← Writing
·9 min read

Ed Zitron Was Wrong About AI. The Argument Keeps Moving.

Ed Zitron is right about AI's ugly economics and wrong about its adoption, efficiency, and technical progress. Here are the claims and the evidence.

AI4 in Las Vegas had the moral atmosphere of a casino at four in the morning. Bad carpet glowed under artificial light while rooms full of people insisted the next machine would pay out. That was where I ran into Ed Zitron.

The man who makes his living carving strangers apart in public wanted out the second I introduced myself. Fine. The hallway doesn't matter. The written record is worse.

I used to agree with Zitron about most of AI.

The labs were spending obscene amounts of money. Adoption looked coerced. Every mediocre demo arrived wrapped in a press release about the end of human work. Earlier this year, I had very little good to say about any of it.

I wanted Zitron to be right because contempt was simpler than uncertainty. His story gave me villains and let me stop thinking.

Then the models kept getting cheaper. The benchmarks kept moving. People I knew started using them for ordinary work. The numbers refused to cooperate.

I didn't enjoy this. I had spent months making the other argument.

Zitron, host of the Better Offline podcast and publisher of Where's Your Ed At?, remains one of the best critics of AI's economics. He has also made categorical predictions about AI adoption, efficiency, and technical progress. Several failed cleanly and publicly.

On efficiency, he even admitted one mistake. Then he folded that mistake into the same broader conclusion.

That's the problem. Zitron's conclusion is built to survive whatever evidence arrives.

The money

Frontier labs are burning capital. Their accounting is opaque. Hyperscalers finance model companies that turn around and buy infrastructure from them. Some employers force tools on workers without publicly demonstrating a return. Hallucinations remain a serious limitation. The frontier labs haven't publicly shown that the economics work at full cost.

Independent evidence supports the tension Zitron emphasizes. Stanford reports historically fast AI revenue and adoption growth alongside record spending on compute and infrastructure. The International Energy Agency finds rapidly improving efficiency per task alongside rising total electricity demand. Better technology doesn't automatically produce adequate returns.

Zitron's reporting on AI economics and infrastructure costs is worth reading because these questions are unresolved. I made a similar case when I wrote about the circular financing underneath the AI boom.

Then he turns good reporting into bullshit certainty. A real problem becomes proof that the technology can't work or that nobody wants it.

My dissertation committee would have skinned me alive for making those jumps. That was the job: take the exciting claim I wanted to make and ask whether the boring evidence could carry it.

Gemini

In December 2024, The Wall Street Journal reported, citing people familiar with the matter, that Sundar Pichai wanted the Gemini chatbot used by 500 million people before the end of 2025.

Zitron called the target "so unrealistic" that someone at Google should have been fired. He named Pichai. This was refreshingly specific. There was a number, a deadline, and a proposed punishment.

Alphabet later reported that the Gemini app had more than 650 million monthly active users in its third quarter. In its fourth-quarter call, the company said the number had reached 750 million.

Monthly active users don't tell us how deeply people use Gemini, whether they arrived voluntarily, or whether the product will justify Google's spending. Zitron hadn't predicted any of those things. He said the user target was absurd.

Alphabet passed it by 250 million. He was wrong.

The efficiency prediction produced something rarer: an admission.

After DeepSeek, Zitron wrote that he had "assumed, incorrectly" that there was no way to make models more efficient.

He was referring to his September 2024 claim that large language models had "effectively plateaued" and that nobody had succeeded in making them more efficient.

Stanford's 2025 AI Index later quantified what was happening. The advertised price of querying a model at roughly GPT-3.5-level MMLU performance fell from $20 per million tokens in November 2022 to seven cents by October 2024. That's a reduction of more than 280 times. The same report found machine-learning hardware price performance improving by about 30 percent per year and energy efficiency improving by about 40 percent per year.

That October endpoint came shortly after Zitron's essay, and API price isn't the same thing as a provider's internal cost. A company can subsidize usage. But smaller models reaching the same capability, better hardware price performance, and rising energy efficiency all point the same way: useful capability per dollar and per unit of compute was improving quickly.

Zitron's admission lasted one sentence. He immediately said his real mistake had been giving American AI companies too much credit. He managed to turn being wrong into another indictment of the same defendants.

