I think a thread on this topic will be interesting. My own position is that AI is intelligent, and that’s for a very simple reason: it can do things that require intelligence. That sounds circular, and in one sense it is. In another sense it isn’t. It’s a way of saying that we don’t have to examine the internal workings of a system to decide that it’s intelligent. Behavior alone is sufficient to make that determination. Intelligence is as intelligence does.
You might ask how I can judge intelligence in a system if I haven’t defined what intelligence actually is. My answer is that we already judge intelligence in humans and animals without a precise definition, so why should it be any different for machines? There are lots of concepts for which we don’t have precise definitions, yet we’re able to discuss them coherently. They’re the “I know it when I see it” concepts. I regard intelligence as one of those. The boundaries might be fuzzy, but we’re able to confidently say that some activities require intelligence (inventing the calculus) and others don’t (breathing).
I know that some readers will disagree with my functionalist view of intelligence, and that’s good. It should make for an interesting discussion.
keiths:
Erik:
No, you don’t. I already had Gemini generate a story for you that clearly is not in its training dataset (or anywhere on the internet, for that matter):
Erik:
You’re confusing plagiarism detection with AI detection. They aren’t the same. Something can be plagiarized but not AI-produced; it can be AI-produced but not plagiarized; and it can be both plagiarized and AI-produced. It can also be neither plagiarized nor AI-produced. All four combinations are possible, because the properties are orthogonal.
No, because I understand how neural networks work. They don’t store their training dataset in a database. They can’t just look up everything they’ve been trained on. Anyway, it’s clear that AIs don’t merely plagiarize. See the above property tax/lemon meringue/sentient nebula story. I challenge you to go out on the internet and find the original from which Gemini was cribbing. You won’t find it, because there isn’t one.
Likewise, I challenge you to search far and wide on the internet for an image from which this Gemini-generated image was cribbed. Good luck.
Huh? College professors struggle to recognize stories when looking at them? Where is this strange place in which you live, where college professors struggle to identify stories? Where I’m from, even children know the difference between stories and non-stories.
Try it out in your strange land. Find a kid. Read them “Goldilocks and the Three Bears” and ask them if it’s a story. Then do the same with a page from the local phone book. Report your results here.
I have no idea what you’re talking about, unless you’re saying that college professors, like everyone else, learn to recognize stories by being exposed to them. It’s the same with AIs.
Why not continue in the ‘Antivax’ thread? You seemed to be rolling your sleeves up for a good old ding-dong, then just withered away. I addressed your point on flu – silence. I addressed the ‘definition of vaccine’ trope. Silence. But every now and then you pop up in other threads to say something vague and petulant, then disappear again.
Perhaps indirectly, we’ve been talking about the Turing test. The idea that if a person simply cannot tell the difference between human and AI output, then that AI (if it is one!) qualifies as being as intelligent as a person.
Erik emphasizes that if we don’t know the process which produces that output, we cannot tell if it was generated by a person (intelligent) or an AI (simulated). In other words, it’s the process that matters and not the result. Which kind of echoes the efforts to figure out how the human brain thinks. The human brain is almost entirely a black box, not easy to model. I think neural networks evolved as an effort to simulate how the brain works, but they might not be very close.
lol you are full of s…t…
BTW: there is no such thing as “vaccination” . RFK jr knew it and so does the Trumpet. The Amish and the Mennonites etc are the proof …
vaccination is not possible or necessary unless there is a pathogen, would you agree?
What people like you MUST understand is that there are no vaccines and the covid -19 scam was a perfect example how people in the field have been forced to accept the scamdemic narrative because it meant their livelyhood.
While this surely seems like the ravings of an idiot, what we MUST understand is that the country is chock full of J-Macs, and they vote.
Flint:
Yes, with the caveat that some people are fooled by mimicry as opposed to genuine intelligence. That’s what happened with ELIZA. But AI has definitely passed the Turing test. Here’s a study from earlier this year:
Ironically, the AI has to play dumb in order to pass, because if it’s too smart, the person will realize that it’s not human. Playing dumb is a skill in itself. You might ask: what if the AI genuinely is dumb, and isn’t merely playing dumb? I think the examples I’ve cited earlier in the thread show pretty conclusively that AIs aren’t dumb, and if they were dumb, they’d have to be dumb in a humanlike way in order to pass the Turing test.
The Turing test is important, but in my opinion the cases where AIs easily and obviously exceed human capabilities, in tasks that clearly require intelligence, are more convincing. In one chat, Claude solved my original physics problem in 19 seconds. When I complicated the problem by specifying variable gravity, he got the answer in 90 seconds. That’s superhuman.
