Transcription
[music] >> In March of last year, a paper landed quietly online. Language models, it said, had passed the Turing test. And if you hadn't heard of it, the test is simple. Can a machine hold a conversation well enough that you can't tell it's a machine? For 75 years, nothing really could. However, last March, there was enough evidence that something did. Some people celebrated. Some called it meaningless. Both sides thought they were having a new conversation. They were not.
This argument actually started in 1637, over 400 years ago. Rene Descartes, writing in Discourse on the Method, posed a question that will sound very familiar. If you had a machine that looked like and moved like a person, how would you know it wasn't one? He came up with two tests. The first was the one we care about. A machine, he said, could be built so that it makes a sound when you touch it. Touch it here, it speaks. Touch it there, it cries out that it's hurt. But it could not arrange those words freely and reply to what was said to it, not the way even the simplest person can. This was the Turing test, but smaller, written 313 years before Alan Turing.
I go deeper in this on a piece called Post-Israel Governance, where I trace some of it all the way back to Aristotle, giving tools like torches and tent. But for now, we stay focused. 200 years later after Descartes, a woman named Ada Lovelace read Charles Babbage's plans for the Analytical Engine, a mechanical computer, and she saw the next test coming. Lovelace knew that this Analytical Engine could do more than arithmetic. Symbols could stand for notes, for colors, maybe even language. She wrote that such a machine, given the right rules, could compose music, scientific music, of any degree of complexity, her words. But she drew a line. The engine could follow rules, it could not make them. The engine, she wrote, has no pretensions whatsoever to originate anything. That line is still holding up a lot of arguments today. And what I love about Lovelace is she didn't just predict computation, she predicted language models. She said that the engine could combine its numerical quantities exactly as if they were letters or any other general symbols. She was describing natural language processing in the 1830s. However, Lovelace drew her line at origination, being able to be original about something. A machine can follow rules, it can't make them.
100 years later, Alan Turing looked at that line and asked a harder question. What does originality actually mean? In 1950, he wrote Computing Machinery and Intelligence, the paper where the actual Turing test appears. And his argument with Lovelace is the crux of the whole thing. Turing said that the problem with Lovelace's line wasn't the machine, it was the measure. What if originality itself is just a reshaping of things we were taught? Think about how you learned anything. School taught you language. Your parents taught you values. You read books, watched movies, had conversations. All of it loaded in. Then one day you write something that feels like yours, but where did it come from? It was built out of everything that was loaded in before. We don't call a student just following their training when they write a good essay, we [music] call that thinking. So when a machine does the same thing at scale, why call it less? If a machine surprises us, that should count, should it not? If a program can hold up its end of a conversation well enough to fool a clear observer, we should treat that as thinking, no? Well, at least that's what Turing believed. He even addressed Lovelace directly and said, "I do not assert that machines have not got the property of thinking. I assert that the evidence available to Lady Lovelace did not encourage her to believe that they had it." Turing didn't claim to know what thinking is. He just chose to test behavior and outcomes instead. And that is a different move than Descartes or Lovelace, and the field still hasn't necessarily decided on which side is better.
Fast forward to March 2025, the paper I was talking about at the beginning shows that language models passing the Turing test with multiple people and peer-reviewed studies. And the response tells you more than the result does. Some people said the score was shallow. Others said the structure of the brain is fundamentally different from a chip, so it isn't a fair match. And I do think the structural argument might be an interesting route to go on. My brain runs on a sandwich and coffee. These models need a nuclear reactor worth the power to train, and then another cloud database to even get it close to getting a question wrong. That's not the same kind of machine at all. But here's the catch. Every argument people make against AI tests, we actually make against humans. An IQ test grades outputs, not inner life, but we use it anyway, and we say it marks intelligence. Turing saw this as well. He said, "You cannot be sure that some causal remark of your own hasn't started the idea off. The remark could come from a teacher or from pre-training data, or just a puzzle of the mind."
