89 Comments
User's avatar
Pat Myers's avatar

Of all the astonishing things reported in this column, I am most astonished by the fact that AI cannot count to 2.

Scott's avatar

Language models are not knowledge models. The only thing they are good at is putting words in an extremely plausible order. They are remarkably good at this, and the most plausible order is often informed by knowledge, but it is in no way *governed* by knowledge.

Pat Myers's avatar

You'd think they'd have programmed a little arithmetic calculator in there, though.

Scott's avatar

As far as I know they do not have the ability to do this. LLMs aren’t programmed in the conventional sense, and the output isn’t derived from procedural instructions that can be readily augmented or modified. Some adjustment knobs can be turned, but the systems can’t exactly be told what to do. The hypothetical horrors abound.

Pat Myers's avatar

So do you think if you asked, “Tell me how many R’s are in this word,” it would reply, “I’m sorry, Dave, but I’m afraid I can’t do that”?

Scott's avatar

If you want that answer I’d suggest asking it to open the pod bay doors. Calling it “Hal” might improve the chances. Otherwise you’ll *usually* get the right answer, but *not* because it counted the Rs.

Lynne Larkin's avatar

It has trouble counting fingers, too.

Sam Mertens (he/him)'s avatar

I noticed once in passing that AI told me that the hexadecimal value 86 (decimal value 134, math which as a computer programmer I can do in my head) was equivalent to the decimal value … 86. It printed up a pretty chart to show me and everything.

Robert Ebbecke's avatar

Mid ‘70s, I was doing reliability tests on large, some new, some very old, power transformers in factories and steel mills along the rust belt. (The age record was a Westinghouse, nameplate said it was made in 1910.) One of the numbers we wanted was the Power Factor, a measure of efficiency derived from the results of a couple of different tests. I had done so many, I could do the calculation in my head. Until I bought my first calculator. Six months or so later, I broke the calculator. The synapses in my brain had changed. I had to go back to pencil and paper. AI is going to do that to us.

Sam Mertens (he/him)'s avatar

My brain used to recall tons of phone numbers. Now, it expects them to go into one address book or another, and pretty quickly drops them unless I make a concerted effort. Same issue, I’m sure.

Robert Ebbecke's avatar

In the ‘80s, I was partnered up with an older Korean engineer. I was 35, he was 72. We became good friends. As a youngster, he’d been educated in a tech university set up by the Japanese when they occupied Korea and built a lot of their heavy industry there. At one point, we were tasked with evaluating a bunch of the Washington Metro’s transformers, and I was back to doing that same old Power Factor calc. On the first transformer we did, I got my calculator out. He pulled a 5”x7” abacus out of his bag AND BEAT ME TO THE ANSWER.

Sam Mertens (he/him)'s avatar

I always admired the ability to use one of those. Evidently not enough to pick it up though.

Dale of Green Gables's avatar

The key question we face with AI is, what parts of human cognition do we want to protect, and what parts are we willing to outsource?

And yes, AI presents us with real cognitive risks:

* Loss of baseline competence --- If you can’t write a paragraph without AI, you can’t judge whether the AI wrote a good one.

* Overtrust --- Humans tend to trust fluent output, even when it’s wrong.

* Reduced cognitive resilience --- When the tool fails, the human may not have the fallback skills.

* Shallow thinking --- AI accelerates everything, including the temptation to skip depth.

But there's also the opportunity for real cognitive gains:

* Enhanced reasoning --- Humans can explore more complex ideas with AI as a scaffold.

* Faster learning --- Personalized, interactive explanation boosts comprehension.

* Creative expansion --- AI widens the space of possible ideas.

Dale of Green Gables's avatar

Recall how you asked it to make the conversion? AI models don’t “know” a number’s base unless you explicitly anchor it. If the prompt doesn’t strongly signal “this is hexadecimal,” the model defaults to the overwhelmingly common base in human text: decimal. So, the most reliable phrase would be something like: “Convert the hexadecimal value 0x86 to decimal.”

Sam Mertens (he/him)'s avatar

I don’t recall the exact query, but it was saying “this is hexadecimal” and “this decimal” and making a clear distinction. It was very transparent that AI was merely pattern matching and filling in blanks, not doing any actual cogitation.

Iowa David's avatar

So, Gemini doesn't know it's a twin? Mind blown....

Dale of Green Gables's avatar

Large language models don’t naturally “see” text as sequences of characters. They see "tokens," which are chunks like: “Trium;” “phal;” “ Arch;” "Trump” and “Hail.” Because of this, tasks like counting letters, checking anagrams, or verifying exact character patterns are unnatural for the architecture. When you ask AI: “What is 17 × 24? You imagine it doing multiplication. What it’s actually calculating is: “Given all the math-looking text I’ve seen, what number usually follows a question like this?” That’s why it can be brilliant on some math calculations and bizarrely wrong on others.

