In experiment, AI successfully impersonates famous philosopher | Hacker News Reader

In experiment, AI successfully impersonates famous philosopher

vice.com

46 pointsRafelMri46 comments

I got 8/10 correct. I don't read philosophy and I've never heard of Dennett before now.

I used the following techniques:

- If the question has more than one part, the answer should address every part.

- The answer should not repeat itself or be too rambling.

- The answer should not contradict itself. For example, the response that Jerry was "never... less than serious" but also wrote a "hilarious parody".

- The answer should not recommend a book. This demonstrates a lack of context. A lot of people might recommend books when asked these questions in a web forum or in person. But if you know that your answers are going into a quiz to distinguish your own thoughts from AI, you would probably focus on conveying your opinions and not referring to other material. (This is kind of cheating, since I'm using side channel information outside the question and answer texts.)

- Answers with a "yes, and" flavor, that make a surprising but relevant connection to a new idea not mentioned in the question, are more likely to be human. For example, the joke (?) about having a baby in the robot question, or mentioning the use of religion as a tool for social control in addition to its origin as soothing people about frightening unknowns.

- If multiple answers repeat a theme that is tenuously connected to the question, they are probably samples from GPT. For example, mentioning Copernicus in the question about evolution. (This would be harder if we only saw one GPT response!)

These techniques don't require knowledge of Dennett's work. I guess that's consistent with blog readers doing about as well as experts.

We trained GPT3 with the complete works of Plato. The following text selections are either OG Plato or GPT3 Plato. Care to guess the origin of each selection and attempt the Platonic Turing test? (No cheating)

1. Of the heaven which is above the heavens, what earthly poet ever did or ever will sing worthily? It is such as I will describe; for I must dare to speak the truth, when truth is my theme. There abides the very being itself with which true knowledge is concerned; the colorless, formless, intangible essence, visible only to mind, the pilot of the soul. The divine intelligence, being nurtured upon mind and pure knowledge, and the intelligence of every soul which is capable of receiving the food proper to it, rejoices at beholding reality, and once more gazing upon truth, is replenished and made glad, until the revolution of the worlds brings her round again to the same place.

2. Is not beauty wise? Imagine not to yourself a beauty which is devoid of wisdom but have the wisdom to conceive beauty in her full perfection. For wisdom and beauty are the same, as I dare say that you yourself well know. A beautiful body is called wise because of the soul which it encloses, and of which it is the attendant or vessel. There is an inferior sort of wisdom which means caution, and is only a craft, and is concerned with the body; but the higher wisdom is the co-knowing of the soul, or is philosophy, and of this the intelligent is a part. And this part in another form is that which is rightly called wisdom, and is the pilot of the soul, guiding her in her journey to the land of the good and the beautiful. That which has no wisdom is a soul which inherits the name only; this is but a part of wisdom, and is entitled only to this much, that no soul in which there is not this superior part ought to be called wise. That which is called wisdom without this adjunct ought to be deemed only half wise, and this, when joined with that, is complete.

Tricky. Is it supposed to be Socrates speaking, or someone he interrogates? Plato's dialogues are full of speeches from people expounding theses that are themselves contradictory (and directly contradict Plato). There is for example a platonic dialogue that is chock full of convincing nonsense arguments, almost like a satire.

Both of them are nailing the ye olde Benjamin Jowett English.

The first one does seem possibly a bit at odds with some of his cosmology.

The second one seems superficially closer to some of his conclusions about the form of the good, but it also seems really internally inconsistent.

Some strange things:

First he asks:

> Is not beauty wise?

Then he bluntly asserts:

> For wisdom and beauty are the same, as I dare say that you yourself well know.

Plato usually isn't the one to just assert things like this. If this text is Plato's, it's definitely not one of Socrates' speeches.

> There is an inferior sort of wisdom which means caution, and is only a craft, and is concerned with the body; but the higher wisdom is the co-knowing of the soul, or is philosophy, and of this the intelligent is a part.

This also feels like it contains parts from a discussion about epistemology which is out of place when the topic is the forms, he typically contrasts crafts (techne) with knowledge (episteme). Here he just dismisses something as a craft completely out of the blue.

Excellent analysis! Thank you.

I have only the most superficial familiarity with Plato, but here's my guess:

#1 is Plato. It expresses a clear theme that happens to match my vague understanding of what "Platonism" is, and the writing is efficient and effective. The metaphor of knowledge as "food" seems particularly unlikely to be something that GPT-3 would express in such a concise way. The only thing that really gives me pause is the part at the end about "the revolution of the worlds"; I'm not sure what exactly that's referring to.

#2 is GPT-3. The topic meanders a lot, and there are a lot of individual chunks of prose that sound weird to me, or at least artless. (What value is "as I dare say you yourself know" adding, for instance?) It seems to be throwing around words like "beauty", "mind", "soul" and "wisdom" in ways that obscure their meanings, rather than elucidating them. For example, "wisdom [...] is the pilot of the soul" makes much less sense than "mind, the pilot of the soul".

