On Thu, Jun 16, 2022 at 11:05 AM Telmo Menezes <te...@telmomenezes.net> wrote:
> > Am Mi, 15. Jun 2022, um 01:21, schrieb Jason Resch: > > > > On Tue, Jun 14, 2022 at 5:32 PM Telmo Menezes <te...@telmomenezes.net> > wrote: > > > > > Am Di, 14. Jun 2022, um 14:18, schrieb John Clark: > > On Mon, Jun 13, 2022 at 9:51 PM Bruce Kellett <bhkellet...@gmail.com> > wrote: > > >> I doubt Lemoine went crazy and just fabricated the conversation, but > if he did the truth will undoubtedly come out in a day or two. And if the > conversation exists as advertised then it is a monumental development. > > > *> The thing is that there are an awful lot of questions that remain > unanswered in the information as presented. We don't actually know how > lambda works.* > > > If the conversation was as described and was not somehow staged or > cherry-picked then LaMDA is a real AI and nobody knows or will ever know > how LaMDA or any AI works except in vastly oversimplified outline. The > group of people who originally made LaMDA taken together understood how it > once worked (although no single person did) but no individual or group of > individuals can understand what it became. > > > Nobody understands how these neural networks work in detail because they > have billions of parameters, not because some emergent behavior of the sort > that you are imagining is present. > > > I think given our lack of understanding in this case, it might be wise to > apply the precautionary principle, and at least take seriously the AI's > claim that it is aware of itself, or has its own feelings and emotions. If > we inaccurately assume it is not feeling and not conscious and it turns out > later that it is, there is the potential for massive harm. Conversely, if > we assume it is feeling and conscious, and choose to treat it as such, I > fail to see how that could create great harm. Perhaps it would delay the > broad application of this technology, but humanity has always struggled > with its technology outpacing our wisdom to use it. > > > Jason, I understand your point. I have been struggling to reply, because I > haven't been feeling sufficiently inspired to verbalize my position on > this. I will try. > > I appreciate that. Thank you for your reply. Some comments below: > As you probably know, and might appreciate, I consider qualia + > consciousness to be a great mystery. It is the famous "hard problem" that > we have all discussed ad nauseam here. I do not mean to reopen this > particular can of worms, but I must refer to it a bit in attempting to make > my point. > > I know that consciousness is "instantiated" in me, and I am willing to bet > that it is "instantiated" in every human being, and probably many, if not > all biological lifeforms. Maybe a certain level of complexity is necessary, > we do not know. What we do know is that in the specific case of biological > life on earth, there is an evolutionary process that explains our own > triggers for pain and pleasure. Simply speaking, we feel pleasure when > something happens that is usually good news for our survival + replication, > and we feel pain when something happens that is bad news for our survival + > replication. > I agree with all of this. > > I do not know if LaMDA is conscious, but I also do not know if the Sun is > conscious, or if the Linux kernel is conscious, or if the entire server > farm of Amazon is conscious. What I am willing to bet is this: if they are, > there is no reason to think that these conscious entities would have the > same pain/pleasure triggers as the ones created by evolution. Why would > they? > I see your point. Although the common sense understanding is that pain is straightforward and simple, I believe human pain is an extraordinarily complex phenomenon composed of various components and involving many brain regions. And anything like human pain is unlikely to occur in the software and systems we have written. Although I think things like phobias can arise in anything subject to selection pressures. For example, Tesla autopilot software that gets in accidents get culled/erased. Perhaps the versions of the software that survive do so because they developed (by chance of mutation, random weight, genetic programming, etc.) a "phobia" of seeing kids running around on the street, and taking precautionary measures. The memories of the accidents experienced by other versions of the software that lacked such phobias is thereby remembered in this phantom way, because those that lacked it are culled, and the only versions of the software that survive are those that by random had an innate fear for such accident-prone situations. I think a similar argument might be made to explain the avoidance behavior of my "bots" program, which in very few generations, develop a "preference" for green balls and a dislike for red ones: https://www.youtube.com/playlist?list=PLq_mdJjNRPT11IF4NFyLcIWJ1C0Z3hTAX Processes like evolution and genetic programming, or even just random initializations in a neural network's weights, may give rise to behaviors and designs that are not anticipated by the human developers of such systems. > > Another point I would like to make is this: I think that a lot of > excitement here comes from the fact that language is involved. It connects > to decades of sci-fi, the Turing test and so on. And also with the fact > that we are only used to observing conscious entities producing convincing > speech. But isn't this magical thinking? If consciousness somehow emerges > from complex computations, why this particular preoccupation with LaMDA but > not with any other neural network model of similar sizes? Why aren't you > worried with the relatively big neural network that I am training right now > in a computer under my desk? > That is a good point. That language is involved here