Note that Weizenbaum was an AI critic: Weizenbaum's intention was not for Eliza to pass the Turing test, but to show to people that a clearly not intelligent program based on primitive pattern matching can appear to behave intelligently.
He failed: His own secretary wanted to be left alone with the software and typed in her personal problems. Work on Eliza (1963-65, paper published 1966) until today is mostly misunderstood.
Weizenbaum even wrote an entire book, Computer Power and Human Reason: From Judgment To Calculation, on how AI is a crank term for computer science. The basic premise is that humans are not machine so stop using language that treats them as such. Especially that computers can decide what comes next, but only a human can choose what to do.
The book also has one of the best and most succinct descriptions of Turing machines and the theoretical underpinning of computer science that I have ever read. Even if you’re an AI maximalist you should read the third chapter of the book.
Aren't we? Casual chains upon our matter produce emergent behaviors using the same physics and chemistry that our mechanistic creations rely upon.
Certainly those behaviors and results do not produce the same repeatable, predictable results as our clockworks but that is the whole point of the field of AI (as opposed to the marketing corruption/term that is currently in vogue, so GAI if you prefer), to produce system and algorithm structures designed with architecture and patterns more like our own.
Perhaps you believe in the ghost in the machine hypothesis? The magical soul that is more than the emergent evolving pattern produced across time by DNA replicators? That this undefinable, unmeasurable spirit makes us forever different?
This is an hypothesis. Of course, it is obvious that a lot of what human/animals are and do can be explained or described by physical processes. But there is the "hard problem of consciousness", to which no one has a satisfying answer.
I feel sarcasm in your last paragraph, but it feels rather dishonest, in the sense that you voluntarily use rather crude and ridiculous formulations, as if it was the only alternative. But there is some arrogance in thinking that we, smart as we are, finally got the final answer about one of the hardest questions in philosophy and metaphysics. Materialism is not proven: it is a basic methodological assumption of modern science - meaning that we do not have the tools to either prove or disprove it. Newton, Gödel and lots of other renowed scientists are knowed to have opposed materialism. Let's accept that the question is open.
On this topic, I always recommend reading the report of the Galileo Commission, which is a manifesto by a wide range of scientists and philosophers to reduce the stigma associated with even questioning this fundamental dogma.
Just because there are two hypotheses doesn't mean that we should assign them equal probability. Why should we believe that we will need to modify the laws of physics to explain consciousness when they already do a remarkably good job of predicting and explaining everything else we know about chemistry and biology? If someone has an alternative theory of consciousness that goes beyond the standard model of particle physics, the burden of proof is on them to produce the equations that explain how this new substance or force interacts with the known fields of the standard model. Of course we should never put 0 probability on anything, but I know where I would put my money given the empirical track record of our current theories.
I actually think it is a bit more subtle than that. I am here really focusing on consciousness as a subjective experience, which seems to be of an entirely different nature than physical processes. We have a lot of examples of emergent properties of complex systems, but as far as I can tell all of them stay in the same "realm": the emergent properties of an ant colony, for instance, are physical, resulting from physical interactions.
With the subjective experience as an emergent property, things such as a society, a country, an economy could become conscious, in the sense of having a _subjective_ experience of their own existence as entities separate from the rest of the world. If we accept the "consciousness as an emergent property", we _have_ to accept that possibility. Which, to me, is not less wild or unlikely than, say, the "theory" of a field of consciousness "received" or "captured" by physical systems with certain properties, the same way a radio can receive radio programs. There are additional reasons to want to consider alternative explanations, but going into them would rrquire much more space - if interested, I would point to the report of the Galileo Commission.
So it does not change anything to physics, really: materialism is pretty much the best methodology to unpack the laws of physics: whatever you observe, see if you can find more elementary physical processes that explain it.
I am just a bit irritated by bold statements which assume that we know for certain that consciousness is an emergent property of physical processes. We do not, and the reason why this is such an accepted fact is more sociological than scientific - Newton and others decided to focus solely on physical processes as a methodological tool, and over centuries, the undeniable success of the approach in making discoveries _and_ building practical tools gave it an ontological status it did not have initially (Newton was for instance a very convinced Christian). Which makes me keen to remind that, because in the current scientific culture it is shameful to even ask the question.