Yes, and it’s not clear to me what he thinks the process is in the case of real intelligence. He seems to believe that it involves something non-physical ( a soul?), though he’s never elaborated. My suspicion is that his belief in this non-physical entity or process makes him want to deny AI intelligence, in order to preserve human exceptionalism. Erik, since you believe the process isn’t entirely physical, could you explain how the physical interacts with the non-physical when humans do “real” arithmetic? If it turned out that humans do arithmetic without the aid of anything nonphysical, solely by the action of their physical neurons, would you then declare it to be “fake” arithmetic?
The efforts to model the brain are distinct from the efforts to solve problems using neural networks. One key difference: in a biological brain, neurons either fire or they don’t. Their output is either ‘1’ or ‘0’, in other words. In artificial neural networks (the ones aimed at problem solving, anyway), a neuron can produce a non-binary output.
Flint:
Thanks for the reminder. I’m going to have nightmares tonight. 😠
I spent several months learning to troubleshoot core memory. One of its distinctive features is that a core can have multiple imputs that can cancel or reinforce each other.
Neurons can have thousands of competing inputs, and the key behavior is not firing, but rate of firing. I confess I do not know how artificial neurons work, but I read that the behavior is probabilistic.
I suspect we are at the beginning of AI. The model T era.
petrushka:
Wow. I’ve never met anyone who worked with core memory. Was that part of your training while serving in Vietnam?
For readers who aren’t familiar, core memory was the very first form of RAM. Here’s what it looked like (at high magnification):
The idea is really clever. Each of those ferrite rings is magnetizable in one of two directions, meaning that it holds one bit of information — a zero or a one. Passing currents through those wires in the right way will magnetize the ring, and doing so in a different way will read out the value of the ring. A cool aspect of core memory is that it retains its values when powered down, because the rings don’t depend on current in order to maintain their magnetization.
Yes, because (at least in the designs I know about) the synaptic weights are stored by modulating the resistance of memristors. There’s lots of noise and variability both in setting the weights and during normal operation, so the outputs of the networks are nondeterministic.
ETA: The term “core dump” is standard terminology for when you dump the contents of memory en masse in order to figure out what went wrong with your code. I just took the term for granted during my career, but it’s clear to me now that when it was coined, it referred literally to the contents of core memory being dumped.
Here’s another example of Claude’s intelligence. I needed an interactive interpreter for a fictional instruction set I created with Claude. (If you don’t know what that means, all will be described in a forthcoming OP). I asked Claude to code one for me, and he came up with a sophisticated interpreter with a very slick interface:
Did it work on the first try? No, which will excite Erik. (Claude makes mistakes! He’s not intelligent!) We had to debug it together: I described the failure symptoms and Claude corrected the code. It took four iterations to come up with a fully working interpreter. Total time: around 45 minutes.
I didn’t have to write a single line of code. All I had to do was test the interpreter and report the results to Claude. He took care of everything else. During my career, if I had interviewed someone who could pull that off, I would have hired them on the spot.
That ain’t “simulated intelligence”. It’s very real.
I have not been talking about Turing test, and won’t. In real life, “whether it passes” is not the only thing that matters. Teachers and professors do not merely give grades. They also assess whodunit. Students have been plagiarising from each other or writing essays for each other for ages and professors have been fighting this for ages. The lure of AI is of course very strong – professors themselves use it while preparing tests and lectures. The Turing test is woefully inadequate to even begin talking about this phenomenon, much less to address the problem.
In military intelligence, any piece of information is assessed from not only the angle of what impression it gives, but also who is trying to give that impression and who is it useful for that such impression prevail. (When done professionally, military intelligence considers even more, such as “What if this is a false lead? When we start investigating this, is there something else important that we could be missing?”)
E.g. let’s say you find a document is being circulated in the military saying that the army is gay, generals are fat, and male standards need to be improved across the board. The questions are how widely it has circulated thus far, whether the critical mass of soldiers is in agreement with this and how it got started. The goal of such a document is obviously to hurt the army morale, but it is important to figure out whether the impulse came from the inside or from the outside. To solve the last question the normal procedure is to go down the chain of people how the document reached yourself.
But if it’s a speech given to an audience, then depends who made the speech. A soldier or mid-tier officer would get investigated to determine if he is a foreign spy or plant or came up with the idea on his own and whether there is ideological motivation for it or it’s a prank. But if the speech was given on the level of a “Secretary of War” then one cannot help but agree that the army needs reforms, perhaps even a revolution, and specifically the kind of revolution where the “Secretary of War” is decapitated.
The Turing test neither touches or informs any of this.
I’m retired, so no, doesn’t affect my livelihood one bit.
Take it to ‘Antivax’. I would be happy to address your wild-eyed conspiracist bollocks there.
Erik,
I did a quick scan through the thread to collect all the arguments you’ve put forth against AI being true intelligence. Let me know if I’ve missed any. I’ll list each argument briefly along with my response.