And this is where I think most of the current AI conversation is actually stuck. We keep arguing about whether a model is really thinking. I think that's the wrong abstraction layer for whether you should use them. I think it's a great question, but the abstraction layer for use, for using these models in an intelligent way, is more at a higher system. What I mean by this is what is underneath the abstraction is mostly hidden and doesn't matter if you focus on the output. When you drive a car, you don't think about pistons firing or the fuel injection timing, you think about where you're going. That's a high abstraction layer. Same here. The useful question isn't is the model really thinking? The useful question is whether the output is reliable for the task in front of you, and whether you're on the hook for when it isn't. Those are ethics questions. Those are engineering and business questions. Thinking is an entirely separate thread. Pulling on it doesn't get you the outcome. Yet, I still think it's an important question. And we have to ask, why do we keep pulling on that thread in the first place? Well, I think it really goes back to Descartes again, the 400 years ago. When we look at Descartes's dualism, dualism is the idea that the mind and body are made of fundamentally different kinds of stuff. One is physical. You can weigh your brain. You can see it on a scan. The other isn't. You can't really weigh a thought. You can't put a memory on a scale and hold it in your hands. Two different substances, two completely different worlds. Descartes needed that gap to eventually get to his famous statement, "I think, therefore I am." 400 years later, that gap is still something we're looking at.
Look at cybernetics, the study of how systems sense and correct themselves. A thermostat is technically cybernetics. Check the temperature, compare it to the target, adjust. Cybernetics shows that purpose, the thing we used to think only minds have, can live in a loop. Norbert Wiener watched anti-aircraft guns track fast planes and saw a kind of intent in the mechanisms. From the outside, feedback and intention look identical. The loop is doing the work. And behaviorism took it step further. B.F. Skinner trained pigeons to guide missiles by pecking at a target. The birds produced steering curves as smooth as calculus. You can call that intelligence and that curve, or you can call it training data, and the curve is the same regardless of what you call it. Extended mind theory goes the other direction. Andy Clark argued that your phone, your notebook, the cloud model holding your calendar and your drafts, all of it is part of your cognition. Not a tool your mind uses, but actually part of your mind. If your cortex can reach into your phone for a memory it outsourced, then mind isn't sealed inside your skull. And if it isn't sealed inside your skull, why would it have to be sealed inside flesh?
But we finally get a resistance to this with a man by the name John Searle. He tries to freeze the debate in 1980 with something called the Chinese Room. Imagine it this way. You're locked in a room. You don't speak a word of Chinese. People outside the room slide in Chinese characters and words through a slot. Inside, you have a rule book. It says things like, "If you see these characters and words, slide these characters and words back." You follow the rules. You have no idea what any of it actually means, though. However, the people on the outside think you understand Chinese perfectly. Searle's question was, "Do you actually understand Chinese if you're just memorizing the rules around it?" His answer was no. You're just shuffling symbols, which means a computer can't really understand, either. It's doing the exact same thing you're doing, just faster and with a bigger rule book. Now, it still kind of holds today, but not if you look at it from a different perspective. The room is not you. The room is a system. You plus the rule book, plus the paper, plus the slots, plus all of it together, when the system answers in fluent Chinese, the understanding has to live somewhere in that system. Sure, it's not in your head alone, but it is in the whole setup. The man inside the room doesn't understand Chinese, but the room does. Mind leaks into matter. Matter hosts mind. The border was never as sharp as Descartes actually needed it to be.
And here's the part that I actually work on and I was thinking about while doing my research at the Edinburgh Futures Institute. You see, I was running psychometric scales on language models. You know, the same instruments psychologists use on people to figure out our personalities. I sent them towards a whole bunch of AI and the scores actually come back pretty consistently. Back then, GPT-4 read as very open and low on neuroticism. Claude was leading as guarded. That doesn't necessarily prove that the models are thinking, but it shows something quieter and actually more useful. A rule-bound system can carry a moral or psychological tone. Every sentence you ask it to finish will drift in the direction of that psychological bias, unless you notice it. And even if you notice it, you have to figure out what you want to do with that bias. [music] That's the practical version of this whole debate. The question isn't is it conscious? Can it think? The question is what value is this thing quietly handing back to you? And are you going to notice before you ship that lesson plan, that policy draft, that court memo? Descartes would not have framed it that way, but he would recognize the stakes, the problem with these systems.
And here's where I keep landing. The discourse isn't really a quest for an answer. I don't think we'll ever find one. It's a mirror. When we ask if a machine can think, we sharpen what we mean by thinking, by what we are. And if we let it, the question turns outward, [music] because I need someone else to read this, to listen to this for any of it to have meaning. I need a point of reference. I need someone else to compare this to. When Einstein said you cannot know how fast something is moving if there's nothing to compare it to. I think the same is true of consciousness, of thinking. It's relational. Descartes said, I think, therefore I am. But maybe it's closer to you think, therefore I am. Stay curious, my friends, and until next time, happy learning.