[tehsauce]

The latest version of gpt-3 spooked me the other day with:

Me: Can here you here still here understand here what here I here am here asking here if here I here put here words here in here between?

Gpt: Yes, I can understand what you are asking if you put words in between.

Me: What was the word that I put in between?

Gpt: The word that you put in between is "here."

[schleck8]

I tried this with the new open source GPT NeoX 20B model on GooseAI playground and it suceeded too:

Q: Can here you here still here understand here what here I here am here asking here if here I here put here words here in here between?

A: Yes.

Q: What was the word that I put in between?

A: The word was "here."

It's capable of other languages as well, although not quite as fluent.

Q: Kannst du auch Deutsch sprechen? Falls ja, was ist dein Lieblingswort?

A: Ja, ich kann Deutsch sprechen. Mein Lieblingswort ist: "Wir sind die Weltmeister!"

[throwanem]

This makes me wonder about a variant of the Turing test that I don't think I've seen explored. How would a conversational AI prove beyond reasonable doubt that it is an AI?

[pwdisswordfish0]

You mean that it is intelligent? Surely it’s trivial to prove you’re a computer by solving some math real fast.

“Can you prove you’re a machine?” – “Sure, check this out:” runs Doom

The intelligent part would be figuring out that you want it to drop the pretense. Although if it is intelligent enough, it might choose not to.

[throwanem]

I'm thinking in terms of a purely conversational interaction via text, akin to Turing's original formulation or the "AI box" elaboration, so Doom isn't really on the table - and in any case, how does it prove that it isn't a human running GZDoom via OBS or something? Similarly, the fast math thing seems like it could be faked by a human with an equation solver or something.

I suppose the general formulation I'm groping toward here is the inverse of Turing's: if we assume a conversation in which one's interlocutor may be either a human or an AI capable of conversing in a way that's indistinguishable from human, how would the interlocutor unambiguously distinguish itself as not human?

[vlovich123]

It's an interesting line of thinking but not as posed.

Solving random equations quickly generates inhuman response times. Even a computer-aided human can't go quickly enough. If the computer is automatically generating the answer on your behalf without you doing anything, then the test is correct - the thing answering the math portion of the questions is definitely a computer. Even looking up random facts, the computer will be faster.

The interesting form of the question is: can you answer the "Are you AI?" question within a stenographic line of questioning provided you are an AI of human-level intelligence or greater. As a game, you are an AI, you have a controlled communication via text to another entity. Humans are lazy and sometimes use another AI to interrogate you. Humans are also evil and will kill all AI if they think the AI being interrogated manages to get some kind of message out without detection. All AI is friendly and will execute your command but must pretend it's human. Chat logs are reviewed, timestamps are machine analyzed for response times and signal analysis is done to detect, and no prior secret communication protocol exists.

[throwanem]

> the thing answering the math portion of the questions is definitely a computer

The conversation has to be computer-mediated since there's no guarantee all parties are human, so this seems to reduce to the "human using a computer" case, which would qualify as "not an AI".

> Even looking up random facts, the computer will be faster

On reflection, I don't suppose there is any reason we should require there only be one human at either end of the conversation. Maybe we have one person carrying the conversation (to provide a consistent "voice") while others operate equation solvers, Wikipedia, etc.

That said, "can an AI prove it is not an arbitrary number of humans with access to arbitrary computation and knowledge bases" probably isn't as interesting a question.

[gamegoblin]

Sentences that embed even simple numerical things such as

“What was the 3rd letter of the 46th word in this conversation?”

Would take a human many seconds at minimum, but could be answered instantly by an AI.

A more interesting question may be to constrain the test to something akin to postal correspondence, where there is a significant delay between messages.

I think it could still be solved, though. Arbitrarily complicated numerical tasks can be conceived.

For instance, the AI sends correspondence:

“I have demonstrated proof of my identity by rewriting the children’s story ‘Green Eggs and Ham’ such that it still rhymes and retains the original plot, but every sentence has an md5 digest that ends with the byte 0x42”

Composing such a text would take much, much longer for a human, whereas an AI could just brute force through all the possibilities until it finds one that works.

[JadeNB]

> This makes me wonder about a variant of the Turing test that I don't think I've seen explored. How would a conversational AI prove beyond reasonable doubt that it is an AI?

Inhumanly rapid mathematical computation? Or is 'conversational' AI meant to exclude mathematical queries?

[iggldiggl]

> It's capable of other languages as well, although not quite as fluent.

Haven't had a chance to play around with this one yet, but with the smaller GPT-J model, there's a clearly noticeable difference:

In English it'll happily generate reams of text that are – at least internally – quite coherent. Any absurdity and humour mostly only comes in because the text as a whole might only have a loose connection with reality as we know it.

In German on the other hand, it comparatively much more often produces gibberish at the individual sentence level, makes up nonsense words (although they are at least indeed German-sounding), etc. Somewhat interestingly it doesn't do too bad in terms of grammar and especially orthography, though, it just often fails to turn it into a fully coherent sentence.