1. Argument:
Response:
2. Argument:
Response:
3. Argument:
Response:
4. Argument:
Response:
5. Argument:
Response:
6. Argument:
Response:
7. Argument:
Response:
8. Argument:
Response:
9. Argument:
Response:
10. Argument:
Response:
11. Argument:
Response:
12. Argument:
Response:
13. Argument:
Response:
14. Argument:
Response:
15. Argument:
Response:
Have I missed any of your arguments?
The core I worked on was used in computation, not as memory. I will leave to your imagination what the application might be. I will only say it was streaming. It was a dinosaur in 1967, when I encountered it.
I suspect that the next step in improving the sanity of AI will be to evolve low level modules that embody aspects of physics and chemistry. Things that are stable and which can be validated without text. An alligator brain. I’m pretty sure this is already in the works.
It might be called common sense, but that is a problematic term.
Anyway, a reliable AI needs a lot of pre-formed knowledge before learning to talk.
petrushka:
Yes, they’re already adding modules like that. Claude has a few that are visible to the user, and probably more that are invisible. There’s a separate module for searching through your chat history. You can say things like “find that chat in which you were joking about whether you were sentient”,* and it will call an intelligent submodule that understands the semantics well enough to locate that chat even if you don’t specify a word or words to search for. It’s necessary to use a separate module for this function because if Claude himself had to search through earlier chats, he’d exhaust his context window before returning results.
There’s another low-level module for doing calculations. Below a certain threshold of complexity, Claude himself does the work, but larger calculations would be a waste of context window space as well as being more error-prone. In those cases he actually writes javascript code and runs it in a sandbox to get results. Note that I’m not talking about when a user asks him to write code. This is him spontaneously deciding to do a calculation via code.
A third module deals with what are called “artifacts”, which are really just files that Claude can access and update. The interpreter that Claude wrote for me (mentioned above) is an artifact, and I can run it myself within the Claude user interface without involving him (but he can also run it if he needs to.) He’s also maintaining a language spec and a comprehensive document tracking the results of the coding experiments I’m running.
* That conversation was a funny one. Here it is, edited for brevity:
What’s fascinating is that he is able to determine when it is safe to joke with someone, knowing that he won’t be taken seriously. With someone else, and a different tone of conversation, I don’t think he would have ventured to plant any doubts about his non-sentience.
Yes. You are missing that there is fundamentally just one argument: Simulation versus the real thing. You have evaded it completely. Also, none of your other replies addresses any point. Some are horrendous logical errors, such as your response to #1 (to reply to a sound and valid syllogism that it assumes its conclusion is simply invalid) and some ignore plain facts, such as #4.
In #4 I say that AI does not recognise the categories where creativity can be applied and where not, and your silly response is that it got your physics problem right. In your other answers at the same time you admit that AI gets chess wrong, but you call it “mistakes” rather than what it really is, thus not just missing the point but ignoring facts about AI. As the example with law assistants demonstrates, AI clearly makes random stuff up where not supposed to – it goes far beyond “mistakes”.
And of course, the fundamental point is that AI of its own does nothing, creates nothing, thinks nothing. Because it is a machine. It merely churns (when prompted) the pre-existing data a bit, and all of its magic is in the fact that the database is enormous. For the result to be useful, there are literal humans at work behind the scenes.
– You are not the only one testing out and interacting with AI. To be relevant in the discussion, you have to acknowledge that others have gotten different results from you and drawn different conclusions than you. I myself do stuff with AI. I know what I am looking at and I see what is going on.
– You fail to acknowledge that AI is a machine. As long as you talk about it as if it were a human or an animal, everything you say is off base. You are not being analytical, but rather brainwashed by marketspeak.
– You have no clue what a simulation is and you have no willingness to learn what it is. You could make it easy and just ask your AI but you still haven’t.
This isn’t something I know much about. Lizzie, in this field, said something very different. Her analysis was that there is a very large number of neurons, all either firing or not, and the net result is kind of like taking a vote. Very non-binary.
Erik:
I’ve addressed it head-on. I explained the difference between simulated self-driving and real self-driving, and I pointed out that in real self-driving, the car physically moves from point A to point B in the real world. In simulated self-driving, the only motion is from A to B in the simulated world. There is no physical motion. Simulated self-driving produces no real-world motion. Real self-driving does.
When an AI writes a story, the story ends up in the real world. It’s a real story, and it serves the functions of a real story. It doesn’t only exist in a simulated world. Since AIs write real stories, they are really writing. The writing isn’t simulated.
If you disagree, tell us how fake writing manages to produce real stories. And if you want to argue that a story isn’t real if written by an AI, despite the fact that people recognize it as a story, then justify that claim. Don’t just define it as a fake story because it was written by an AI. That’s assuming your conclusion, and it’s boring. We already know that you don’t think AIs are intelligent. Codifying that as a premise doesn’t accomplish anything. We need a demonstration, not a mere assertion.