I'm not inside the scientific community so I can't verify the degree to which it's a shameful question. My guess is that it has a lot more to do with simple heuristics than dogmatism. A researcher has finite time and resources. They need to decide what to work on based on the likelihood that they can make progress on certain well-defined problems. We already have a centuries-long track record of making progress by studying things in terms of physical and chemical processes. That doesn't mean this approach can solve every problem, but there's not much else we can do until a new Einstein comes along and proposes an alternative that's compelling enough. I believe there are a lot of young scientists who would be willing to jump on a new paradigm if it was obviously leading to novel insights and breakthroughs, but that hasn't happened. It's not because these ideas are being suppressed, it's because nobody has put them forward in a rigorous and convincing manner.
I can't say I find the Galileo Commission to be particularly compelling. It is authored by people who are inclined toward non-materialism, and I get the sense that they care more about consciousness existing beyond the physical brain than discovering what is true. This isn't to wholly dismiss it, there is good work and smart people behind it, but I think it has a pretty heavy agenda.
And, this may be petty, but the self-comparison to the maligned Galileo is rather off-putting.
You've written your response thoughtfully, thank you.
Of course my description is of a hypothesis rather than ground truth.
It turns out that there isn't consensus over whether "The Hard Problem Of Consciousness" is actually a problem[0] but I need to admit I didn't know that before reading your comment. You may have been too kind in your assumption of my competence level on the subject. As adjunct, my writing was about my own currently highest probability expectations rather than a mature reporting of the state of the field of experts. It did not seem worth (then or now) writing to such a standard on a public discussion forum, even this one.
My last paragraph wasn't written with sarcasm (sorry to have given easy opportunity to read that in it) but it was a bit rushed, clumsily written, and unimaginative. Poorly written as it was, my curiosity was sincere (its been some time since I engaged with the state of the art) as was some of my dismissal. I have found it common that people have a belief in the uniqueness of humans as a platform for intelligence that frequently has its roots in sentiment and/or unexamined beliefs. Of course there are individuals whose thoughts meet a much higher bar. It is also the case that faith as a basis of belief does not invalidate the belief.
There is so much we don't know. We only recently explained how aspirin works[1]. That didn't stop it from working well during our long period of usage prior to our understanding. I'm comfortable with that an its analog here, that I don't understand consciousness and its mechanisms completely. I seem to experience it and it seems to result from what is, despite my incomplete knowledge of what that "what is" comprises. However, there is an imbalance of evidence for the material hypothesis and it seems plausible that emergent dynamics are sufficient to explain.
So... Yes, materialism is not proven and yet I currently hold it as the explanation that makes at least a partial contribution to the more complete truth. Further, that it is the explanation with the greatest volume of evidence and support. I suspect that it is sufficient for the emergence of intelligence and even consciousness in ourselves and as such sufficient for (through the same mechanisms) the emergence of intelligence and consciousness in our artificial constructs. Note that I also suspect we are still some distance from that inflection point.
Thank you for the reference to the Galileo Commission. I had not heard of it and am always happy to consider new perspectives and challenge those which I have held.
> Weizenbaum makes the crucial distinction between deciding and choosing. Deciding is a computational activity, something that can ultimately be programmed. It is the capacity to choose that ultimately makes one a human being. Choice, however, is the product of judgment, not calculation. Comprehensive human judgment is able to include non-mathematical factors such as emotions. Judgment can compare apples and oranges, and can do so without quantifying each fruit type and then reductively quantifying each to factors necessary for mathematical comparison.
Okay, so what is judgement? I haven't read that particular book and I don't quite remember his argument from interviews and lectures I saw, so this might be wrong, but I'd say it's for example saying "this is fair" when you measure the slices of pie you cut a cake into. That is, calculating that they're of equal size is pure computation; but there is no way to compute that when sharing cake with your friends, the slices should be equal.