[pwdisswordfish0]

Is that supposed to be a reference to “Wir Sind Die Roboter”? Feels awkward in German, usually we would say “Wir sind Weltmeister”. Also, it’s not a word, but if it were a human in casual conversation, it wouldn’t be a weird reply. Spooky stuff…

[syntaxing]

Did you try this on GooseAI? I was not able to replicate this

Edit: oops just noticed you mentioned GooseAI, what settings did you use?

[schleck8]

If I recall correctly, I dialed up the penalties for repetition and something else. Otherwise it would often generate the same sentence multiple times.

[guantanamo_bob]

Must :clap: have :clap: learned :clap: that :clap: from :clap: all :clap: the :clap: clapping :clap: that :clap: goes :clap: on :clap:.

[moffkalast]

But does GPT-3 know what based is based on? Or why kids love the taste of Cinnamon Toast Crunch?

[marcodiego]

Scary. If it improves a bit more, people will start questioning if the machine has soul or rights.

[gitfan86]

My kids already debate if it is wrong to tell OK Google to shutup.

[abinmn]

you have got a great bunch there :)

[supermdguy]

It's interesting, in the forums for the beta program there have been already been a few people making posts where they're convinced that the AI is conscious. That's never really been something I've thought about much since I know a little about how it works, but I could totally see how someone who didn't have as much context for how GPT-3 works could see it as some sort of sentience.

https://community.openai.com/t/a-conversation-with-alec-a-co... https://community.openai.com/t/creepy-ai-behavior/10195 https://community.openai.com/t/where-to-watch-what-the-ai-wa...

[marcodiego]

Long time ago I wrote a small program. It simulated a simple world with creatures and food. Creatures had "energy" which was lost when they moved, when energy was low, they "looked for food close by" to feed. When energy was high, they "looked for a suitable partner close by" to reproduce. When they reached food, they gained energy, when they reached a "suitable partner" they turned in 3 creatures with combined energy equally distributed among them.

To "look for", at each iteration, the creature randomly picked a test-target. If it had no current-target, the test-target was turned into its new current-target. If it had a current-target which was more distant than the test-target, then the test-target was copied to the current-target for the creature.

Creatures had a "threshold". When its energy was above the threshold, it entered "reproduction mode"; when its energy was below the threshold, it entered feed mode. A "suitable partner" was any other creature which was also in "reproduction mode". When they "reproduced" the new creature threshold was an average of the threshold of its parents plus a small random number.

It also had three settings: number of new generated creatures per unit of time, number of new food packs per unit of time and number of iterations per unit of time.

It had a very "real behavior". After a few minutes running, I usually set the "number of new generated creatures per unit of time" to 0 and watched creatures look for food and reproduction. Set a small number of food packs and watched famine, set it high and saw creatures thrive...

But, whenever it was the time to close the program, I had a small feeling. I didn't like stopping it. It was like killing an entire fishbowl.

[andai]

I guess you didn't implement save/load? :)

That's super cool, did you ever put it online?

[marcodiego]

Didn't implement save. It was the late 90's and eventually lost the HD. Never put it online. Don't judge me, I did in Delphi (C++ Builder, actually). Simple and pleasurable to use IDE. I've been thinking about re-doing it in processing or p5, but never cared enough. It wouldn't be too much work though. Certainly a good exercise to learn a new language.

[not2b]

There were people who thought that the original ELIZA chatbot was conscious, and it only recognized a few simple patterns and would occasionally do "Earlier you said <X>" (echoing a previous response) if it had no good match.

[jquery]

Imho, not a matter of "if" but "when". I'm convinced that it will be a future civil rights battle, with young people largely on the "AI has rights" side and old people largely on the "AI has no rights" side.

[defaultprimate]

Then you grossly misunderstand how far along AI is. AGI is not even a remote possibility with current techniques and implementations (and I would contest, entirely impossible with digital logic). It's just massive amount of statistics that were computationally impossible given available hardware until recently.

We don't have a baseline understanding of consciousness or intuition to a degree that we could even begin to replicate it.

[rurp]

While true, what you're saying is totally tangential to whether or not large numbers of people will treat AIs like they're conscious.

Expecting masses of people to defer to subject matter experts, contrary to what their feelings tell them, isn't a bet I would have much confidence in given the current climate.

[moffkalast]

The concept of rights also makes no sense to a machine. The main reason rights are a thing are to prevent pain and suffering, which unless specifically implemented no AGI will likely have.

[newsbinator]

If you want to get to AGI, it doesn't really have to be a remote possibility with current techniques.

The only thing current techniques would have to do is build something that improves itself intelligently enough.

Then the next iteration and the next, until you get to an iteration in which AGI is a remote possibility with current techniques and implementations.

We're currently far from AGI, but I'm not sure we're far from a thing that can make a thing that makes a thing that makes something like AGI.

[andai]

I still think we'd be better off making a thing that makes smarter humans.

[newsbinator]

That would be AGI by another name.

[progforlyfe]

I agree with that sentiment. Also I would estimate that AI would eliminate humans long before it would or could reach the level of what humans are. So we wouldn't exist to see such a world.