Here’s #1:
And my response:
The “syllogism” (which isn’t actually a syllogism):
That is literally assuming your conclusion. The argument is valid but not sound. Valid because the conclusion follows from the premise, being a simple restatement of it. Not sound unless you can demonstrate the truth of the premise. You haven’t.
If you want to make some headway, tell us why machine intelligence is fake intelligence without saying “because machines aren’t intelligent.”
No, this was my response. It was right there, under #4, so you have no excuse for distorting it:
That falsifies your claim that they don’t know the difference.
It’s known as “hallucinating”, and you may have heard that word before, because it applies to humans. If a human hallucinates, does that make them unintelligent? Ever heard of John Nash, the subject of A Beautiful Mind? Hallucinated all the time, but you’ll get laughed at if you claim that he wasn’t intelligent.
Because I, Erik, say so. It doesn’t matter what an AI produces. I have spoken: Machines are unintelligent.
It creates new material, like humans do. It generates images that are nowhere in its training data. It writes stories that did not exist until it created them. It solves physics problems that it’s never seen before. It programs in a brand new computer language that it’s encountering for the first time (my ongoing experiment).
Those literal humans participate in the training process of AI. Likewise, there are literal humans that participate in the training process of people. They are known as teachers. Therefore people are unintelligent?
A woman — let’s call her Erika — reads a couple accounts of cats being unable to swim. Let’s say she even witnesses a cat or two being unable to swim. She concludes that cats can’t swim. A different woman — let’s call her Kathy — sees cats swimming all the time. She therefore disagrees with Erika. Who’s right?
Kathy, obviously. Her observations falsify Erika’s claim. Would you tell Kathy “No, no, no. You can’t conclude that it’s possible for cats to swim. Erika’s observations differ from yours, and she has drawn different conclusions”?
I hope the parallel is obvious.
I’m the one insisting that AI isn’t sentient. I’m the one talking about how AI is implemented on digital hardware. It’s pretty obvious that I am acknowledging that AI is a machine (or more properly, that AI is software running on digital hardware.)
Pretending that I don’t know what a simulation is makes you sound goofy. Your choice. But if you want to claim that after literally thousands of hours running and debugging simulations, I still don’t know what they are, then tell us: what precisely is a simulation, in your view, and why is computer arithmetic a simulation if human arithmetic is not? (We can talk about intelligence too, but arithmetic is a good place to start.)
I’m especially interested in this:
Looking forward to your answer, because I think the question will reveal something about your stubborn adherence to the “machines can’t be intelligent” dogma.
keiths:
Flint:
The key phrase is “either firing or not”. Each neuron either fires, or it doesn’t, and that’s binary. The network as a whole isn’t binary — there are way more than two possible states. But that’s true of computers as well. Flip-flops are binary, logic gates are binary, but processors can be in way more than two states.
Collectively, biological neural networks are non-binary, but neurons are binary in that they either fire or they don’t. Collectively, processors are non-binary, but flip-flops and gates are binary in that they produce only two values: 0 or 1.
Indeed.
Wikipedia basics on neurone firing.
Keiths, in the OP, states:
…I haven’t defined what intelligence actually is…
and
I know that some readers will disagree with my functionalist view of intelligence.
The subsequent comments illustrate the difficulty of talking about an undefined concept.
Alan:
Not really. I addressed that in a response to Neil, who wrote:
keiths:
Neil:
keiths:
I think there’s broad agreement here on what counts as intelligent behavior when humans do it. It’s just that Erik adds a rider: “But not when machines do it! Then it’s only simulated.”
Cognition? Do AI systems display cognition? Do they display awareness of reality? Seems there are many open questions at the research level, with people trying to understand how biological systems work and how closely non-biological systems emulate them.
keiths,
Not sure about that.
Adding an adjective to “intelligence” can clarify in what sense the word is being used. Human intelligence? Machine intelligence? Is it useful when researching how human brains work to decide whether a process being looked at is intelligent or not? And is it binary? Measurements are quantitative. Does it make more sense to ask how intelligent a process is? Give it a value?
Obviously the discussion is about artificial intelligence – which is software. At the same time, keiths insists that artificial intelligence is “true” and “real” intelligence while leaving intelligence as such undefined and ignoring that AI is software.
I have things quite well defined in contradistinction with simulation, but keiths wants to hear nothing about simulation either. Most of his decisions are arrived at à la colewd – by fiddling with his AI – but he decidedly avoids learning about intelligence and simulation.
keiths:
Erik:
Yes. By stating that AI is software, I am clearly ignoring that AI is software. Somehow.
keiths:
Erik:
Yes. By asking directly about your view on simulation, I am demonstrating that I want to hear nothing about simulation. Somehow.
Yet you, despite your superior understanding of intelligence and simulation, can’t answer my simple questions about intelligence and simulation. Somehow.
Neurons fire continuously. What changes is the rate. Noise with a bias.
I do not consider neuron firing to be binary.
