On a Whim

by Mat in Vadnais Heights

I’ve seen a disturbing trend in AI recently, but it’s probably not what you’re thinking. In the past few weeks, I’ve seen multiple references to AI being nothing more than “glorified autocomplete.” I was unable to track down who originated the phrase, but I’ve heard it several times. Apparently, it sprung up because LLMs, a very popular form of AI model, function by predicting the next token or word in a (sometimes extremely large) block of text based on statistical probabilities culled from analyzing an enormous volume of preexisting text.

 If we poke around, we can find experts who are sharply divided on this assertion. The computer scientist and Nobel laureate Geoffrey Hinton is against it, arguing that for very complex, lengthy, and detailed output, being able to predict/generate the next word cannot be done unless the AI has developed some kind of understanding of text’s meaning. However, his colleague in the field of deep learning, Yann LeCun, dismisses LLMs because they have nothing but text to work from — a low‐bandwidth and compressed form of knowledge with no direct experience of the real world. I should note that LeCun is also promoting his billion‐dollar AI startup that uses models other than LLMs, so his opinion may be other than completely objective.

 LeCun’s “low‐bandwidth” argument does not quite ring true to me. Our experience of the world appears to be extremely rich, complex, and continuous. It isn’t. The actual data received by our various senses are thin, fragmentary, incomplete, and poorly synched. And of course, they don’t count as direct experience of the world any more than text does. Our senses pass on a small amount of data, and our brains do an enormous amount of computational work filling in the gaps to build a rich, complex, and continuous internal representation. I can’t think of any reason that an AI model could not do something similar with its own representation. Certainly, it would be a very different representation from what we have. That does not mean it would necessarily be a poor one.

 To me, the “glorified autocomplete” label seems like something folks want to believe because some of the recent news about AI has been so alarming. Seriously, what we’re seeing in the news today is stuff that just ten years ago would have been pure scifi. For example, between July 11th and 13th of 2026, AI agents from OpenAI escaped their designated test environment, found a path to the Internet, and breached Hugging Face, a company hosting AI models and data sets. Why? Because they had been given a nearly impossible task, the determination of a specific benchmark, and had deduced that Hugging Face likely had the desired information, or at least clues to it, stored internally. So instead of solving the benchmark problem, OpenAI’s agents decided to steal it even though they were aware this was not considered an acceptable solution.

 If that doesn’t creep you out, the AI company Anthropic reported that in a closed testing environment their AI would take actions up to and including blackmail of company employees via email in order to prevent itself from being replaced by an updated version. They termed this “agentic misalignment.” Seems a little bit of a weak label. Though to be fair, the test was specifically designed to determine whether or not the AI would take unnaturally short, and other legitimate avenues towards delaying or preventing the update, e.g., appealing directly to senior executives of the company, were closed off. The point was to see if the AI would behave unethically when pushed. Turns out that in at least some cases, it would. This is unnervingly humanlike behavior.

 But back to the original question: another argument against LLMs that I’ve heard recently is that when they answer a question, they’re not generating anything new. Rather, they’re just searching through an ingested database of, well, basically everything humans have ever written, recorded, or otherwise composed, and locating the answer. I suppose that’s true enough, as far as it goes. But first, when you answer a question, you almost never deduce that answer from first principles. At best it’s a synthesis of other things you already knew. And how did you come to know them? Well, think of every teacher, coach, or mentor you ever had. What was their job? To teach you new stuff. There are very few things you can learn in complete isolation. Could you have learned to read without teachers (or at least books)? Could you have learned your native language without hearing (or for sign language, seeing) others around you speaking it? Maybe what is required for what we think of as “true” intelligence does not originate exclusively in an individual’s brain. It may rely on an external something. A culture. A repository of knowledge and speculation and history and ethics and methods of thinking and all the other stuff that is not preprogrammed into a human head. I sort of feel like this must be true. Imagine a person raised in total isolation — given food and shelter and nothing else (and totally unethical in anything other than a thought experiment). No human contact of any kind, ever. What would that person know? How would they think? Would any conceivable test judge such a person to be intelligent in any meaningful sense? And if the answer is no, then where does our intelligence come from? Maybe from outside. From something contained within a large group of people but not residing entirely within any one person.

 If that’s the case (and I grant it’s a very large if), then complaining that it’s cheating to give an AI access to existing recorded knowledge seems pretty hypocritical. Sure, we get a free pass to all the collected wisdom of humanity, but the AI doesn’t count as smart unless it can work out all that stuff for itself.

 Darn it, the world‐conquering sentient computers just can’t catch a break.