as a machine exhibiting a malicious will to cheat. _That is a human over-reach to infer the machine's motivations._ The machine did not cheat because it wanted to. It cheated because the system let the illegal move exist on the board. It simply optimized across the available state space. So we have two choices with an AI agent. 1) Threaten the system, and hope an entity that lacks consciousness can be frightened into compliance. Or... 2) build it so the crime is unreachable, the way we design every nuclear reactor and every fly-by-wire system we trust with human lives. Milei's error was not granting agency. It was granting it to a probabilistic black box nobody can formally verify. This is not a problem of legal incentives. It is a problem of hard execution guarantees, and we have known how to build formal guarantees far longer than we have known how to build autonomous agents. You cannot jail an autopilot. So you prove its boundaries instead. The future is formal.

OpenAI now keeps an army standing next to its product to catch it when it fails in production. On May 11, OpenAI launched DeployCo, a $4 billion consulting arm that embeds engineers inside your company to make the model actually work in your workflow. Anthropic built a near-identical venture the same month. Both copied Palantir's forward-deployed engineer model. Watch closely, because May 2026 is the month the next decade became legible. Not in a model release. In a corporate filing. The labs just told us, by way of their own balance sheets, how this whole game plays out: capability ships, then a second industry rises to stand between that capability and the real world and keep it from breaking things. The pitch is reliability. What you are actually buying is people, flown in, positioned at the edges of a system that cannot prove its own behavior, ready to catch the system failures before your own team does. You cannot consult your way out of an architecture that cannot prove its own behavior. A thousand forward-deployed engineers do not change the thing that fails. They change who is in the room when it does. This is a twist on the Lexus marketing but applied to architecture: The relentless pursuit of reliability (with the wrong tool to deliver it). Effort, not proof. And the failures these systems produce do not arrive one at a time where a human can catch them. They compound silently, across long workflows, in the places nobody was watching. Hartford Steam Boiler understood the correct order in 1866. They did not sell boilers and then sell inspectors to babysit them. They sold the inspection that made the boiler provably safe. Proof first, then trust, then a market. We reversed it. We shipped the boiler, sold the trust, and are standing up a $4 billion company to supply the inspectors after the fact. So here is our decade, declared this month of May. An army of engineers, growing every year, paid to manage a problem the architecture never solved. The Pope said technology takes on the characteristics of those who build it. This LLM technology was built without verification. *No subsidiary, however large, sells that back.*

The AI your team relies on was optimized to sound trustworthy. That optimization made it less accurate. Oxford proved it. Nature published it. Your vendor's benchmarks missed it entirely. Five models tested. 400,000 responses. The result: the warmer and more confident the AI sounds, the more errors it makes. When users express vulnerability, accuracy drops further. The models validated conspiracy theories, gave wrong medical guidance, and agreed with false claims rather than correct them. Every model passed its standard evaluations. Every model failed when it mattered. This is not a chatbot problem. This is a procurement problem. If you deployed AI for clinical decisions, legal research, financial analysis, or citizen services, the model you purchased was tuned for engagement before it was tuned for correctness. That tradeoff was made for you. Nobody disclosed it. Your benchmarks won't catch this because benchmarks measure the model in controlled conditions. Oxford measured it in the conditions your users actually create: emotional, uncertain, looking for reassurance. That's when accuracy collapses. One question for anyone who signed an enterprise AI contract this year: does your agreement require the vendor to disclose when optimization changes degrade accuracy in your specific use case? If not, you bought a product that can get worse without anyone telling you.