Just like you can compute how much clean drinking water an average or specific person needs a day, with at least some accuracy, but when it comes to the question "should there be life in the universe" or "should people die of thirst", no computation could answer it. You could choose to write a program that decides it based on a random seed or a super complex algorithm taking a billion factors into account, but and then that program would decide the question, but it's essentially still something you did / chose.
It's basically a religious view. For a "judgement" to be non-computable, it'd need to come from some factor in the human brain which violates know physics and can't be reproduced outside a human brain.
It's little more than arguing for a "soul" with no evidence for any effect that can't be explained by cause and effect.
> For a "judgement" to be non-computable, it'd need to come from some factor in the human brain which violates know physics and can't be reproduced outside a human brain.
You say this as if we are even close to understanding much less reproducing the human brain completely, which probably would have to include the web of relations with all sorts of other living things that also go into the judgements we make, and the emotions we have while making them. Until you actually do draw the rest of the owl, it's not exactly "religious" to say there's no owl.
You cannot compute what you don't understand, and even if you did by accident, you wouldn't know you computed it, as long as you don't understand what you're trying to do. That seems obvious to me.
And "computable" and "computable for us" are very different things. It's not about the machines or algorithms we might make one day, provided that we fully understand everything that goes into our our thoughts and emotions with nothing left unaccounted for, and everything turning out to be countable; it's about the ones we are actually making, back then and today, and then in some cases outsource our decisions to.
You're misunderstanding the terms. For something to be computable is very different from whether or not we know or are presently able to compute it.
For something to be computable, it only needs to be possible to show that it is logically possible, by e.g. decomposing the problem into elements we know are computable or showing an example.
The existence of the human brain absent any evidence of any supernatural element is strong evidence that human reasoning is computable, and it's a reasonable, testable, falsifiable hypothesis to make: If you want to counter it "all" you need to do is to show evidence of any state transition in even a single brain that does not follow known laws of physics. Just one.
Alternatively, even just coherently describing a decision-making process that it is possible to construct a proof wouldn't be computable using known logic.
Either would get you a Nobel Prize, in either physics or maths. Absent that, even just a testable hypothesis that if proven would increase the likelihood of finding either of the above would be a huge step.
In the absence of all of that, it's pure faith to presume human reasoning isn't computable.
You are talking about "comprehensive calculation indistinguishable from human judgment with the ability to include factors such as emotions". Even if that might be possible once we can fully compute all of physics (that in itself I wouldn't assume), that's just not what we're actually dealing with.
Am I ever doing things because I can and not because I have to? Also, what mechanism determines what things I want to do because I can do them? And isn’t that mechanism then just not just another part of the machine.
Just because it feels as though I do things because I can doesn’t mean that is actually true.
As long as you can imagine different possible futures and decide upon which one you want to try and realize, I think you have choice.
Choice stems from uncertainty, partial knowledge. It might be an illusion for an observer outside of the system, but as far as a participant within the system is concerned, there is choice, then there is free will.
I am writing this because I ca n but I don't need to do it. I have futures where I don't do that and do something more rewarding instead and still. As long as I am aware of the choices, then I have free will.
Traditionally, a program is a series of instructions. The program doesn't really act on its own.
Now, a program which is objective driven and can infer from new inputs might be something else.
Just like humans try to maximize the stability of their structures via a reward system. (it got slighty complex, faulty at times, or the tradeoff between work vs reward is not always in favor of work because we do not control every variable, hence procrastination for example, or addiction which is not a conscious process but neuro-chemically induced).
But what does "act on its own" mean? If I give the program some randomness over its next action, is that "acting on its own"? When I'm at work, I act according to a series of instructions. Am I not acting on my own?
This is a very philosophical discussion, but if I had an infinitely-powerful computer and could simulate an entire universe based on a series of instructions (physical laws), would the beings in that universe that created societies not be "acting on their own"?
Yes, as long as the computer chooses its next set of instructions in order to maximize a given value (the objective), I would say that it acts on its own. Instruction set that was never defined by anyone that is.
If the instruction set is limited and defined by someone else, I believe it doesn't.