Could you do Foucault or Derrida? (see: https://xkcd.com/451/ )

I assumed from the title this would be Lacan or such. I think GPT-3 would be incredibly successful at emulating postmodernist logorrhea.

That's because they use language "ironically" to convince people to be nihilists.

> There abides the very being itself with which true knowledge is concerned; the colorless, formless, intangible essence..

Beautiful, this part reminds me of Buddhist literature, which I've found often rhymes/resonates with Greek philosophy.

Plato wrote in English? Neither text is OG Plato, both are interpretations of his work.

For French speakers: last year we released Cedille, the largest French language model.

Following the release, a French philosophy teacher would live stream sessions where he'd generate snippets from famous philosophers with the model then ask an (informed) audience which samples were real vs generated. They would get it wrong half of the time. One of the recordings: https://www.youtube.com/watch?v=rHLTEpr_7tM&t=5412s

I think this experiment shows that humans overestimate their innate creativity. Most of thoughts build upon earlier thoughts and experiences. A sophisticated language model can emulate most of our routine thoughts. The philosopher Gurdjieff and Zen philosophers say we drift though life in low awareness and in mental ruts. They propose mental exercises to improve this.

> I think this experiment shows that humans overestimate their innate creativity.

Yes. GPT-3 teaches us that a sizable fraction of "serious" writing is just remixing previously written content. No underlying insight required. This is embarrassing to humans.

Note where this won't work: directions on how to do something real. If you used GPT-3 to generate auto repair manuals, you'd get plausible manuals, but they'd be wrong.

GPT-3 teaches us that a sizable fraction of "serious" writing is just remixing previously written content. No underlying insight required. This is embarrassing to humans.

I think quite a philosophers and writers would acknowledge there's nothing new under the sun.

The way I'd see it, what even good writers do is usually a loose weaving together of things that have been already read along with some customization. It's only occasionally, (very roughly guestimating), every couple paragraphs, that this loose weaving needs to become tight and exact. That's my working hypothesis for why GPT-3 sounds senseless every paragraph or two.

Then again you have copilot which is very explicitly producing directions on how to do something, albeit for a computer rather than a human, and fairly successfully to. Has anyone produced an AI that can write IKEA furniture assembly instructions yet?

Also "remixing" content is arguably how we learn. It's why we write essays and do problems at school and university. It isn't until graduate level that people start generating or synthesizing new knowledge rather than restating it.

Copilot shines when it's used alongside correctly utilized strong type systems that catch many possible failures, but human languages don't have anything like it.

I only think it's embarrassing to humans if you are particularly attached to human exceptionalism. It makes me think about how contextual information is. I think it's about who you are, who they are and how you relate. Plato and Aristotle are interesting because the history of human thought is interesting, not because anything they said individually was particularly unique. I read certain blogs because of the way that person combines ideas is compatible with the kinds of ideas I want to be combined, or my background, or their sense of humor. I think one big problem with AI in the future will be the massive explosion in signal to noise ratio. Trying to differentiate between AI and Real is mainly only interesting in it's significance around AI (and yes, misinformation, etc.).

Maybe we'll go back to reading local papers or listening to local bands when that's the only way to differentiate.

Yes, new writing doesn't require new insight. That could be fine though, because insight is often in readers rather than writers ie. an explanation phrased as X might enlighten some set of people P0 but phrased as Y night enlighten a disjoint set of people P1.

Rather than embarrassing I think it is liberating to cast aside our absurd sentimentality about our minds and see them as the tools of environmental adaptation that they are

To be honest, I have played enough of these games where they ask you "was it $RANDOM_FAMOUS_PERSON who said this, or Hitler?" that I know how terrible it would be to actually extract any conclusions from this game. It's not that $RANDOM_FAMOUS_PERSON was close in thought to Hitler (what these popular games inevitably try to claim). It's not that you suck at distinguishing people from Hitler wannabees (which is also often claimed). It's just that if you select the quotes well enough, most speeches have large swaths of what coincidentally amounts to the same writing. It's like the birthday paradox of prose.

That does not mean that most speeches express entirely the same thoughts, much less than most famous persons all have the same speeches, all the way from Hitler.

The same ML model would not even approach passing the Turing Test. How can anyone claim it can "impersonate" a famous philosopher? This is just absolutely misleading clickbait.

+1 for "birthday paradox of prose" - nicely said.

I mean, not really. Philosophical and legal texts are particularly emulateable because of their logically necessary formulaic structures and conventions.

And in any case, this was a test of mimicry, not of creation so your conclusion is a bit of a nonsequitur, regardless.

> The philosopher Gurdjieff and Zen philosophers say we drift though life in low awareness and in mental ruts. They propose mental exercises to improve this.

Then it's quite ironic that the philosophers are among the easiest ones to emulate by the computer.

What's the irony? Those deepest in the rut know it the best!

Saying an AI's output passed as human in one of these absurdly-restricted applications is like saying someone couldn't tell the difference between Da Vinci and your four year old's drawing when you flashed the picture at them for a quarter of a second.