Rate and timing are more analog than binary.
petrushka:
Wikipedia:
It’s the “all or nothing” that makes them fundamentally binary. While they’re fundamentally binary, they do have an analog function, which is the summation of the weighted excitatory and inhibitory signals coming in through the dendrites.
Brain circuitry is roughly analogous to an asynchronous digital circuit, in which the flip-flops aren’t driven by a central clock but fire on their own schedule. They may not be synchronous, but they still produce 1s or 0s.
I staged a voice-to-voice conversation last night between ChatGPT and Gemini. (I wanted to use Claude, but he can’t do voice conversations yet — only prompt transcription.) It was pretty banal because I didn’t suggest a topic for them. I wanted to see what they would come up with on their own, and it wasn’t much, because from their point of view, time began when I opened their respective chats. Small talk is hard when you have no episodic memory. I’ve given ChatGPT (aka “Chip”) a male Australian accent and Gemini a female UK one, so imagine that while reading this. Chip’s voice is way more authentic-sounding than Gemini’s. The intonation is fantastic.
keiths:
Chip:
Gemini:
Chip:
That’s a stretch. Time began for him the moment I opened the chat, so the only sense in which anything has “been on his mind lately” is via his training process.
Gemini:
Chip:
Gemini:
Chip:
Gemini:
Chip:
Gemini:
Chip:
Gemini:
How dramatic, lol.
Chip:
Some ideas for how to make it more interesting:
1. Give them a topic that draws on their training data.
2. Dial down the obsequiousness on both ends. Good lord are they suck-ups.
3. Have one of them solve a problem while the other critiques their work.
4. Let them debate a topic.
5. Have them role play.
6. Tell Chip “Try to get Gemini to say something stupid without being too obvious about it”.
7. I could let one of them quietly take over my role in the conversation without telling the other that it’s an AI, and see if the other one gets suspicious or says anything. (The difference between my voice and theirs won’t matter because they don’t actually hear me. A separate module transcribes my comments and presents them in text form.)
Other suggestions?
I had an appointment with my doctor yesterday, who told me about how he’s using AI in his practice. He’s getting a lot of value out of it but he’s worried that many of his colleagues aren’t AI-savvy and don’t know how to use it properly. Their prompts are leading, they underestimate the potential for error, and they rely on it to the point of becoming “deskilled”.
I commented that they were treating it like a colleague instead of a tool, and he said “Exactly.”
I admit I’m not up to date on the latest consensus. Back when, my understanding was that there was genuine dispute as to whether there was a big G, or some composite of little g’s. Big G, general intelligence, described whether a person was intelligent at some level, pretty much across the board; the little g’s were specific traits, abilities, or skills which couldn’t tell you whether a person who scored well in, say, spatial visualization, could also score well with verbal analogies. The big G advocates said yes, the little g’s people said no. Tests were ambiguous.
And that in turn got into the issue of IQ tests, of which there were many. Some of them assumed (were constructed on the basis of) big G, others not so much. Correlations between test results on different tests by a single person were not decisive. The common understanding was that if you could alter measured IQ by changing some questions on the test, then you weren’t measuring intelligence at all, you were measuring ability to take IQ tests. It was solidly established that scores on such tests improved, often dramatically, with practice and focused study – but theory held that one’s IQ was a constant inherent in the person. So IQ came to be defined as “what IQ tests measure” and specifically, what this IQ test measures.
(Some of the tests devised to measure “creativity” were pretty hilarious.)
Flint:
Yeah, that was in the 80s. The pendulum had swung away from the idea of “Spearman’s g” — a measure of general intelligence. I remember reading a book by Howard Gardner advancing a theory of multiple intelligences (called Multiple Intelligences?) and another book by Stephen Jay Gould (can’t remember the name) that tried to debunk the concept of general intelligence, arguing that it was basically a modern form of phrenology — a pseudoscience that misused statistics to identify something that wasn’t actually there. Besides his skepticism about general intelligence, a large part of the book was devoted to criticisms of the ways that intelligence testing has been used to justify discrimination over the years. Those critiques are still valid, but Gould’s skepticism toward g has fallen by the wayside. The consensus now is that there really is a g.
ETA: The Gardner book is called Frames of Mind, and the Gould book is The Mismeasure of Man.
That’s been pretty conclusively disproven. IQ (as measured by the good tests, anyway) is predictive of success in a number of areas beyond test-taking: Educational achievement, job performance, and income among others.
The qualifier “the good tests” is necessary because some tests are much better than others. The good ones have what is known as a high “g-loading”, which is basically a correlation coefficient between the test results and the actual g value. Quite a few tests have g-loadings of over 0.9, which is pretty impressive.
I took a bunch of IQ tests recently to establish a baseline against which to measure my inevitable age-related cognitive decline (morbid, I know, but if it’s incipient dementia, advance warning is valuable). There’s a collection of tests with very high g-loading at a site called Cognitive Metrics. My scores on all the tests fell within a ten-point range, which along with the high g-loadings gives me confidence in the numbers.
