64 Comments
User's avatar
Gerben Wierda's avatar

The Turing test proves that people are gullible, not that computers are intelligent.

The paradigm that we have to address is the underlying assumption that people are intelligent. We *are* the most intelligent species on the planet, but that is a *relative* distinction. We're all gossiping 'verbally-fleecing' primates and we're all potential flat-earthers. In an absolute sense not that intelligent at all. This is (after Copernicus and Darwin) probably the hardest paradigm to shift of them all, as our belief in ourselves is basic to our psychology. (That this is our next major paradigm shift (if we succeed, that is) was a core message of my 2020 (IASA) and 2021 (DADD) talks about what the information revolution is doing to us and teaching us. See https://ea.rna.nl/2021/11/02/masterorservant-dadd2021/)

Expand full comment
Denis Poussart's avatar

Let's not forget that this is a prime example of a genuine "Imitation Game" and that it is is recursive: it is the industry making the public beleive that the $$$ it has received puts it on the verge of AGI. So the $$$ can continue to flow in (including from governments everywhere). Gullibility is the critical enabler. Unfortunately general critical thinking does not seem to be in sufficient supply to counter it.

Expand full comment
Jurgen Gravestein's avatar

Interesting, but why does this research surfaces now? This paper has been out for over a year.

Not to pad myself on the back or something, but I actually covered in my newsletter in Dec 2023: https://jurgengravestein.substack.com/p/gpt-4-performs-worse-than-a-coin

Had some interesting email exchanges with one of the researchers, Cameron Jones, too.

Expand full comment
Dakara's avatar

"Since 2014 there has been great progress in human mimicry"

I agree the relevance isn't intelligence; however, what this likely points to is ever increasing ability to manipulate public opinion. Its capabilities for influence are increasing. This is the concerning part. We are the algorithmic society now.

We are going to get a lot more of this: "Hundreds to thousands of simulated accounts are now typical packages being offered to businesses to grow their brand or product."

https://www.mindprison.cc/p/dead-internet-at-scale

Expand full comment
Larry Jewett's avatar

LLMs = humimicry

Expand full comment
Stephen Schiff's avatar

I continue to be amazed by the fascination about various Turing tests given the lack of understanding of what actually constitutes thought. It is awfully hard to measure something if you don't know what it is.

Expand full comment
Larry Jewett's avatar

There are no real measurements in AI “research and development”

That’s why it’s not real science.

Expand full comment
Stephen Schiff's avatar

Well, yes and no, depending on what you include in the definition of AI. Almost all machine learning [sic] techniques are firmly based on mathematical principles and have useful properties such as the ability to predict error probabilities a-priori. That’s because ML includes the process of validation, whereby portions of the data set are held back and used to test the algorithm. Also, the algorithmic forms are inherntly discoverable and thereby subject to comparison with known physical laws, inter alia. In LLMs, that process does not exist, and as far as my weak mind can comprehend, cannot exist. [As an aside I really despise the terminology. Machine Learning is bad enough, but Artificial Intelligence?]

Expand full comment
Larry Jewett's avatar

When you say that “ML includes the process of validation, whereby portions of the data set are held back and used to test the algorithm” I am reminded of the now infamous “test” of a machine learning algorithm to (supposedly) distinguish between photos of huskies and wolves whose identification was actually based on whether images had a snowy background or not.

Expand full comment
Larry Jewett's avatar

Black box methods are inherently unscientific because one can not know whether one is “measuring” what thinks one is measuring.

Expand full comment
Larry Jewett's avatar

I would also draw a distinction between math and science.

Expand full comment
Chad Woodford's avatar

Agree that this doesn't matter so much. But I would also argue that the only Turing Test that should be taken seriously is the one designed by Kapor and Kurzweil for Longbets https://longbets.org/1/

Expand full comment
TheAISlop's avatar

I found the need for customized personas to pass the test was interesting. This, a LLM on its own doesn't measure up. And if you want to test yourself just use voice mode with any model, and see how much you really like it. Voice mode still feels like a novelty act.

Expand full comment
Ian Douglas Rushlau's avatar

'To do well the system had to be coached to adopt the persona of “a young person who is introverted, knowledgeable about internet culture, and uses slang”'

So, the tech bros who are certain they are geniuses design an automated parrot that... reminds them of themselves. Upon listening to themselves on a playback loop, they declare the automated parrot to be intelligent. Because, you know, **they sound intelligent to themselves**.

Self-involved, self-aggrandizing and obtuse.

Unfortunately, these same tech bros are well-funded by similarly self-involved, self-aggrandizing and obtuse finance bros, and have been given free reign to impose their stupidity on our entire society.

What could possibly go wrong?

Expand full comment
Patrick Senti's avatar

Corollary: a system built to imitate human-like conversation will eventually convince some humans that it is, in fact, human.

No proof of intelligence.

Expand full comment
Amy A's avatar

Why does the evasive persona sound so much like a Republican at a Senate confirmation hearing? Did Hegseth have an AI whispering non answers to him?