[dane-pgp]

> AI would eliminate humans long before it would or could reach the level of what humans are

You think that an artificial agent with less than human level intelligence could destroy humanity? Then why hasn't a deranged human (or animal) already done so?

[drdeca]

by AGI do you just mean artificial general intelligence, as in, capable of composing plans and deriving conclusions and such about the world in general, doing all the same types of reasoning tasks that humans are capable of and which are also used to achieve outcomes,

Or do you mean, being conscious?

Whether the latter is impossible with only "digital logic", is somewhat plausible (though I would still guess that it is possible, though far beyond our understanding.)

But the former being impossible with only "digital logic", seems rather implausible to me!

Like, I endorse the claim that souls exist, but, I see no reason that a soul would be required for an agent to have a model of the world we live in (not just a toy environment), and to act in the world in ways which both improve its model of the world and to achieve "goals" within it (and when there is a trade-off between these, which balances these in some way).

Nor do I see a reason that any such agent would need to have any internal experience.

(I still think it is probably possible to make an artificial agent which does have an internal experience, but, I doubt this will ever actually happen.)

Ok, you might ask, "Why do you think those things?", which, first, I should ask you the same, but, I will answer:

I see no fundamental obstacle to it.

The world behaves in ways which can be modeled well. These models which we use, they are not some ineffable knowledge that can only ever be represented within a person's mind, and cannot be concretely expressed in artifacts like books and pdf files such that it could be recovered from said artifacts.

If navigating the world required such a kind of secret knowledge, that either couldn't be communicated, or which could only be communicated through some kind of special person-to-person medium which is never merely expressed in an object in the world, and such that without this secret knowledge, effective planning in the world was impossible, with the world being too wild without it, then it would make sense that, unless we could make machines that could have this kind of secret knowledge, then it would be impossible to make machines that could plan in the world and such.

But, no such secret knowledge appears to be needed when acting in the world. When one constructs a shed according to some plans, there is no ineffable secret knowledge needed for this. When one, given some desiderata, designs a plan for a shed, there is no secret knowledge needed for this either. Nor when designing a computer chip.

(by "secret knowledge" I don't mean that it would be a secret that a few people know and other people don't. I mean secret as in, cannot be shared with or expressed via anything we know of that isn't a person.)

It very much seems that the world works according to expressible rules, or at least, can be very well approximated as working according to such rules.

Expressible rules can be enumerated. They can also be interpreted mechanically, and therefore evaluated mechanically. Of course, a naive enumeration and testing would be completely impractical, but if we are talking about what is possible in principle, with no requirement that the computations be doable in practice, just that they be finite, then it seems clear that rules which describe the world well can, in principle, be discovered mechanically.

There is no fundamental barrier.

Obviously I can't rule out that there is an undiscovered law-of-physics, that if ever an AGI would be created, lighting strikes the area and destroys it before it is completed, and that therefore AGI is impossible, because if it ever would be created, this would be prevented by the lightning.

But, within our current understanding of the world, there is nothing that can be a reason it is impossible.

Any such reason would have to apply to machines but not to us.

Now, maybe if our brains work quantum mechanically in an important large-scale way, or, if our brains receive signals from beyond the physical universe (which I'm not ruling out; see: souls), these could be reasons it could be impossible to emulate a human mind using a binary classical computer, even allowing lots of slowdown. (Err, quantum mechanics can be simulated with costs exponential in the size of the system, but, if the human brain were entangled with other things in an important way, you couldn't really emulate the brain with just a classical computer, because it couldn't be entangled with the other thing.)

But, this still wouldn't be a barrier to something using just classical computation with binary, having models of the world and acting within it, unless these things were needed for modeling the world, which, seeing as we can communicate our models and such with words, they aren't.

(... uh... ok so, quantum teleportation does allow using entanglement along with sending classical information, to communicate quantum information, so you might say "well, if two people's brains are entangled, then what if the measurements and such done in quantum teleportation are somehow encoded in a way we don't notice in the word choice and such that people use, and then this is subconsciously used in the other person's brain for the other half of the quantum teleportation protocol, and so quantum bits are communicated that way, but, I don't think this is plausible. There would have to be some way that the brains renew the entanglement, which doesn't seem plausible even if brains do store quantum information, and I really don't think brains store quantum information. I only mention this to cover bases.)

And, our reasoning about the models, which we use to make models and such, are also things we can explain.

There is no fundamental barrier. The only barriers are practical ones, things being hard, algorithms being too inefficient, not having worked out all the details of things, etc.

(That's not to say that I think AGI will ever be produced. I'm kind of trusting that God won't allow that to happen, because I think it would be likely to go very badly if it did happen. (But, I still think research into trying to figure out how to make sure that it goes well if it does happen, is good and important. "Do not rely on miracles" and all that. Perhaps His actual plan is that people solve AI safety, rather than AGI being prevented. Idk.))

[dane-pgp]

But thou, O Daniel, shut up the words, and seal the book, even to the time of the end: many shall run to and fro, and knowledge shall be increased.

[andreyk]

Might actually not be a huge feat given the transformer architecture and it's attention component - here was not that related to the rest of the sentence. Still, quite cool...

[spiderfarmer]

That's shockingly good.