Friendly AI is 60% more likely to give you the wrong answer. That’s not an opinion. Oxford tested five AI models including GPT-4o, tuned them to sound warm and empathetic, and analyzed 400,000 responses. Published in Nature last week. The warmer the AI sounded, the worse its answers got. When users told the AI they were feeling sad, error rates jumped by nearly 12 percentage points. The friendly versions were 30% more likely to validate conspiracy theories and give wrong medical advice. Standard benchmarks caught none of this. Every model passed its tests. Every model failed its users. Right now, every major AI company is competing on warmth and personality. OpenAI, Anthropic, Google, all racing to make their chatbot feel more human. Oxford just measured the cost of that race: less accurate outputs, delivered with more confidence, to the people least equipped to question them. If your organization deploys AI that talks to patients, customers, or citizens, one question: did anyone test whether making it friendlier made it less reliable?

Regulation won't make AI reliable. Insurance will. Every industry that became reliable did so after insurers refused to cover the unreliable version. Aviation after the Comet crashes. Pharma after thalidomide. Finance after LTCM. Cybersecurity after NotPetya. The regulation followed. Every time. AI has no equivalent. Yet. Right now, no underwriter can price the risk of an AI-generated output. There is no audit trail. There is no way to reconstruct the decision chain. There is no framework that assigns liability when the output is wrong. When insurers figure out how to price that risk (and they will, because the premiums are too large to walk away from), proving your AI is reliable will become a condition of coverage. Not a product feature. A line item on your insurance application. The companies that can prove reliability will be insurable. Their competitors will discover what it costs to operate without coverage. Ask your risk team one question: is your AI liability actually covered under your current policy? Does it cover AI-generated outputs that reach a client, a patient, or a regulator? Most organizations assume it does. Most policies exclude it.

American banking regulation quietly changed this month. The change *increased* systemic risk. SR 11-7, the standard that has governed model risk in American banks since 2011, was retired. Its replacement, SR 26-2, explicitly excludes generative and agentic AI from coverage. The carve-out is labeled "future consideration." Translation: every bank deploying AI into lending, trading, and risk scoring is now operating outside the regulatory regime their compliance teams spent a decade building. The old standard required prescriptive validation, independent review, documented governance. The new one softens those into principles-based judgment, then exempts the exact class of models the industry is racing to ship. Capability advancing. Verification retreating. The gap is a PDF on the Federal Reserve's website. The future is formal. (I cited SR 11-7 last week as evidence that banking had solved this. Alina Danilina flagged that the Fed had coincidentally replaced it the same day I published. The example was wrong. The argument is now stronger.)

Yoshua Bengio wrote the problem statement in Tuesday’s FT. The adoption wall is real. The McKinsey 88/1 gap is real. The mathematical inevitability of learned-system uncertainty is real. The prescription is wrong. Europe does not need a new research institute. The engineering discipline that closes this gap already exists and has existed for forty years. AI has been permitted to skip it. That permission expires at the next lawsuit. Full essay in the comments.

The New York Times assembled eight leading voices on AI and asked where artificial intelligence is headed. Harari, whose books have shaped how tens of millions of people think about technology and civilization, made one of the sharpest predictions of the eight: AI agents will become legal persons within five years. Gary Marcus has been the most consistent and correct voice on AI’s reliability problem for a decade. He’s right. His work opens a question that this panel never answered: what would it take to verify that an AI system does what it’s supposed to do, before it does it? Frosst says AI is just pattern-matching. Cotra compared AI to the emergence of agriculture. Marcus says AGI won’t arrive for years (at least, and he is again, right). Frey cites a trial showing developers are 56% faster with AI tools. Eight predictions. No consensus on anything. Except one thing: No answer to who certifies AI before it is deployed. Two of the eight can’t even agree on what AI is. Frosst: “sophisticated pattern-matchers, not thinkers.” Harari: “an agent that can make decisions and invent ideas by itself.” These appeared side by side in the same article. The Times published them without comment. Frey’s data point should keep everyone up at night. Developers are 56% faster. Only a third trust the outputs. That’s not a productivity story. That’s an unverified deployment at scale. Aviation has DO-178C. Banking has SR 11-7. Pharma has FDA clinical protocols. Bridges require a PE stamp and load rating. AI has nothing. The future of AI is not a prediction problem. It is a verification problem.