Expand full comment
Peter Byrne's avatar

And we are supposed to allow the Chats to operate lethal weapons? See this new report on Military AI: https://www.projectcensored.org/military-ai-watch/

Expand full comment
Ken Kahn's avatar

From https://academic.oup.com/book/42030/chapter-abstract/355746927

This discussion between Turing, Newman, R. B. Braithwaite, and G. Jefferson was

recorded by the BBC on 10 January 1952 and broadcast on BBC Radio on the

14th, and again on the 23rd, of that month. This is the earliest known recorded

discussion of artificial intelligence.

...

Braithwaite: Would the questions have to be sums, or could I ask it what it had

had for breakfast?

Turing: Oh yes, anything. And the questions don’t really have to be questions,

any more than questions in a law court are really questions. You know the sort

of thing. ‘I put it to you that you are only pretending to be a man’ would be

quite in order. Likewise the machine would be permitted all sorts of tricks so

as to appear more man-like, such as waiting a bit before giving the answer, or

making spelling mistakes, but it can’t make smudges on the paper, any more

than one can send smudges by telegraph. We had better suppose that each jury

has to judge quite a number of times, and that sometimes they really are

dealing with a man and not a machine. That will prevent them saying ‘It must

be a machine’ every time without proper consideration.

Well, that’s my test. Of course I am not saying at present either that

machines really could pass the test, or that they couldn’t. My suggestion is

just that this is the question we should discuss. It’s not the same as ‘Do

machines think,’ but it seems near enough for our present purpose, and raises

much the same difficulties.

Expand full comment
Saty Chary's avatar

Hi Gary you had hit the nail right on the head... when you'd said that 'the Test' isn't a measure of intelligence.

Biological beings exhibit intelligence, without explicit computation. Given that, it's rather absurd to equate that with computing machines' computational outputs.

In other words, no AI that calculates anything, can claim to be intelligent like us - that encompasses AI that does symbolic reasoning, NN and RL calcs.

Expand full comment
Linda M. Holliday's avatar

Not only is the touring test kind of stupid. It was obvious during my my 30 years in tech, the test. set up a competition between humans and computers where developers try to be as good as or better than humans instead of assisting humans. We passed by so many amazing useful technologies in pursuit of BS. A huge unintended consequence that nobody atones for.

Expand full comment
David Roberts's avatar

So the takeaway here is that the Turing Test was never a good test because it's subjective and is at best a sliding scale. Some humans are fooled by ELIZA; others are fooled by ChatGPT; still others by no existing AI. It all comes down to the skill of the human in asking probing questions and the length of the conversation. Once you know what to look for, it's very easy to trip up any existing LLM. Surely, I find the existing LLMs much more advanced than ELIZA, but that's not saying much. I think one of the worst aspects of modern LLMs is that they sound sure of themselves even when they are spewing hallucinatory sewage. A human will tell you that it doesn't know or at least hedge. Current LLMs will confidently assert nonsense.

Expand full comment
Larry Jewett's avatar

LLMs would make very good politicians — and billionaires.

Expand full comment
Larry Jewett's avatar

And AI salesmen

Expand full comment
mpsingh's avatar

ASI- Artificial Sam Altman Impersonator

Expand full comment
Larry Jewett's avatar

Then again, maybe ChatGPT is not impersonating but IS Sam Altman.

After all, Altman has AI in his name.

And have we ever seen them in the same room together?

Expand full comment
Contextual Mind's avatar

Hey Gary. I'm a bystander to your posts and I find them endlessly fascinating in role in pushing AI development. Thank you for that!

Question - and you've likely answered this elsewhere, but at its core, isn't mimicry the core of what it means to be a functionally intelligent being? The duality of nature plays a role here it seems to me. Would love to hear your thoughts.

Thanks again for your writing!

Expand full comment
Gary Marcus's avatar

mimicry is one of many things we can do. it’s a long way from the whole enchilada.

Expand full comment
Contextual Mind's avatar

Thanks Gary, appreciate the clarification. Point taken that it's one piece of a much larger puzzle.

Expand full comment
Forrest's avatar

I have listened to many hours of dungeons and dragons podcasts. I can mimic a dungeons and dragons game host quite well.

Despite this, I don't own, nor have I read or understood, the core rulebooks for the game. So if I tried to run a game, I would flail and fail as soon as I met a truly novel situation.

Mimicry only gets you so far, without understanding.

Expand full comment
Bruce Cohen's avatar

Yes. I think you can describe a large part of human learning as a process of mimicry, followed by abstraction of the mimicked behavior/information and breaking it into modules which can be interchanged with other modules already learned and which have similarities to the originals but allow modification of outcomes, application to different stimulations, and/or further development of the abstractions. Abstraction and modularization require knowledge, I would argue of at least some level of symbolic representation (not necessarily embedded in. language).

Expand full comment
Contextual Mind's avatar

Great point! Let me clarify mine. The duality of the thing, actor and mimic, is where the seed of creative intelligence might be born. Asymmetries in the mimic results in slight changes to the action (no exchange of information is 100% efficient at the molecular level in my hypothesis; so I’ve scaled the idea to intelligence building outcomes through asymmetric mimicry), and those changes could be the seed of creativity.

You play DnD in your own creative way by mimicking the mechanics without knowing the way of the thing. You have a creative game unique to that gaming session. As LLMs develop the capacity for more creative outcomes through asymmetric mimicry - maybe you see the path I’m outlining.

Expand full comment