AI did not peak in 2024

Zitron wrote in December 2024 that he had been warning since March that generative AI had "already peaked". He called transformer architecture a dead end and described generative AI products as trapped in place.

He was right about one part. Making pretrained models larger was producing diminishing returns. The naive recipe of more data, more compute, and more parameters was getting brutally expensive.

Scaling stalled. Research didn't. Reinforcement learning improved, models started spending compute at inference time, and tool use became normal. Smaller models got much cheaper.

Stanford's 2026 AI Index reports that performance on the original OSWorld computer-use benchmark rose from roughly 12 percent to 66.3 percent. Scientific, mathematical, and multimodal evaluations improved too.

Benchmarks need warning labels. OpenAI stopped reporting SWE-bench Verified after an audit found flawed tests and evidence of contamination. A newer and much harder computer-use benchmark, OSWorld 2.0, tests workflows that take humans a median of about 1.6 hours. Its best tested system completed only 20.6 percent of tasks.

The capability curve looks drunk. A model can operate a desktop, then fail to read an analog clock. That isn't solved intelligence. It is still measurable progress.

I experienced this while changing the address on my Arizona driver's license. I opened the MVD site in an agent's browser, logged in, and told it to change my address. It navigated the government's ancient portal, paid the fee with my approval, opened a support chat, and clarified an issue with the representative. The new license is in the mail.

That's an anecdote, not a benchmark. I also wouldn't have trusted the same workflow six months earlier.

I saw the same split when Kimi K3 built a working game and still needed me to play it. The models can do something astonishing, then fail a task that would embarrass a seven-year-old.

One scaling recipe had stalled. The field had not.

In April 2026, Zitron argued that AI "isn't actually doing very much" and called agents chatbots plugged into APIs. That description is mechanically accurate and practically useless. An agent is usually a model in a loop with tools and permission to touch something outside the chatbox. What matters is whether the loop gets anything done.

Not reliably. A randomized METR study found that early-2025 AI tools made 16 experienced open-source developers 19 percent slower across 246 issues. Anyone who has used an agent for serious work has watched it confidently walk into a wall, apologize, and then walk into the same wall at a slightly different angle.

That is a serious limitation. It is not "nothing."

John

My friend John goes to my church. Shout out, John, if you're reading this. John isn't technical. He's never written a line of code and he wouldn't describe himself as a PC power user. Sorry, John. I'm not putting you down. I think you'd agree.

When John gets stuck using an online service, he throws the problem into ChatGPT. It gives him a decent how-to and gets him unstuck. Nobody forced him to do this. No enterprise deployment team put a license in his inbox.

In May 2026, Zitron wrote that "nobody wants this, nobody wanted it since the beginning, it was forced upon everyone".

Microsoft and Google do stuff AI into products people already use. Employers mandate tools. Enterprise licenses sit unused. That doesn't explain John.

Three months before Zitron published that claim, OpenAI reported more than 900 million weekly ChatGPT users and more than 50 million consumer subscribers. Those numbers come from OpenAI. Nationally representative surveys published by the National Bureau of Economic Research found that by late 2024 nearly 40 percent of Americans aged 18 to 64 had used generative AI. Twenty-three percent of employed respondents had used it for work during the previous week.

Push me, pull me, destroy my facts. John uses it. So do a lot of other people who have nothing to do with tech. Zitron's claim doesn't leave room for them, but they exist.

After a miss

Zitron isn't selling skepticism. He's selling the enormous relief of never having to change your mind.

When Gemini passed the target, the question became whether those users were valuable. When models became more efficient, the question became whether the labs could turn efficiency into profit. Each miss is promoted into a broader argument that is harder to test.

If every outcome supports the conclusion, the original prediction did no analytical work.

The broader economics argument may eventually be correct. It can't retroactively make the earlier predictions correct.

The circular financing deserves scrutiny. The environmental costs are real. Enterprise buyers can waste fortunes on software nobody uses. None of that rescues Zitron's predictions about Gemini, efficiency, technical progress, or demand.

I changed my mind because the evidence changed. That was irritating. It was also the job.

If you make your living demanding accountability from everyone else, you don't get immunity when the evidence turns on you.

Sources

-Dr. J