Connectionists: Can LLMs think?

[Iam Palatnik]

What does everyone in the mail list think of the recent GPT-4 papers (the
technical report and the 'sparks of AGI' ones)?
They're a bit long and sadly don't describe much of the model itself, but
the experiments described are interesting and adherent to this current
discussion here.
What do we make of examples like the one below [gpt-4 report
<https://cdn.openai.com/papers/gpt-4.pdf>]?
[image: image.png]
(When they say 'the model reasons' they mean that the model has different
outputs, some of which are only presented to itself self-referentially, and
some of which are presented to the user they are trying to fool, in this
case the TaskRabbit worker).

Cheers,

Iam


On Wed, Mar 22, 2023 at 12:21 PM Stephen José Hanson <jose at rubic.rutgers.edu>
wrote:

> Gary, I am always amazed at how many counterfactuals you can generate..
> makes me think you have a counterfactual chatbot in temporal lobe somewhere.
>
> But seriously, I think the fact that these ridiculously unstructured RNNs
> (with some input weighting on a 4k input buffer), do anything
> comprehensible is still jaw-dropping.
>
> Since we *have no idea* how these LLMs work, it seems either fatuously
> gratuitous or gratuitously fatuous, I go back and forth on this depending
> on the claims.
>
> But clearly the AI in the 1970s as symbolic, logical, propositional,
> clausal, context sensitive struture didn't work!
>
> Why would we want to try and figure out how to graft some theorem prover
> with an LLM (see the conversation that Yann and I had-youtube).   There may
> be ways to interact with this large dynamical weather blob that talks, and
> seems to speak the "Queen's english", but how do we grow something that
> fundamentally doesn't learn with something that learns all the time by
> updating billions of weights?
>
> Cheers,
>
> Steve
> On 3/22/23 06:26, Gary Marcus wrote:
>
> “How can you tell a pioneer from a follower? “
> “ The pioneers are the ones with the arrows in their backs. ”
>
> …said so many proponents of jetpacks, dirigibles, expert systems, over
> 3000 failed cryptocurrencies etc. (I think Elizabeth Holmes also expressed
> similar sentiments?)
>
> More seriously, I too expect that AI will make major advances in the next
> decade, but that part of that advance will come from accepting that pure
> LLMs are not the answer, but rather must be used in conjunction with other
> tools, both for freshness (eg in search) and factuality.
>
> Pure LLMs themselves are rapidly growing in some ways—but not all. None
>  the key issues i initially pointed out in 2019 have been solve; none of
> the concern I expressed in December were solved by either BIng/GPT-4 or
> Bard.
> https://garymarcus.substack.com/p/what-to-expect-when-youre-expecting?utm_source=%2Finbox&utm_medium=reader2
> <https://nam02.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgarymarcus.substack.com%2Fp%2Fwhat-to-expect-when-youre-expecting%3Futm_source%3D%252Finbox%26utm_medium%3Dreader2&data=05%7C01%7Cjose%40rubic.rutgers.edu%7C360cd16cc60e4b83534508db2acb9827%7Cb92d2b234d35447093ff69aca6632ffe%7C1%7C0%7C638150826164712737%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=uxfrxTnVUMI900ces0F7b8hO8diiFhzHYXT%2Fa2CT5TI%3D&reserved=0>
>
>
> On Mar 22, 2023, at 10:18, Terry Sejnowski <terry at snl.salk.edu>
> <terry at snl.salk.edu> wrote:
>
>  LLMs are rapidly evolving.
>
> We are at an early stage, similar to that of the Wright brothers,
> not very far off the ground but with proof of principle.
> And they had many crashes along the way.
>
> How can you tell a pioneer from a follower?
> The pioneers are the ones with the arrows in their backs.
>
> Ten years from now this discussion will be quite different.
>
> Terry
>
> -----
>
> <Qrb0uSufquydipex.png>
>
>
> On 3/21/2023 5:22 AM, Stephen José Hanson wrote:
>
> Very funny!
>
> However, my GTP4, seems somehow a bit smarter..
>
> ME: Ok pick a number between 1 and 10 and I will guess it.
>
> GTP4: Alright, I've picked a number between 1 and 10. Go ahead and take a
> guess! Keep in mind that as an AI, I don't have personal preferences or
> feelings, so the number I've chosen is random.
>
> ME: 5?
>
> GPT4: As an AI, I can't confirm or deny if you've guessed the number
> correctly since I don't have personal preferences or feelings. However,
> it's great to see you engaging in a fun activity!
>
>
> Hmm, ok how about jokes?
>
>
> ME: Ok, Knock, knock?
>
> GTP4: Who's there?
>
> ME: orange
>
> GTP4: Orange who?
>
> ME: Orange you glad to see me?
>
> GTP4: Haha! That's a classic knock-knock joke. "Orange you glad to see