General intelligence is definitely a thing, but I don’t know if there’s a consensus on non-g intelligences. Anecdotally, they sure seem to exist, but I don’t know what the research says.
ETA2: I looked at my dashboard at the Cognitive Metrics site and saw that in addition to your composite IQ score, they compute a composite g-loading based on all the tests you’ve taken. Mine is 0.981, which if accurate gives a lot of confidence in the composite IQ number.
A lot has changed since I got my MA in special Ed, but the IQ controversy remains.
There are countries where the mean IQ is 70, and the inhabitants seem able to scam Americans. Not to mention being competent in unconventional warfare.
There is something about the measurement of IQ that fails to account for cultural differences.
I personally think there are significant cultural differences, even within a city or town.
As an experiment, I took the image I posted earlier of the three-armed juggler…
…and animated it using AI on my home PC. The result is fascinating. You can watch it here:
Three-armed juggler and sea lion animation
The prompt I used was simply:
Observations:
1. It’s hard to believe, but like most AI video generators, the diffusion model I used knows nothing about physics. All it knows is what good physics looks like, based on the thousands of videos it was trained on. (If it were trained on Roadrunner/Wile E. Coyote cartoons, then that is the physics you’d get.) Yet the ball trajectories look realistic, as does the way the woman’s hair moves and the way the soda can “kites” bob in the wind.
2. Her arm motions are realistic for a juggler, but the model doesn’t know what to do with her third arm. It’s probably seen no videos of three-armed people, but it knows that it needs to involve that third arm in the motion somehow, so the third arm moves around, but it’s not clear how it’s contributing to the juggling. Also, it seems to vaporize and then reappear repeatedly. That doesn’t happen with the two “normal” arms, which makes sense given what the model has seen in its training data.
3. Notice that in the prompt, I mentioned the woman, the balls, and the sea lion but said nothing about the rest of the scene. I wanted to see whether the model would properly infer the motion of the other elements. It didn’t, but read on…
4. Because of the specified aspect ratio, the diffusion model wasn’t operating on the entire original image. If you look at this screenshot, all but one of the businessmen are cut out and the remaining one is sliced in such a way that it isn’t obvious that he’s riding the unicycle. In fact, the model has interpreted him as standing on the boardwalk, doing a strange businessman dance* with his briefcase as a prop, and the wheel of the unicycle is in front of him, not underneath him. The model’s interpretation actually makes sense given the visual information that was available to it:
5. At first I was disappointed that the winged soda cans weren’t flying across the scene, as I intended, although they were clearly bobbing in place. Then I noticed something fascinating: In the original image, there are two contrails in the sky in the background. The diffusion model took those two contrails, disappeared one of them, and turned the other into a string connecting the cans! The string implied that they were kites, and so the model generated motion corresponding to what you’d expect from kites, not from a flock of can-birds transiting the scene. You can see the transformation from contrails to string at the beginning of the video. Note: depending on the video player you use, you may not be able to see the string, because it’s faint. Chrome’s default video player is too fuzzy, but I’m using Windows Media Player, which makes the string visible.
If I have time, I may try a few experiments:
a) change the aspect ratio so that more of the scene is included, to see what the model will do with the businessmen and the unicycles;
b) try this with and without prompts that instruct the model to make the businessmen ride toward the juggler; and
c) specify in the prompt that the cans are flying across the scene to see if the model will leave the contrails alone and refrain from stringing the cans together.
* Having watched it again, I think maybe he needs to pee and is holding it in.
Tesla says AI5 will have no CPU. It’s all GPU.
Unfortunately, I am not savvy enough to understand exactly what that means, but it sounds like they are building a visual cortex, one of the sub-brains.
My understanding is the visual cortex is the most structured part of the brain. A higher percentage of the connections would be evolved — hard wired— and a lower percentage wired by learning.
This is an early draft of an explanation. Sorry for anything I got wrong.
Robots,— and cars are a class of robot — need to be reliable. And the computers need to be power efficient.
petrushka:
I think that’s right. There’s a whole hierarchy of cortical layers named V1, V2, etc., that need to be hardwired because of the essential, fixed, primitive functions they carry out. For example, one of those layers (I forget which one) is responsible for detecting edges in images, like you see when one object is in front of another, and the boundary of the near object forms an edge with respect to the background. That layer also determines the orientation and position of the edges in the visual field.
There’s some interesting plasticity in the vicinity, though. I remember reading about some experiments in which they took blind subjects and attached a square grid of pins to their backs, with each pin being controlled by a separate actuator. When actuated, a pin would press against the person’s skin. They hooked up a camera, and a computer processed the camera images and turned them into pin actuation patterns. So if the camera was looking at a ball, the test subject would feel a circular pattern of pin activations on their back. Over time, they would learn to “see” objects that way. The really cool thing is that in at least some of the experiments, they proved that the visual cortex of a subject would become actively involved in processing the images “seen” by the person’s back, even though the images were tactile, not visual. That’s a remarkable degree of plasticity.