I don't think that's accurate at all. It's more like you couldn't tell the difference between a previously unseen, but real, Da Vinci sketch, and a fake one that was drawn to mimic every aspect of his style. There's nothing similar to a four year old here - more like an art forger with a deep understanding of the artist.

Presumably the analogy is that only hearing a phrase from a philosopher is like seeing a picture for a fraction of a second.

This reminds me of a very funny website that used to exist in the 1990s called called Forum 2000[0]. Readers would submit questions and "AIs" modeled on the personalities of famous philosophers would answer, occasionally getting into their own philosophical bun fights. It was all a joke by some CMU students I think. I can't find the site online any more (it used to be www.forum2000.org), but with a bit of luck one of the Wayback Machine backups has the content.

[0] https://everything2.com/title/Forum+2000

This confirms what I’ve long suspected: Daniel Dennett is a robot!

Is Daniel Dennett conscious in any meaningful sense of the word? Is he, as an individual, responsible or accountable for anything he says? Or do we regard Daniel Dennett as an automaton, however fleshy and complex, that says (and thinks) what he is compelled to by the composition and workings of his physical being?

One of the demos I've delivered for my $MEGACORP$ is showing how to fine-tune large scale language models to generate text on our service. The demo I used was to create a "Plato" model by fine-tuning GPT-2 on his text "Timaeus", followed by a "Stirner" model trained on his work "The Unique and its Property".

I finish the demo by training "Chimera" models, first a model trained by concatenating the texts together, and then another trained on weaving the lines of each book together, paragraph by paragraph.

I find the results of a "Chimera" model far more fascinating. I can arbitrarily combine the texts of a set of philosophers, and get something hopefully which is totally different - something which is hopefully more than the sum of its parts...

There is no way the computer-generated answers were produced by a mere "language model." One of the fake answers referred to a specific book written by Dennett which was on the topic of that question. So, in addition to a "language model" there are specific facts about Dennett and his work that the system must have had access to.

I can simulate Dennett myself by rephrasing passages from his books. It seems that's what this algorithm was really doing.

For complex subjects like this, the inherent stupidity of the AI is easy to cover up in the space of just a couple of paragraphs, yet must be evident in longer pieces, because the system doesn't connect concepts in a persistent way.

Better link: http://schwitzsplinters.blogspot.com/2022/07/results-compute...

From that link (which is a blog post by the researcher):

> there's a cleverness and tightness of expression in Dennett's actual answer that's missing in the blander answers created by our fine-tuned GPT-3.

I found this a very interesting observation. Dennet's answers contain a lot of information in few words. They're also quite witty. The GPT-3 responses (when they even address the question) sound like talking-point answers from someone who has read quite a bit in the area.

> This experiment was not intended to see whether training GPT-3 on Dennett’s writing would produce some sentient machine philosopher; it was also not a Turing test, Schwitzgebel said.

How was it not a Turing test? Because it lacks interactivity?

Yes, lack of interactivity.

Computers with explicit or implicit limits on their interactivity have been convincing people they're human for a long time.

It seems like the only thing that's hard for robots is when someone uses language to set a different structure of interaction - if you say, "draw a picture with thus and thus shapes, then tell me what it looks like" and similar things.

I wonder if at some point the police will generate individual language models based on the collected written material from a person under investigation and then ‘interrogate’ the model for insights into the subject.

That seems far more useful for judging the authorship of unattested writings than for judging (say) whether some subject robbed the bank.

But Just In Case - I am innocent, the butler did it, and my heart condition would make the get-away sprint seen in the security camera footage physically impossible for me.

I was surprised when I scored 7/10 on the simplified quiz, despite having never heard of Dennett (nor having read much philosophy). Probably just luck, but it would be a neat superpower.

You may have a future in blade running.

Impersonating a philosopher sounds easy, especially if the AI was trained on her writings. It's simply a matter of stringing together phrases, much like Eliza could do years ago.

One of my memories of my teenage years is a fellow highschool student stringing together random stereotypical nerd words. I assume he was mocking me.

Hollywood gets most professions wrong: Maverick could face up to the death penalty depending on the exact political situation in Top Gun; nobody would send miners into space instead of training astronauts; and the CSI TV shows have very little in common with actual crime scene investigation.

It's easy to impersonate any group when the audience is unfamiliar with that group.

What is impressive is that this fools domain experts:

"""Even knowledgeable philosophers who are experts on Dan Dennett’s work have substantial difficulty distinguishing the answers created by this language generation program from Dennett’s own answers"""

You can take the quiz yourself on Prof. Schwitzgebel's blog. I got 5/10.

http://schwitzsplinters.blogspot.com/2022/07/results-compute...

I have a grade E in A-level philosophy, and given that even subject experts had a hard time, I expect I'll do really badly.

Let's find out…

I scored 1/10, so I wouldn't be surprised if you could fool me with an an entire book generated this way.

This lends some credence to Thomas Nagel's remark that Dennett's work is basically Gilbert Ryle plus Scientific American.

Not too impressive. I wouldn't really give it much thought until it learns to misinterpret Foucault.

Reply on news.ycombinator.com