Sasquatch's avatar

Neither can Trump.

Nelsonsdad's avatar

Worse - I've found it cannot keep time effectively. It doesn't accurately keep track of days or weeks - as if it doesn't know how to read a calendar. I'm referring to the skills of Claude.ai, which is extremely useful in many ways, but just doesn't seem to understand how time passes in human lives.

Stephen Rockower's avatar

Gemini doesn't know how to count??? Why would it lie like that? Will it be the next Presidential Spokesman?

Jennifer Elsea's avatar

It must have been fed a lot of presidential statements on the economy to machine-learn alternate math. Odd that it would provide a lower number rather than vastly exaggerate it.

Steve Newman's avatar

Even so, Gemini is smarter than Trump. Sad. HUGILY Sad.

Sam Mertens (he/him)'s avatar

Just two days ago we had a tricky technical question at work about a piece of hardware. As is typical, manufacturers talked up certain features of their chips while completely omitting others. I sighed, and fired up the workplace -approved Gemini. I asked it my question - looking for confirmation, mostly. And Gemini went on to confidently tell me I was wrong. And it explained in convincing technical detail why a certain design decision had been made and why.

Well, this changes everything. It means I need to look somewhere else completely for an answer to the problem.

Except it didn’t feel quite right. I tried to find corroborating evidence on the web and failed, until I finally found that my original premise was correct and Gemini was full of shit, and led me on a wild goose chase. For goose shit. In fact, Gemini itself eventually told me.

Net work saved through the use of AI: -1.5 hours. Plus the loss of however many wetlands.

Donald  Ross's avatar

Paul Krugman wrote - "So, I almost never use dictation software. But I was in a mood today and tried for a bit. Strait of Hormuz" came out "street of hormones", which kind of makes sense.... "

Ed Rorie's avatar

“Meet Me Down on Hormone Street” would be a good title for a song.

Dave Zarrow's avatar

I don’t want to brag but … YES I DO!!

————————————————

Dave Zarrow • Apr 19 •

• All-star contributor

Surely others have noticed that TRIUMPHAL is an anagram for HAIL TRUMP. If he builds it in violation of the rules and laws that would confirm he is an ARCH criminal.

© DaZarCo Enterprises, 2026

————————————————

APRIL 19th, BABY!!!

kenneth gallant's avatar

Yep—that’s where I heard it before.

Jerry Slaff's avatar

I haven't seen or heard the word "roustabouts" used in years. Was he also prone to lollygagging?

gene weingarten's avatar

No, but they engaged in rapscallion pursuits.

Jerry Slaff's avatar

Is a rapscallion a hip-hop green onion?

Sasquatch's avatar

ehhh....don't be a blatherskite.

Julia Griffin's avatar

I think I would have accepted “Pizzabomba” as the confessor to Cardinal Pizzaballa …

Trevor Stone Irvin's avatar

Maybe Billy is working with Jeff Bezos at Blew-it-up Origins now.

Dale of Green Gables's avatar

Notice Lauren has also blown up. Must have been that everything, everywhere all at once package.

Gary E Masters's avatar

At least Jeff is working on a better path. And it is work t blow up a space craft. That is the development path. I remember the early explosions with Eisenhower as President. Just as if it were yesterday.

Sasquatch's avatar

When I saw today's headline, I thought Gene was referring to the New Glenn explosion. Too bad Jeff wasn't attached to the rocket.

Not Simple, Ever's avatar

TRIUMPHAL ARCH = HAIR MULCH TARP

Jennifer Elsea's avatar

The other day I was looking for a reliable source to cite for a paper, so I googled. The AI bot summary provided exactly what I wanted, so I clicked the links below that looked potentially authoritative. None of them really supported the assertion. Finally, for the first time ever, I clicked on the AI summary to review the full AI explanation. It provided an answer that was the polar opposite of the summary.

Pat Hendryx's avatar

ChatGPT diagnosed a medical condition I’ve been dealing with for a few years. I uploaded the scans, biopsies, lab reports and dr notes from multiple doctors. Thus far, it appears it is accurate. I use it as an aid when speaking to the drs (who don’t communicate with each other.) I’m not replacing my drs but I can see where the uninsured will.