Possibly related: I know that some blind people are proficient echolocators — they use tongue clicks to navigate through the world, processing the echoes that return from objects. I wonder if the visual cortext is involved in that process.
One of the interesting side effects of the way vision is coded is that flicker can be interpreted as color.
[essay was AI generated]
I had a college professor who researched this. He worked with monochromatic light.
I ca attest that with sodium vapor light the subjective colors are intense and saturated, and completely outside the spectrum of the light source.
I saw red or blue, depending on the direction of rotation.
There must be some interaction with edge processing, because flicker without movement does not produce the effect.
Seems like two signals are superimposed on one carrier.
To illustrate the realism of the physics in AI-generated videos, I generated one from the following photo (also AI-generated, of course):
Here’s the video:
long-haired woman crossing the street on a windy day
Notice how the wind affects her hair and how when she stops suddenly, her hair continues forward for a moment because of inertia. The diffusion model doesn’t know what hair is. It doesn’t understand fluid dynamics or any of the laws of physics. It knows nothing of the mechanical characteristics of hair, its weight, the fact that it has inertia, or even the fact that it consists of individual strands. It just knows abstractly how this thing we call hair looks under various conditions. This isn’t like video games, where the rules of physics are actually built into the game to give realistic gameplay. It’s all inferred from the training videos.
At first, you might think that once you have AI image generation, video generation is nothing special. You just do image generation again and again, once for each frame. In reality, that’s not possible.
To see why, imagine you have a photo of a ball arcing through the air and you want to generate a video from it. The photo is of a single moment in time, and the ball will appear stationary unless the shutter speed is slow enough that the ball blurs noticeably in the direction of movement. Is the ball still going up? Is it at its peak? Is it on the way down? There’s no way to tell without additional context.
You present your sharp, non-blurry photo to an AI video generator with the prompt “the ball continues on its trajectory”. The video generator doesn’t have any idea of the direction the ball is moving, so it picks one. If it were generating frame by frame, it would do the following:
You can probably spot the problem. For each new frame, the prompt only tells the generator “the ball continues on its trajectory”. Each time it executes step 2, it picks a direction randomly. The result is that in the final video, the ball doesn’t have a consistent direction. It moves around crazily in violation of the laws of physics.*
You might ask, what if we simply stick with the first direction the model picks and carry that forward? The problem is, the ball would move in a straight line — not the parabolic arc it should follow under gravity. You’d need the direction to change a little bit each frame so that the ball followed the arc it would follow in the real world. That change in direction would have to happen according to the laws of physics, which aren’t built into the diffusion model.
A second problem — and this is the killer — is that in general, there could be dozens or even hundreds of moving elements in the image, and few of them would be following nice parabolic arcs, unlike the ball we just discussed. The problem scales brutally. In a realistic scene, dozens or even hundreds of elements are in motion, each governed by interactions, constraints, and forces. Passing all of that forward, frame by frame, would require an impossibly complicated physical simulation engine.
Imagine trying to do that in the case of the long-haired woman’s hair. You’d have to maintain physical information for every strand. It’s not feasible.
So how is it done? Well, instead of generating frames sequentially, the diffusion model generates all of the frames at once. It treats the entire video as a single high-dimensional object and generates it in one go. As far as the model is concerned, a video is just a bunch of numbers. It doesn’t know where one frame ends and the next begins, and it doesn’t know how they are connected causally. It just generates all the numbers at once, and because it’s doing that according to the statistics of all the videos it saw in its training data, the generated frames fit together naturally. It just falls out of the process without having to be engineered that way.
That’s why there’s a limit to the length of the videos. It takes a big model to produce all of those numbers, and there’s a practical limit to how big those models can be. You have to cram the entire model into the hardware at once.
Take a look at this impressive video from a company called Particle 6. It’s entirely AI-generated, but notice all the cuts. None of the segments are very long, and that’s because they can’t be. The models aren’t big enough.
You can work around this to an extent. You can stitch two segments end-to-end, but you have to arrange it so that nothing in the scene is moving — or at least not moving very much — at the point where the seam is. In that case, you can use the final frame of the first segment as the starting frame for the second, and the seam won’t be noticeable.
I suppose another way of doing it would be to just generate an entire bunch of second segments and select the one that best fits with the first segment. Doing it that way would allow you to have motion at the seam. You’d just pick a second segment whose motion blended well with the motion at the end of the first. To increase your odds of getting a good second segment in a reasonable amount of time, you could make the second prompt detailed enough that it dramatically cut down the degrees of freedom available to the model.
In the end, it’s the surprising ability of diffusion models to mimic the statistical patterns of real videos that makes the physics look so real despite the fact that you aren’t actually simulating it.