I think, re. the simulated universe, that for us, they wouldn't have free will because we know causality (as long as we are all knowing about the simulation). But as far as they would be concerned, wouldn't they have free will if they know that they don't know everything and whether the future they imagine is realizable?
If they knew with certainty that something was not realizable, they wouldn't bother, but since they don't know, either they try to realize a given future or they don't.
Partial information provides choice of action, therefore free will.
Computable by who?
Because you don't have the full list of correlations and there are superlinear things (tail events) you'll get a probabilistic estimation at best.
Just because we have quantum RNG in our heads that doesn't make us automatically better. If anything it makes us worse since we don't act on reason alone.
I don't know if there is a quantum rng or just an inference machine that manages to recognize patterns within input data and can do some math sometimes.
A girl would like to ask a boy to the high school dance.
A computer can do all the calculations to decide on if it's a good idea. Given the inputs of the time they have spent together, the number of glances that are passed between then in the halls between classes, if he doesn't have a date yet or not, etc. The probability adds up to ask.
So the machine decides to ask.
The girl feels it. Has all the time they've spent together has made her feel a certain way? Maybe a weird tingle each time their arms touch. Is that glance in the hall this morning not just an accident, but him going a little out of his way for her to notice? She's asked around and knows that no one else has asked him, but doe he really not have a date yet? Can she overcome the bit of anxiety about asking a boy to the dance? Will she be able to accept the risk of rejection knowing that the chances may be high he says yes?
All the tingles, feelings, anxieties and hesitations are activities triggered by little programs that work autonomously and are fully deterministic.
The girl is fooled
Even if you accept a strong determinist position, there is still a distinction:
The determining factors driving a computer program can be fully quanitified; the sets of inputs and conditions is finite, can be reasoned over, and described fully.
That's basically the fundamental description of computing, in fact, and what makes a Turing machine.
The determining factors "IRL" are effectively infinite, a causal "chain" of infinite (or near infinite) complexity that expands backwards to the Big Bang, (or whatever) and sideways around the planet and beyond. There is no catalog you could make of all the "causes" that could isolate things enough to truly reason over and describe them all.
And so, yeah, to say it's all just "little programs" is the most ridiculous reductionism, that basically purposefully neglects to see the complexity and depth of the world around us.
I personally take a strongly determinist, materialist philosophical position. But I would never ever express that in terms of "programs" or anything similar.
This post of yours sent me looking for this and I ended up with (now reading) Agassi’s provocative (& brutal) takedown of this book. As to the content of the book, Agassi pointedly mentioned Weiner. I will of course read Weizenbaum after this. (thanks.)
All this “AI” hype is a constant reminder to me that you cannot reason anybody out of something that they want to believe. People's need to believe in miracles is obviously stronger than all reason.
There is a ton of arrogance everywhere on this subject. Anyone who doesn't see how complicated and nuanced and difficult figuring out what the hell is going on here really needs to sit down and think a little more carefully. There's absolutely no room for out-of-hand dismissal. We are all babes in the woods right now. We need each other, like always.
Take a look at this project that has rediscovered and recreated the original ELIZA source code. You can even try out an accurate reimplementation of ELIZA
Many years ago I spent a lot of time on a website called Personality Forge, where you would create chat bots very similar to this and set them loose. They could then talk to other people or each other and you could review the transcripts. At one point I even entered my chat bot in The Chatterbox Challenge. It was so much fun to work on this thing at the time, but when I rediscovered it years later[0] I was mostly just disappointed in how "fake" all that effort was.
Now here I am talking about life, the universe, and everything with ChatGPT. It makes me both inexplicably happy/hopeful and simultaneously weirdly melancholic.
The great thing about being a teenager or kid is that you don’t know why the grownups don’t think your project is worth doing, so you just do it. Even if it doesn’t change the world (most things don’t, after all) you can still learn something and have fun.