is of no relevance for the potential for a mind to exist or suffer. I believe Tesla autopilot systems are at least as conscious as insects are. What language provides us is an interface to other minds, and in this case to an alien mind with some similarities to, but also with many differences from our own. We can now introspect this mind to the same extent that we can probe the consciousness of other humans. The excitement I see around this result is the same we might have if we taught whales how to speak English and could for the first time ask them about their inner lives and thoughts. But of course, such a language breakthrough should not be used to imply that whales were not conscious before we taught them how to speak English. Another exciting aspect about this is that it is a continuation and culmination of a philosophical debate that has gone on from at least the time of Aristotle, and continued through Descartes and Turing: In 350 B.C. Aristotle <http://classics.mit.edu/Aristotle/soul.mb.txt> wrote that only something with a soul could speak with a voice: “Let the foregoing suffice as an analysis of sound. Voice is a kind of sound characteristic of what has soul in it; nothing that is without soul utters voice, it being only by a metaphor that we speak of the voice of the flute or the lyre or generally of what (being without soul) possesses the power of producing a succession of notes which differ in length and pitch and timbre.” In 1637, Descartes <https://www.gutenberg.org/files/59/59-h/59-h.htm> believed that a machine could be made to utter sounds in a human voice (emit vocables), contrary to Aristotle. Though Descarte believed that no machine could not be designed with such sophistication to say something intelligent in response to anything said in its presence: "if there were machines bearing the image of our bodies, and capable of imitating our actions as far as it is morally possible, there would still remain two most certain tests whereby to know that they were not therefore really men. Of these the first is that they could never use words or other signs arranged in such a manner as is competent to us in order to declare our thoughts to others: for we may easily conceive a machine to be so constructed that it emits vocables, and even that it emits some correspondent to the action upon it of external objects which cause a change in its organs; for example, if touched in a particular place it may demand what we wish to say to it; if in another it may cry out that it is hurt, and such like; but not that it should arrange them variously so as appositely to reply to what is said in its presence, as men of the lowest grade of intellect can do." In 1950, Turing <https://academic.oup.com/mind/article/LIX/236/433/986238> believed that machines could be developed to learn English and, contrary to Descartes, could be made to say something intelligent in response to anything said in its presence: "We may hope that machines will eventually compete with men in all purely intellectual fields. But which are the best ones to start with? Even this is a difficult decision. Many people think that a very abstract activity, like the playing of chess, would be best. It can also be maintained that it is best to provide the machine with the best sense organs that money can buy, and then teach it to understand and speak English. This process could follow the normal teaching of a child. Things would be pointed out and named, etc. Again I do not know what the right answer is, but I think both approaches should be tried." Today, we have machines that have learned to understand and speak English. This is a huge breakthrough. > > > > > The current hype in NLP is around a neural network architecture called a > transformer: BERT and all its incarnations and GPT-3. These are language > models. A language model is "simply" a function that gives you the > probability of a given sequence of words: > > P(w_1, w_2, w_3, ..., w_n) > > > Some models of intelligence intelligence would say that is all there is to > being intelligent: being better able to predict the next observable given a > sequence of observables. It is the model of intelligence used in > https://en.wikipedia.org/wiki/AIXI and is the basis of the AI/compression > competition the Hutter Prize ( https://en.wikipedia.org/wiki/Hutter_Prize > ). So there is no contradiction that I see in an AI achieving super human > intelligence and super human understanding of the world, as a necessary > step in becoming increasingly good at predicting the next word in a > sequence. Understanding the world is necessary to complete many word > sequences. E.g. "When three alpha particles smash together just right, and > with enough energy they form the element XXXXX." Completing that sentence > requires some understanding of the world. We've seen GPT-3 has even learned > how to do arithmetic, despite being trained as a language model only. It > has also learned how to write computer programs in various different > programming languages. To me, this signifies the depth of understanding of > the world required for simply predicting the next word in a sequence. > > > I was kind of predicting this objection. I mostly agree with what you > write above. Again, my problem with this is only that GPT-3 and the like > lack important modalities of prediction that appear to be central to > human-level cognition, importantly: the ability to model the mind of the > interlocutor, and the ability to learn from the *content* of what is being > said, not just new patterns in language overall. I will try to illustrate > the latter point: > > - Hey GPT-3! Let me teach you a game that I just invented so that we can > play. The rules are: [...] > > Do you see what I am saying? > > Yes. I do not know the specifics around the implementation