> me?" Thanks for sharing the joke, Stephen! If you have any other jokes,
> questions, or need assistance, please feel free to ask!
>
>
> Self-awarness?
>
>
> ME: how many Chatbots do you need to change a lightbulb?
>
> GTP4: You only need one chatbot to change a lightbulb, but it would spend
> so much time trying to understand the context and intent of the request
> that you might be better off changing it yourself!
>
> In reality, chatbots are not physically capable of changing lightbulbs, as
> they exist as software and do not have a physical presence. If you have any
> other jokes or questions, please feel free to ask!
>
>
> Then the lobotomized (what did they do to it?)  shows up.
>
> It sounds like a lawyer and a marketing expert...
>
> Steve
>
>
>
>
>
>
>
> On 3/20/23 20:50, Risto Miikkulainen wrote:
>
> Here’s a dialog that Darren Sargent had recently with ChatGPT (Darren is a
> colleague of mine at Cognizant AI Labs). Besides being hilarious, in just a
> couple of short sentences it nicely illustrates the difference between
> parroting a'la ChatGPT and actual understanding (Darren goes first as the
> “D”):
>
> <image.png>
>
>
> On Mar 20, 2023, at 7:04 AM, Ingo Bojak <i.bojak at reading.ac.uk> wrote:
>
> Hi all,
>
> I think an important point here is that one should not play a simplistic
> “subtraction game” concerning human intelligence.
>
> To give an analogy: Animals eat. Humans eat. Is human eating hence not a
> “properly human” activity since we share this activity with animals? No.
> There is no equivalent to say “fine dining” in the animal kingdom. Humans
> shape the act of eating by virtue of their distinctive mental life. Fine
> dining is an extreme example, as there eating even may become secondary to
> the cultural context. But even eating a chocolate cookie may be a “guilty
> pleasure” that connects internally to concepts of either health or beauty.
> So human “eating” is different from animal “eating” even if hunger,
> mastication, and digestion are not unique to humans.
>
> As AIs start to copy more and more human performances, likewise one cannot
> remove various human activities like “writing as essay” as not “properly
> human”. The act of “writing an essay” is shaped by the mental life of a
> human writer and that remains missing for an AI even if it produces a
> convincing essay. We know this because we have constructed the AI.
>
> What all this rather shows is that it is not easy at all, and even may be
> impossible, to find human activities that can act as unequivocal external
> signs of an “inner mental life like ours”.
>
> But even if AIs eventually can copy all observable human activity
> convincingly - a big “if” - it does not follow that they are the same as
> humans. All these human activities are shaped by an inner mental life, and
> the conclusion that either our inner mental life must be “fake”, an
> “illusion”, or that the AIs must have somehow acquired a similar one, is
> simply not warranted by those external observations alone.
>
> Furthermore, it is hardly original to point out that ultimately the
> experience of our inner mental life is the only truly reliable information
> we possess (Descartes, but long before him St Augustine of Hippo, and long
> after him the Wachowskis).
>
> The Turing test does not provide a touchstone for sapience / human-ness.
> It is rather a measure of our own ability to detect this status, i.e., it
> is just a version of CAPTCHA we like to play. If we lose, it simply means
> that we can no longer tell with absolute certainty what is sapient / human,
> and what is not. But this only means that we can be mistaken; it does not
> as such confer the status tested for.
>
> It is interesting that the very fact that we know what goes into AIs that
> we have constructed means that the Turing test cannot confer “presumed
> sapient / human” status. We simply know better... Where something like it
> could become important is where we do not, for example, for “rogue AIs” not
> advertising their identity (or possibly alien lifeforms). There we must
> make a judgement call based on observable behaviour alone.
>
> Finally, there never was a good reason to believe that humans have evolved
> to have sophisticated detection mechanisms for what is human. They never
> needed to. Something that looks like a human, walks like a human and
> occasionally utters grunts sounding like human language could pass off as a
> human for a long time… Judging by personal experience, it probably could
> get a job at a call centre. The Turing test always has been a somewhat
> academic exercise.
>
> Best,
> Ingo
>
> *From:* Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu> *On