Henry Cohen's avatar

What do you know? I attended JHS 82 and the Bronx HS of Science too. I'm two years older than Gene, so I attended JHS 82 in 1962 and 1963, rather than 1964 and 1965. The fact that Gene mentions only two years rather than three tells me that, like me, he must have been in an "SP" class and did 7th and 8th grade in one year. As soon as I read "My high school was for brainy nerds. His was for brainy delinquents," I knew that he was referring to Bronx Science and DeWitt Clinton. I graduated from Bronx Science in 1966, and since Gene apparently started in 1965, his first year overlapped with my final year. I don't remember meeting him.

gene weingarten's avatar

I was in the 3 year SPs program, but only knew Billy for two of those years, I think

Henry Cohen's avatar

I read at Gene's Wikipedia page that he was "the son of an accountant who worked as an Internal Revenue Service agent and a schoolteacher." That is exactly the case with me as well. This is getting eerie.

gene weingarten's avatar

Do you also refuse to dance, ever?

Henry Cohen's avatar

Did your father ever audit Captain Kangaroo, who then said hi to me and my brothers on tv?

Jennifer Elsea's avatar

Wow, the local tv clown (Mombo the Clown) once wished me and my sister well when we both came down with the measles! Right on TV! What kind of TV did you have?

Henry Cohen's avatar

There is a paragraph on Mombo in the Wikipedia entry for "The Dr. Max Show."

Jennifer Elsea's avatar

How we loved Dr. Max and Mombo! Come to think of it, it might have been Dr. Max who mentioned us. We must have been 4 at the time. Convinced the duo were actually inside the TV set, we demanded to inspect the back of it to figure out how this information might have made its way to Dr. Max. Dad played along. We were mystified. If we knew then what we know now about the availability of personal info, we might have been scared.

Jack Ohman’s You Betcha!'s avatar

I found this column through AI, and I’m on the lam for bank robbery.

Dale of Green Gables's avatar

Problem is Dear Leader you (and far too many others) have bought into the hype or conventional wisdom (take your pick) that a technology that depends entirely on fallible humans is itself somehow infallible. Consider the technologies that fail consistently but don't get AI-level outrage:

* GPS systems

* Autocorrect/predictive text

* Medical diagnostic tools

* Airline reservation and baggage systems

* Credit bureaus/financial scoring systems

* Weather forecasting models

* Search-engines

* Self-checkout machines

Probably boils down to people fearing unfamiliar failure modes more than familiar ones and the fact AI does touch on purely human creativity, identity and cognition. GPS can’t write a poem or impersonate your boss. AI can. That raises stakes. And that AI is tied to existential narratives doesn't help either. Certainly pisses it off, but no one (okay, relatively few...) believes autocorrect will overthrow humanity. History is replete with similar tech "panics," starting maybe even with Socrates who himself panicked about writing, arguing it would destroy memory, give only the appearance of wisdom, and undermine real dialogue. Modern AI is powerful pattern‑matching machinery, not a thinking entity --- so ask once and check (at least) twice and especially with "hard" questions. Hard questions aren't usually about patterns. When there’s no “most likely” answer in the training data, the model can wobble or hallucinate. So yeah, kinda like fire, the wheel or slider storage bags, but a multiplier in a different way.

AAM's avatar

The technologies you mention are programmed to perform specific tasks using defined data sets. They certainly were detrimental to some jobs and businesses, such as cashiers and physical maps. (Maps! The lost joy of road-trip navigation. Asking now nonexistent gas station attendants for directions, or refusing to). The AI data set includes whatever some wise guy feeds into it. AI errors are called "hallucinations" Many people are losing jobs on the assumption that this flawed technology will improve productivity, but employees stuck with it are told, "Check the results because they could be wrong". It convinces fragile people to commit suicide. It will lessen our ability to think, which has already started: my co-worker begins his prompts with, "please". It will harm life in a way other technology hasn't.

Sorry to go on. Nothing personal.

Robert Ebbecke's avatar

Reminds me of a comment I read in the NYT when GPS was new. “The problem with GPS is it tells you where to go, but it doesn’t tell you where you are.” I still keep a collection of actual roadmaps of the states I am likely to travel so that I can see what interesting places (to me, not google) might be near my planned route.

Dale of Green Gables's avatar

Nothing to be sorry about, although while you're correct that some of the technologies on the list do use fixed, defined datasets, others use learning systems that behave more like AI (e.g. search engines, autocorrect/predictive text and some medical diagnostics). But my point is AI per se is not the issue --- it is a tool and very good cognitive multiplier if used properly. The issue or issues are the reaction to it, or acceptance of it as some all-knowing genie (as I attempted to point out) and the need to consider its social and economic impact.

Mark Asquino's avatar

You’re getting such a wildly wrong answer is the morale of the story.