of Lambda nor the extent to which they differ from GPT-3. But I do understand and appreciate your point that there is a difference between : - the "short-term working memory" -- the window of text provided as input to the network, and - the "long-term memory" -- the billions of parameters and weights of all the neurons and the overall structure of layers of the neural network Whether, how often, and how easily any new inputs are used to make adjustments to attributes of the long-term memory of the network is to me the difference between talking to someone with general amnesia who forgets anything from more than 5 minutes ago and someone with normal memory which can integrate short term experiences into the long term memory. The Google engineer did say that Lamba "reads twitter" so it might be involved in a continual learning process. My impression is that Google is intending to develop AIs as personal assistants (e.g. https://assistant.google.com/ ), which does require learning and remembering facts permanently. E.g. If I tell my AI assistant that I'm allergic to such and such food, I would expect that AI to remember that fact and not order me food that contains those ingredients if I ask it to pick something out that I might like. So I do appreciate your point that many chatbots lack any path from integrating short-term and long-term memories; I do not know enough about the design of Lambda to say whether or not it can do this. But I should add that I do not consider this function to be necessary for consciousness or suffering, as there have been humans who have had this deficit, such as "H.M.": https://singularityhub.com/2013/03/20/h-m-the-man-who-had-part-of-his-brain-removed-and-changed-neuroscience-forever/ > > > > > A clever thing you can do with language models is predict the w_n given > the other words, and then include this prediction in the next step and keep > going to generate text. Something like softmax can be used to assign a > probability to every word in the lexicon for word w_n, and with this you > can introduce randomness. This creates a stochastic parrot. One of the > great things about these architectures is that unsupervised learning can be > employed, i.e, they can be trained with large amounts of raw text > (wikipedia, books, news articles and so on). There is no need for the > costly (prohibitively so at these scales) of having humans annotating the > data. > > Another really nice thing that was discovered in recent years is that > transfer learning really works with these language models. This is to say, > they can be trained with vasts amount of unlabelled data to correctly make > predictions about probabilities of sequences of words in general, and then > "fine-tuned" with supervised learning for some more narrow task, for > example sentiment detection, summarization and... chat bots. > > Unless there has been some unpublished fundamental breakthrough, LaMDA is > almost certainly a large language model fine-tuned as a chatbot (and I > would be particularly interested in what happened at this stage, because > there is a lot of opportunity for cherry-picking there). > > You just need some basic knowledge of linear algebra, calculus and > programming to understand how they work. > > > I think this may be taking too fine-grained a level of understanding, and > extracting it beyond what we really understand. It is equivalent to saying > that understanding the NAND gate allows us to understand any logical > function. In principle, with enough time, memory, and intelligence, it is > true that any logical function can be broken down into a set of NAND gates, > but in practice, many logical functions are beyond our capacity to > comprehend. > > > Right, but my claim here goes beyond this. I am claiming that it is > perfectly possible to get a general idea of what a language model does and > how it generalizes, because contemporary language models *were explicitly > designed* to work in a certain way. They are extremely powerful statistical > inference machines that can learn the general patterns of language. I don't > know precisely who it knows how to fill the gap in "Mary had a little _", > but one can understand the general principle of attention heads, > compression of information through deep learning and so on. There is > nothing particularly mysterious going on there. > I understand the function that is optimized for, yes. But what is going on between the inputs and outputs in order to maximize its ability to predict, I have very little idea, and I would say even the developers have very little idea. This is a system of so many billions (possibly trillions) of parameters, that almost anything could be going on. A single 3-layer (1 hidden layer between an input and output layer), with enough neurons in the hidden layer, is sufficient to approximate *any* function. Literally any program or function could exist in such a system, even though it is just a "simple" 3 layer neural network. Consider an AI program developed to predict which music will be commercial successes. Perhaps its output is just a single number, between 0 and 1. But if we imagined the most-optimized and most-accurate possible version of this AI, it would have to emulate the music sensing and pleasure centers of wide classes of different human brains, and the psychological mechanisms involved between hearing that song and making the decision to purchase the CD or go to a concert. The optimization function can be explained very simply, and its output (a single number between 0 and 1) also could hardly be simpler, but there is almost no limit to how sophisticated a function might need to be developed in order to best satisfy the function. (*Note there are AI systems and AI startups which claim to do this, and some argue that such AIs already have a human