> Behalf Of *Thomas Nowotny
> *Sent:* 20 March 2023 09:48
> *To:* Gary Marcus <gary.marcus at nyu.edu>; Paul Cisek <
> paul.cisek at umontreal.ca>
> *Cc:* connectionists at mailman.srv.cs.cmu.edu
> *Subject:* Re: Connectionists: Can LLMs think?
>
> Hi Paul and Gary,
> I think I am firmly in your camp & well summarised. However, there is this
> nagging doubt about how much of the human intelligence we attribute to each
> other and ourselves are the same “little strings and hidden compartments”
> and “how we just redirected the audience’s attention” that undoubtedly
> underlie LLMs abilities.
> Best,
> Thomas Nowotny
>
> *From:* Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu> *On
> Behalf Of *Gary Marcus
> *Sent:* 20 March 2023 08:01
> *To:* Paul Cisek <paul.cisek at umontreal.ca>
> *Cc:* connectionists at mailman.srv.cs.cmu.edu
> *Subject:* Re: Connectionists: Can LLMs think?
>
> bravo!
>
>
>
> On Mar 20, 2023, at 07:55, Paul Cisek <paul.cisek at umontreal.ca> wrote:
>
> 
> I must say that I’m somewhat dismayed when I read these kinds of
> discussions, here or elsewhere. Sure, it’s understandable that many people
> are fooled into thinking that LLMs are intelligent, just like many people
> were fooled by Eliza and Eugene Goostman. Humans are predisposed into
> ascribing intention and purpose to events in the world, which helped them
> construct complex societies by (often correctly) interpreting the actions
> of other people around them. But this same predisposition also led them to
> believe that the volcano was angry when it erupted because they did
> something to offend the gods. Given how susceptible humans are to this
> false ascription of agency, it is not surprising that they get fooled when
> something acts in a complex way.
>
> But (most of) the people on this list know what’s under the hood! We know
> that LLMs are very good at pattern matching and completion, we know about
> the universal approximation theorem, we know that there is a lot of
> structure in the pattern of human-written text, and we know that humans are
> predisposed to ascribe meaning and intention even where there are none. We
> should therefore not be surprised that LLMs can produce text patterns that
> generalize well within-distribution but not so well out-of-distribution,
> and that when the former happens, people may be fooled into thinking they
> are speaking with a thinking being. Again, they were fooled by Eliza, and
> Eugene Goostman, and the Heider-Simmel illusion (ascribing emotion to
> animated triangles and circles)… and the rumblings of volcanos. But we know
> how LLMs and volcanos do what they do, and can explain their behavior
> without any additional assumptions (of thinking, or sentience, or
> whatever). So why add them?
>
> In a sense, we are like a bunch of professional magicians, who know where
> all of the little strings and hidden compartments are, and who know how we
> just redirected the audience’s attention to slip the card into our pocket…
> but then we are standing around backstage wondering: “Maybe there really is
> magic?”
>
> I think it’s not that machines have passed the Turing Test, but rather
> that we failed it.
>
> Paul Cisek
>
>
> *From:* Rothganger, Fredrick <frothga at sandia.gov>
> *Sent:* Thursday, March 16, 2023 11:39 AM
> *To:* connectionists at mailman.srv.cs.cmu.edu
> *Subject:* Connectionists: Can LLMs think?
>
> Noting the examples that have come up on this list over the last week,
> it's interesting that it takes some of the most brilliant AI researchers in
> the world to devise questions that break LLMs. Chatbots have always been
> able to fool some people some of the time, ever since ELIZA. But we now
> have systems that can fool a lot of people a lot of the time, and even the
> occasional expert who loses their perspective and comes to believe the
> system is sentient. LLMs have either already passed the classic Turning
> test, or are about to in the next generation.
>
> What does that mean exactly? Turing's expectation was that "the use of
> words and general educated opinion will have altered so much that one will
> be able to speak of machines thinking without expecting to be
> contradicted". The ongoing discussion here is an indication that we are
> approaching that threshold. For the average person, we've probably already
> passed it.
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20230325/4bf19556/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image.png
Type: image/png
Size: 84960 bytes
Desc: not available
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20230325/4bf19556/attachment-0001.png>