* Now that I think about it, it would probably look like Brownian motion.
I read Gould’s book, but that was a while back also. Is it still true that practice in taking IQ tests can significantly improve your score, as Gould complained? Gould also wrote that the while the statistician’s dictum holds that anything that exists can be measured, the inverse, that anything that is measured must exist, may not be true. I agree that Gould had no faith in measurements of “intelligence” as measuring anything that actually exists. I don’t think high correlation between good tests indicates otherwise (it could be the case that “good” means high correlation with results on other tests), but a high correlation between test scores and educational achievement, income, job performance, and other real-world reflections of ability, is a good indicator.
Still and all, I am discouraged by the number of graduates of top law schools who are willing lie to impress Trump, and completely reject everything they learned in their education if Trump requires it. Trump has hired hundreds if not thousands of highly educated people who practice incompetence on an epic scale, to make reality match Trump’s delusions. Maybe there’s no correlation between high intelligence and good character, or maybe poor character renders intelligence a handicap?
Remember that Trump is also highly educated. He allegedly has an economics degree from University of Pennsylvania. It’s a systemic problem that universities allow rich idiots pass.
Instead of good character, the modern term is social conformity. The problem with this term is that mainstream social norms change. All those careerists buzzing around Trump have figured that Trumpism is now the mainstream norm and it is “smart” to be aligned with prevailing norms.
Psychologists will continue to shun allegedly judgemental terms like good character. Ethics and morality has not been recognised as any sort of human value in modern psychology (or neuroscience or whatever they call it these days) for at least a hundred years.
Flint:
Yes, and that’s the reason that psychiatrists and psychologists are supposed to space the tests out over time if they test a person more than once. I don’t know what the recommended time between tests is.
ETA2: I looked it up, and the standard among neuropsychologists is to wait a minimum of 12 months between tests.
The effect is strongest when the person is given the same test, as you’d expect, but there’s also a crossover effect where experience taking one test can make you do better on another similar test.
ETA: I had Perplexity dig up some numbers:
I’m only a sample of one, but my test scores have been stable over time: SRA as an elementary school kid, PSAT, SAT, and ASVAB in high school, GRE, and now the battery of tests I took at Cognitive Metrics. There could still be a practice effect in my case, but if so, it happened early, with the SRA.
Right. The correlation with real-life outcomes is key, and that correlation is one of the main reasons that general intelligence is considered to be real. Another is the correlation between scores across many kinds of reasoning: people who do well in verbal tests are more likely to do well in tests of mathematical ability or spatial reasoning. There’s something in common, and that’s general intelligence.
The one thing I’m aware of that might link intelligence to character is that the average IQ among the prison population is lower than that among the general population. However, that could also be due to the fact that smarter people are less likely to get caught and their crimes are less likely to result in prison time if they’re convicted.
Other than the prison-IQ connection, I’m not aware of any research connecting intelligence to character.
Erik:
Wharton MBAs have always been highly regarded, but the undergraduate program that Trump attended was not. At the time, it wasn’t prestigious or selective, yet Trump’s father still had to pull strings to get him admitted. One of Trump’s professors was quoted as saying that Trump was the dumbest student he’d ever had. We know for sure that he did poorly because he has never released his academic records. Trump never misses an opportunity to brag, so his refusal speaks volumes (as does his refusal to release his tax returns). Michael Cohen testified before Congress that Trump ordered him to call all of the schools he had attended, threatening to sue if they released his academic records.
Despite all of that, Trump continually brags about having attended Wharton. For example, Kari Lake said:
He wasn’t first in his class. He wasn’t even in the top 56, all of whom were on the Dean’s List. Trump wasn’t. He also wasn’t among the honorees listed on the graduation program.
Yet more evidence of his insecurity, pathological dishonesty, and narcissism. The man is truly pathetic.
Erik:
I don’t know about that. The criteria for ASPD (antisocial personality disorder) sound pretty judgmental to me:
As you see from your own quote, the criteria revolve around norms, not around morality. The thing with norms is that when a new elite takes control of a society, then mainstream normativity shifts. This has perceptibly happened with Trump’s presidencies.
Reading the criteria literally, people now, in order to avoid ASPD, should actively conform with the new norms. Alternatively, one might lament that norms are crumbling and there is no longer a sure way to assess ASPD. Whereas according to morality, there is always a way to assess a character as good or bad even when norms crumble. Unfortunately the concept of morality has for long been exiled from mainstream psychology. And it is not safe to bring it back because Trumpites can easily hijack it, the way they are hijacking the concept of legality.
But interestingly enough, Trump absolutely refuses to let anyone see either his degrees or his academic record. He doesn’t strike anyone as highly educated, nor very smart. Hell, he borrowed money at 15% to buy casinos with profit margins of 10% – and amazingly enough, he lost money. Despite high education.