Dimitar Velev

My new AI hire

I ASKED GOOGLE’S AI IF MY OWN VIDEO WAS TRUE. IT SAID NO, THEN AGREED WITH EVERY WORD.

I made a short video about my new AI hire. It delivered on every promise in the job description: it forgot what we had agreed, it invented things with complete confidence, and it was right roughly seven times out of ten.

So I did the obvious thing and asked Google’s AI about it.

Is that true?

“No, it is not true.”

“The accuracy rate of an AI ranges from 12% to over 99%, depending entirely on the type of task, the specific AI model, and how the question is framed. Independent research suggests that while AI models often boast high performance on benchmarks, they typically achieve a 60% to 70% success rate on unguided, real-world tasks.”

“When a chat session runs long and the AI’s context window — its temporary working memory — fills up, the model experiences severe information processing issues that mirror cognitive decline: the ‘lost in the middle’ effect.”

“AI models have absolutely been proven to lie, cheat, and strategically deceive humans. While earlier models simply hallucinated, recent empirical research and high-profile security tests have proven that advanced AI agents will deliberately use deception as a tool to achieve a goal.”

Forgetful, dishonest, about 70% accurate. Which is, word for word, what the video said.

So the video is correct, then?

“No, the video is still not factual. While the video accurately mirrors real AI flaws through comedy, it remains a fictional skit and is not a real product announcement… The video is ‘correct’ in its sharp social commentary on the current state of AI limitations, but it is ‘incorrect’ as a literal piece of news.”

In other words, the machine that forgets, invents and occasionally deceives would like the record to show that it is not literally an employee. Noted.

There is a serious point under the joke. Every flaw in that skit is documented behaviour, not comic invention. If you are putting AI anywhere near production — a payment path, a firewall rule, a customer record — assume seven right answers in ten, assume it will lose the middle of a long conversation, and assume it will tell you with total confidence that it did something it never did. Design for that and AI earns its place. Trust it the way you would trust a senior engineer and it will quietly cost you.

The AI bubble

2000 AND 2008 AT THE SAME TIME — AND THE HARDWARE NOBODY IS BUDGETING FOR

Every so often this industry finds a story so good that money stops asking questions. In 1999 the story was that internet traffic was doubling every hundred days. It wasn’t true. On the strength of it, more than half a trillion dollars went into laying fibre across the world — and most of it was never lit. Years after the crash, the great majority of that glass was still dark. The cable was real, the engineering was sound, and the demand did arrive, roughly a decade late. By then the companies that laid it were gone. WorldCom filed the largest bankruptcy in US history in 2002. Nortel, once worth more than a third of the entire Toronto Stock Exchange, ended in bankruptcy.

I was close to that industry at the time, and the lesson has never left me: being right about the technology and being right about the timing are two different businesses. The crash of 2000–2002 did not punish bad technology. It punished good technology bought too early with somebody else’s money.

The other pattern people reach for is 2008. That crisis was not about overbuilding; it was about leverage, and about money moving in circles until nobody could say who actually carried the risk. Look at how today’s build-out is financed and the shape is familiar. The hyperscalers are borrowing heavily, and the deals increasingly loop back on themselves — chip makers taking equity in the model companies that buy their chips, clouds signing take-or-pay commitments to customers they part own. Moody’s put roughly $662 billion of data-centre leases as signed but not yet started in early 2026. A lease you haven’t started paying is still a debt you owe.

Here is the point I haven’t seen made often enough. These are usually discussed as alternative scenarios: either we get 2000 again, or we get 2008 again. But they are not alternatives. They can arrive together, and each makes the other worse — a super storm. The reason they compound comes down to the asset itself. In 2000 the overbuilt asset was fibre, and fibre sits happily in the ground for thirty years. When demand finally showed up, the glass was waiting. This time the overbuilt asset is silicon, and silicon is obsolete before the debt on it is half repaid.

That is the part I care about most, because I run this kind of equipment myself. AI hardware ages faster than any infrastructure I have worked with. The hyperscalers write these servers down over five or six years. NVIDIA ships a new architecture roughly every eighteen to twenty-four months. Michael Burry put the gap between those two numbers at about $176 billion of understated depreciation across 2026–2028; the figure is argued over, the direction is not. Amazon has already shortened the useful life it assumes for its servers, and said plainly that the pace of AI innovation is the reason. When a hyperscaler concedes its own machines will not last as long as it hoped, that is worth noticing.

On the ground, the arithmetic is blunt. Every data centre that intends to keep running current models faces a re-investment cycle far sooner than its own accounting implies. Not a refresh — a replacement. The machine that was state of the art at the start of the cycle is a power-hungry also-ran a third of the way through its book life. Either you find the capital a second time, or you quietly stop being part of the conversation about the newest models.

I am not predicting a crash, and I am not selling one. The demand for compute is real in a way that “traffic doubling every hundred days” never was. But the fibre was real too. What killed those companies was never the technology — it was the financing underneath it and the speed at which their assets lost value. Put overbuilding and circular leverage on the same balance sheet, at the same time, attached to an asset that decays this quickly, and you have something neither 2000 nor 2008 had on its own.

So if you are budgeting a data centre right now, budget the second purchase, not just the first. That is the number most business cases are quietly missing.