Eliza's meant to us to be an illustration of the problem. In good old fashioned AI sentiment, it illustrates the fact that you need another if statement for every new kind of construct you want to simulate. But you deign to simulate each thing, like say turning a verb into a gerund, by writing a specific "gerundification" routine. Another to swap the Mes to Yous, etc. this isn't how people think nor most modern AI. This is totally different from just looking at the world distilling patterns from it and using those patterns as the basis for a response. To teach a modern AI new stuff, you don't have to write another if statement. They have similar intentions and at some resolution or distance, they are trying to do similar things. However, they work and totally different ways and the new dynamic generative AI strategy that learns from input as opposed to just symbolically transforming it syntactically is a totally different paradigm. I don't care whether you call it AI or not.
With the small reservation that this is not how Eliza works. Eliza sits on top of MAD/SLIP which does all the heavy work and provides lists and integer indexes, which is what is processed by Eliza. This allows Eliza to work on decomposition rules, which isolate keywords per position and context, and transformation (composition) rules to recombine elements and links between those two. Meaning, the model is much more topological than this. (Arguably, this is closer to regular expressions than to if-else trees.)
However, this isn't what Eliza is all about. It's rather about the question, how little do you actually need in terms of knowledge, world model, or rule sets to give the impression of an "intelligent" (even sympathetic) participant in a conversation (as long as you're able to constrain the conversation to a setting, which doesn't require any world knowledge, at all.) To a certain degree, it is also about how eager we are to overestimate the capabilities of such a partner in conversation, as soon as some criteria seem to be satisfied. Which is arguably still relevant today.
>how eager we are to overestimate the capabilities of such a partner in conversation, as soon as some criteria seem to be satisfied. Which is arguably still relevant today.
Honestly, AI shouldn't be the takeaway point here, but how we do the same for politics.
On a not so serious note, most political careers seem to be built on public utterances that seem to be generated by a rule set that could fit into 3 pages, triggered by a handful of keywords or trigger phrases, also known as talking points. With the advent of the so-called culture wars, most of this is also increasingly context-free and doesn't require much of world knowledge. Users, err, voters will fill in the gaps eagerly, each according to their respective phantasies and understanding. To the point that Eliza may eventually become a worthy contestant. An approval rate of 27% is certainly a good starting point…
> Interestingly, Eliza still outperforms ChatGPT in certain Turing test variations.
I see we have a new entry for the 2024 Lies of Omission award.
The article linked to plainly shows that Eliza only beats ChatpGPT-3.5 and is in the bottom half when ranked against a variety of different system prompts. An excellent ass covering strategy that relies on the reader not checking sources.
An honest author would have actually quoted the article saying:
> GPT-4 achieved a success rate of 41 percent, second only to actual humans.
instead of constructing a deliberately misleading paraphrase.
"GPT-4 achieved a success rate of 41 percent, second only to actual humans" also feels like a (much bigger) lie of omission looking at the original paper. GPT4's performance was in the range of 6% to 41%, Eliza's 27% score sat in the upper middle of that range, and considering the bots tested consisted of 8 GPT4 prompts, 2 GPT3.5 prompts and a naive script from the 1960s, GPT4 would have had to be remarkably consistently inhuman not to finish "second only to humans" with its highest scoring prompt
The blog appears to have been updated to specify GPT3.5, but the original version was accurate.
The paper itself is interesting as it covers the limitations (it has big methodological issues), how the GPT prompts attempted to overridei default chatGPT tone and reasons why ELIZA performed surprisingly well (some thought it was so uncooperative, it must be human!)
https://arxiv.org/pdf/2310.20216.pdf
The example ELIZA responses in the paper are so laughably bad and trivial to pick up, I'm not convinced the human interrogators were sober/conscious/awake during the experiment.
tbf the human side of those conversations isn't much better. I think if someone tried prompt injection hacks on me I'd be tempted to be politely obtuse to troll them right back.
Turing's version involves experts who definitely aren't in the same room waving to each other, but the fundamental problem is it isn't a particularly good test
That's actually the critical part and much more relevant than the exact scores achieved.
According to the linked article, the main reason Eliza got such a pass is because the testers were looking for a ChatGPT-esque giveaway. Long-winded, frightened of controversy, prone to hallucinations.
