A share price that never existed
Ed Zitron was about to paste a table into his newsletter when a number stopped him. He had used Ask B, the assistant built into his Bloomberg terminal, to pull the growth rate of Microsoft, Google, Meta and Amazon shares over what he recalled as five years. Ask B saves him from writing BQL, Bloomberg's own query language: he types a question, the tool writes and runs the query, and the results come back with their sources attached. That design, he said, "deals with hallucinations real well." Then he looked at the spreadsheet. "Microsoft stock's never been $575 a stock."
He caught it, and the cost was nothing. "When it's a cute little thing like, oh, it's a stock price and I can't afford it... It was no harm, no foul." His worry is the same error somewhere with no one looking: a transcription tool used by a doctor, a financial model a hedge fund depends on, or a refactoring job on a codebase that quietly leaves a security hole open. Someone who has been vibe coding for six months, he said — describing software behaviour without writing much of it by hand — may have lost the habits needed to spot it. The problems, in his phrase, "become multiplicative." He added that he does not know how the industry trains models out of this, and that it has "certainly not succeeded."
The host of the episode reported a different experience. He remembered showing ChatGPT to his fiancée when it launched, watching it get things wrong now and then. "I actually don't have that experience anymore," he said. "I have moments where I believe its reasoning is weak, but I don't have outright hallucinations anymore." Zitron: "See, I disagree."
What the leaderboards actually measure
The host's evidence was a benchmark. Citing historical data from Vectara's hallucination leaderboard, he said hallucination rates on simple summarization tasks had fallen from roughly 21.8% four years ago to about 0.7% on today's leading models — adding his own caveat that these are simple tasks, which matched what he saw day to day.
What the Vectara leaderboard scores is narrower than "how often is it wrong." It gives a model a document and asks for a summary using only that document's content, then uses an automated evaluator to judge whether the summary stays factually consistent with that source. The maintained methodology runs over 7,700 privately held articles, ranging from 50 to 24,000 words, generally at temperature zero, and reports factual consistency, the complementary hallucination rate, summary length, and an answer rate — reported separately because some models decline otherwise ordinary source material. Because the evaluator and the dataset have gone through versions, comparing percentages across years means matching benchmark versions.
Zitron's complaint was about this kind of number in general: the industry evaluates progress with tests "adjusted specifically for large language models." His example was METR, whose chart of lengthening tasks he paraphrased with a sting in the tail — look how long it can run, an hour, and then you read the small print: succeeding half the time. METR's measure is built the other way round from how it is usually repeated. The "hour" is the time an appropriately skilled human takes on the task, not how long the model runs unattended; a fitted curve gives the human task duration at which a model succeeds with a chosen probability, such as 50% or 80%. On that scale the March 2025 analysis of software and reasoning tasks found the 50% horizon roughly doubling every seven months over six years, put Claude 3.7 Sonnet at around an hour — meaning predicted success on half of comparable tasks — and found models much stronger on very short tasks and weak on tasks taking humans several hours. The results depend on which tasks were selected and on the human baselines used.
What he pays Matt Hughes for
The host's strongest objection was that perfection is the wrong comparison. For a simple writing task, he argued, the alternative is his own time and his own gaps — or an intern who is also "prone to hallucination." If something hallucinates 0.7% of the time but knows more and is faster, "maybe on a net basis, that's a good trade."
Zitron answered with a person. "Someone I love dearly. Matt Hughes, my editor." Hughes lives in Liverpool and is, in Zitron's description, a decorated tech journalist with empathy and joy for the things he loves and venom for the people he hates. "I don't pay Matt Hughes because he knows everything. I pay him because he has incredible context and a ton of knowledge." He named the willingness to extend that knowledge, the moral support and the editing itself. "I can't get that from a large language model."
He also refused the premise underneath "knows everything": "What good is something that knows everything when it sometimes doesn't know anything?" And the intern comparison, he said, gets interns backwards. Nobody hires one to look up the date; you hire one to train them out of what he called Pinocchio status.
Files are not memory
The host said an AI also gets context and learns. "No, it doesn't," Zitron said. The way it learns, in his account, is that you write a giant instructions file that the model sometimes reads and sometimes doesn't, and build a harness around it — an arrangement he compared to Pee-wee's breakfast machine, a contraption of contrivances assembled to keep the hallucinations down.
The host offered a deliberately small demonstration: ask Claude his dog's name and it knows. ("Jesus Christ," Zitron cut in. "This company raised $95 billion this year.") The point, the host said, was only that it can remember things from the past. Zitron conceded the files and denied the conclusion: it has files it can access, "but that's not the same as memory." Does it know your moods, he asked, or what is happening in the world around it, or have good days and bad days. An intern is something you invest in and that grows; human memory, he said, accrues as a mix of emotion, feeling and fact.
Process or output
Here the host made his sharpest move. There is the process and there is the output, he said, and people pay for the output. Dump your files into Claude and you need not care how it processes them, as long as it has the number when you ask for your revenue. Train a person for years and the question at the end is the same: do they know the revenue number when you ask?
Zitron disagreed about the process, and described a day of it. He and Hughes had been working through research on asset managers, Blackstone among them, going back and forth in disbelief at what they were reading; Zitron then sat for a day-long session and wrote 11,000 words from the notes Hughes had given him. "We were both learning together," he said. "And the learning process was as much about creating the output as the output itself." A model, by contrast, produces what he called the average of the documents it could find. He said he had used the higher-end harnesses hedge funds run, and "they all give the same shite" — the same generic reports, the same noticed analysis.
The host granted that this session bought learning and connection as well as an edited blog post, but held the line for his own example: when he wants his revenue number read back, he cares about the output. What if the number was wrong once, Zitron asked. You should have deterministic ways of knowing such numbers and not rely on a model for them; even with BQL, which he trusts, he double- and triple-checks. And the longer the report, he added, the more there is in it to be wrong.
What trust is made of
The host tried to isolate the ingredient. What is your trust in Matt based on — historical performance? Yes, Zitron said, and the fact that they had learned half of this material together. Tenure alone cannot be it, the host countered; there are people you have known for 15 years and do not trust. His answer was continual delivery on commitments — and on simple tasks, he said, the hallucination leaderboard shows Claude continually delivering.
"I mean, is that what that board says?" Zitron asked.
It says whether it is getting things wrong, the host replied. On simple tasks. How are those defined?
"I don't know."
That, Zitron said, is the illustrative thing about the industry: "They are the whataboutist masters." A benchmark says they are good at something and the number goes up. "What's the number mean?" He said he did not mean it as a criticism of the host — but when a company wants to prove its worth, he said, the answer cannot be to point at the product and tell you to use it.