Which Eliza is not. And (presumably) not being familiar with Eliza, they thought it was another bored human test subject putting in the absolutely bare minimum effort.
Eliza didn't pass for a human - it passed for not-a-LLM.
Yeah GPT4 is not trained to beat turing test, it is trained to be an AI assistant.
Imagine you take a human and train them to be an AI assistant since they were a baby. I imagine their behaviour would also be very odd compared to average people.
Unfortunately, the paper only provides the text of one of the prompts they used (Juliet) and it happens to be one of the worst performing ones that scored lower than ELIZA. I suppose you could qualify the quote by saying that the best prompt used with GPT-4 had a 41 percent success rate. I don't think that's more of an omission than excluding the GPT model, excluding the prompt used, and ignoring the fact that other GPT models beat ELIZA.
Hum, note that this was not an argument about or against GPT, but about the "unreasonable" success of a, by all standards, primitive algorithm that manages to get (somewhat) away by crafting the pre- and context of the conversation. By no means, on the other hand, could I read this article and understand it as claiming any superiority over modern applications.
(Nobody with even the crudest understanding of the principles of Eliza could claim this, and the article clearly demonstrates a detailed understanding. Disclaimer: I wrote the JS implementation linked in the article, many years ago.)
Edit: The question rightfully raised – and answered – by Eliza, which is still relevant today in the context of GPT, is: does the appearance of intelligent conversation (necessarily) hint at the presence of a world model in any rudimentary form?
> One of the first computer programs that successfully passed the Turing test was Eliza. Created in 1966 by Joseph Weizenbaum, Eliza skillfully emulated the speech patterns of a psychotherapist in its conversations.
Why was the Turing test still relevant after this? Didn't this indicate it was very flawed test? Or it was hard to come up with a better test?
I can find no reference of an actual Turing test being done for Eliza. If you look at the link from the article it is clearly demonstrably failing their (different, and more difficult to interpret, but still fair I thiiiink) runs today as well. Note that people constantlywillfully misinterpret what a turing test is.
A turing test means you enter into two conversations. Then you pick which one was with a computer. If people answer wrong 50% of the time, the computer is indistinguishable, hence it passes. Note that it is not "People get wrong whether their single conversation is talking to an AI >50% of the time" and it is definitely not "sometimes people don't realize they're talking to an AI". In particular people constantly write about the latter because it generates clicks.
The variant of the Turing test that Eliza has passed is the easiest one. That is: convince a novice interrogator that what is on the other side of the screen is a human.
The real turing test (the imitation game) involve a computer and a human subject, both talking to a human interrogator. The interrogator must determine who is the human. It is an adverserial situation, both the human subject and interrogator do everything in their power for the interrogator to correctly identify the computer. The computer has to not only convince the interrogator, but also do it better than the human subject. Furthermore, both humans are supposed to be experts in the game, just like the computer that is designed to pass the test. So, not just random people.
He originally made the argument about gender, not intelligence. I think he was arguing for a whole class of properties for which there's no difference between authenticity and convincing fakery.
I think the point is less that there is a truth and we're too dumb to figure it out, and more that in certain circumstances we'll just have to accept a lower bar for evidence about whether those properties apply.
It reminds me of how no class of computer can solve the halting problem for itself. No matter how intelligent you are, there will be holes like this in your certainty about some things.
Or another way to put this, it's not a binary problem, it's a probability continuum.
Even the definition of 'human intelligence' is a continuum from the smartest to the dumbest of us, that doesn't even stop there and descends thought all animal life.
I did some research to prove you wrong, because I don't think continuum is the right concept, but it turns out that Turing seems to agree with you. Quoting him in "Computing Machinery and Intelligence":
> In short, then, there might be men cleverer than any given machine, but then again there might be other machines cleverer again, and so on.
So now I think you're both wrong :) Particularly I take issue with the assumption that the "cleverer" relation is transitive. We've only really studied a few relations in this space:
- pushdown automatons are cleverer than finite state machines
- turing machines are cleverer than pushdown automata
- humans are cleverer than turing machines (I'd argue for this, but others would disagree)
Presumably there are other points which we have overlooked or not yet discovered. For instance, maybe something which has the "memory" quality of a pushdown automaton, but lacks the "state tracking" property of a finite state machine. When compared with an FSM, such a thing would not be more or less clever than it, it would just be clever in an orthogonal way.
I strongly suspect that two intelligences (of greater power than the theoretical machines that we yet have) could meet and discover that they each have a capability that the other lacks. This would be a situation that you couldn't map onto a continuum--you'd need something with branches, a tree or a dag or a topological space: something on which the two intelligences could be considered cousins: neither possessing more capabilities than the other, but each possessing different capabilities. (Unlike the FSM example, they would have to share some capabilities, otherwise they couldn't recognize each other as intelligent).
Further, I suspect that in order to adequately classify both intelligences as cousins, you'd have to be cleverer than both. Each of the cousin-intelligences would be able to prove among themselves that theirs is the superior kind, but they'd also have to doubt these proofs because the unfamiliar intelligence would be capable of spooky things which the familiar intelligence is not.
I mean an evolutionary tree where intelligence features are added in some branches makes sense.
I guess part of what I was trying to address is that we like to think of intelligence as what people do and are the pinnacle of, and discounting anything that is not covered by that.
I definitely agree that defining intelligence as what humans do is a problematic practice. I guess I just wanted to nit pick a little.
There's definitely a lot of "it's not real intelligence because it's not human intelligence" going around these days. Doesn't seem like it's going anywhere useful though.
To me, Turing's argument had always been that the attribution of intelligence (or not) doesn't make much sense: rather than being a question of any substance, it merely diffuses into a matter of appearance. However, as things usually are, it had become the holy grail for claiming "intelligence" (which really should be used in this context in quotes only).
In actuality, it primarily tests the knowledge of the user, not their intelligence. I'm sure that GPT-written paragraphs seemed really impressive when you first encountered them, but nowadays half the Internet has seen enough of them to recognise the default ChatGPT style in less than the space of a Tweet. People aren't significantly smarter than they were a few years ago, but I bet GPT-3.0 will perform significantly worse on a Turing test now than it did the day it was released.
Similarly, I believe a lot of early Turing Test successes kind-of cheated and had their bots pretend to be ESL, on the grounds that the interrogators would interpret their unnatural responses not as a robot but as a second-language speaker's human mistakes. But people who teach English as a second language, or interact with language-learners a lot will learn the types of mistakes each group of learners make, and will spot unnatural mistakes a lot faster.
Now that I think about it, that a major factor in determining Turing test performance isn't the intelligence of the testers but their knowledge does highlight why it's not a great measure of intelligence in the first place.
I've often felt that a better version is not whether a person can guess that it's AI or a human, but whether people behave and feel differently with an AI or human.
That's vague and covers a universe of criteria — mood, satisfaction with the conversation, actual behavior and so forth — but it also I think is a more realistic gauge of AI performance. It's probably unattainable but that's not necessarily a bad thing. If it is attainable within confidence then it's a pretty powerful AI.
There are probably some people who would be ok with some AI for some purposes.
In a sense, the question of the "intelligent machine" is somewhat self-contradictory: To us, the question of intelligence matters as a preposition or qualifying term, for to what extent, probability and prospects we may pose an appeal to sympathy, moral and ethics. (In other words, it is not about trust in any realistic faculties, but about judgement – and then, to what extent we may trust in this.) However, this prospect doesn't fit well our expectations towards machines, which are all about repeatability and reproducible results in given tolerances… (Compare ChatGPT's so-called winter depression and the arising need to plead and argue with the device for any complex results. As the device gains in the emotional domain, its worth in the application domain radically decreases.)
I remember this being written in basic for the c64. Not sure if I had the real Eliza or a clone. But it was fun to look at all of the canned responses and try to get it to respond with them.
Seems like the modern definition is something along the lines of an algorithm whose behavior depends on data which was itself machine generated, rather than hand-created by a human like Eliza's rules.
I'm stoked to be learning about the real Eliza, after being surreptitiously exposed via the Zachtronics game [0]. The game has a fascinating dystopian/scifi take where a company provides AI therapy through human "proxies" that simply vocalize the AI response.
Then later on there was Azile[1], the evil twin to Eliza. Instead of being trained with data to make it sound helpful and supportive, it uses hostile and insulting language.
We had it running in terminals when I used to work in the national museum of computing in the UK (on machines where you can just pull the full source up from floppy disk)
>I didn't write the original ELIZA program, although my Lisp class was taught by Joseph Weizenbaum, who did. I later wrote a very elaborate program of similar kind, which I just called DOCTOR, in order to play with some of the ideas.
>At some point, I noticed there was a program at Stanford called PARRY (the paranoid patient), by Kenneth Colby. I understand from Wikipedia's PARRY entry that Weizenbaum's ELIZA and PARRY were connected at one point, although I never saw that. I never linked PARRY with my DOCTOR directly, but I did once do it indirectly through a manual typist. Part of my record of this exchange was garbled, but this is a partial transcript, picking up in the middle. Mostly it just shows PARRY was a better patient than my DOCTOR program was a doctor.
>I have done light editing to remove the typos we made (rubbed out characters were echoed back in square brackets).
>Also, I couldn't find documentation to confirm this, but my belief has always been that the numeric values after each line are PARRY's level of Shame (SH), Anger (AN), Fear (FR), Disgust (DS), Insecurity (IN), and Joy (J).—KMP
DonHopkins 44 days ago | parent | context | favorite | on: The Revival of Medley/Interlisp
That's right, it's just a throw-away quip, but if you want the deep nuanced story and inside history of Common Lisp and comparison with Scheme, Kent Pitman is the one to read:
>In 1983, I finished the multi-year task of writing The Revised Maclisp Manual (Saturday Evening Edition), sometimes known as The Pitmanual, and published it as a Technical Report at MIT's Lab for Computer Science. In 2007, I finished dusting that document off and published it to the web as the Sunday Morning Edition.
Not to be confused with David Moon who wrote the "MacLISP Reference Manual" aka the "Moonual", and who co-authored the "Lisp Machine Manual" with Richard Stallman and Daniel Weinreb, which had big bold lettering that ran around the spine and back of the cover, so it was known as the "LISP CHINE NUAL" (reading only letters on the front).
The cover of the Lisp Machine Manual had the title printed in all caps diagonally wrapped around the spine, so on the front you could only read "LISP CHINE NUAL". So the title was phonetically pronounced: "Lisp Sheen Nual".
My friend Nick made a run of custom silkscreened orange LISP CHINE NUAL t-shirts (most places won't print around the side like that).
I was wearing mine in Amsterdam at Dappermarkt on Queen's Day (when everyone's supposed to wear orange, so I didn't stand out), and some random hacker (who turned out to be a university grad student) came up to me at random and said he recognized my t-shirt!
The reference manual for the Lisp Machine, a computer designed at MIT especially for running the LISP language. It is called this because the title, LISP MACHINE MANUAL, appears in big block letters -- wrapped around the cover in such a way that you have to open the cover out flat to see the whole thing. If you look at just the front cover, you see only part of the title, and it reads "LISP CHINE NUAL"
toomanybeersies on Sept 7, 2017 | parent | next [–]
>Here's Kent Pittman's :TEACH;LISP from ITS, which is a MACLISP program that teaches you how to program in MACLISP. (That's "Man And Computer Lisp" from "Project MAC", not "Macintosh Lisp".)
In a very rough sense, it actually seems more similar to what LLMs are doing than I would have assumed. That might mainly be because they are both attempting to generate more text from an input.
Note that Weizenbaum was an AI critic: Weizenbaum's intention was not for Eliza to pass the Turing test, but to show to people that a clearly not intelligent program based on primitive pattern matching can appear to behave intelligently.
He failed: His own secretary wanted to be left alone with the software and typed in her personal problems. Work on Eliza (1963-65, paper published 1966) until today is mostly misunderstood.