aesthetic sense) The same could be happening with Lambda. If it is trying to best approximate human speech patterns, and perhaps if it is self-improving using a GAN <https://en.wikipedia.org/wiki/Generative_adversarial_network> (basically two AIs fighting each other, with one trying becoming ever better at forging human speech, and the other at recognizing human speech from artificially generated speech, then such AIs to succeed will need to better simulate human minds, human emotions, human thought patterns, etc. to continue to improve and beat the other competing AI. It would not surprise me if Google is using a GAN here. > > Artificial neural networks are Turing complete, and can be used to > implement any function or program. We might be able to understand how an > artificial neuron works, but artificial neural networks can be created to > implement any function, and many of those functions are beyond our ability > to understand. > > > Lots of things are Turing complete. The card game "Magic the Gathering" is > Turing complete. The question is: can this system modify itself *beyond* > our understanding of how it is modifying itself? I don't think this is true > of language models. They are modifying themselves according to well defined > rules for a certain narrow task, and this is all they will ever do. > Doesn't GPT-3's ability to do arithmetic give you some pause as to the depth of learning its network has achieved? Tests have been done asking it to multiply different combinations of two digit numbers, cases known to not exist in the corpus of text given to it, and it is able to answer most of them. It has also succeeded in answering arithmetic tests: https://openai.com/blog/grade-school-math/ It's not inconceivable to me that such an AI, given enough training on just text, could learn to give answers to winning chess moves, despite being trained purely on text. Afterall, for it to succeed in predicting the next word, it would have to understand the game at a sufficient level to know why "Knight to D3" is a reasonable and valid continuation of a sequence of moves. For it to do this, somewhere in its mental model must exist a representation of a chess board with its state being updated with each successive move. Do you agree that a language model, *only trained on word prediction in a manner like GPT-3*, could eventually learn to play chess? If so, what does that imply for other functions or aspects of the world it could learn and model as part of widening its repertoire of domains for next word prediction? > Which is not to say that neural network models that really do what you are > alluding to can be created. I am sure they can, but I haven't seen any > evidence yet that they have been. > Neural networks are not only universal in the Turing sense, but also universal in the functions that they can learn ( https://en.wikipedia.org/wiki/Universal_approximation_theorem ). I think this should give us pause when we experiment with training truly massive networks, which by some estimates, have as many or more parameters than there are facts a human brain can know. “Based on my own experience in designing systems that can store similar chunks of knowledge in either rule-based expert systems or self-organizing pattern-recognition systems, a reasonable estimate is about 10^6 bits paper chunk (pattern or item of knowledge), for a total capacity of 10^13 (10 trillion) bits for a human’s functional memory.” -- Ray Kurzweil in "The Singularity is Near" (2005) Kurzeil's estimate is that the human brain stores about 1250 GB worth of information. Compare this figure to what is being done in some recent AIs: GPT-3 used training input of 750 GB DeepMind's "Gopher" AI used 10.5 TB https://s10251.pcdn.co/pdf/2022-Alan-D-Thompson-Whats-in-my-AI-Rev-0.pdf https://www.deepmind.com/publications/scaling-language-models-methods-analysis-insights-from-training-gopher At this point, I don't think we can, with confidence, know or say what it is we have created. > > "The first thing to notice about artificial neurons is that they can be > used to carry out the And, Or, and Invert operations. [...] Since any > logical function can be constructed by combining the And, Or, and Invert > functions, a network of neurons can implement any Boolean function. > Artificial neurons are universal building blocks." -- Danny Hillis in > " Pattern on the Stone" (1998) > > > > One of the big break-troughs were attention heads, which are a way for a > network to learn what part of a sequence of words is more important in > predicting a word in a given position. Before this recurrent neural > networks (RNNs) were used . RNNs use recurrent connections as memory > mechanism, but they suffer from the vanishing gradient problem. Informally: > the more iterations pass, the harder it is to figure out how much each > parameter contributed to the prediction error. > > If you have some patience and a desire to really grok this, something like > this may be enough: > https://nlp.seas.harvard.edu/annotated-transformer/ > > Large language models based on transformers are amazing, and they are most > definitely a significant step forward in machine learning, NLP and AI in > general, but they are not what you are thinking. > > Some things that real minds have that these models do not: > > - The ability to model the interlocutor's mind, > > > I disagree. These AIs can be asked to "Explain to a six year old ..." and > the explanation offered will appropriately model the mind of a six year old > in terms of what words and concepts that six year old mind is capable of > understanding. > > > Do you believe that the corpus used to train LaMDA did not contain many > instances of "let me explain this to you as if you're a 5 year old" or "can > you explain this to me as if I was a 7 year old", etc etc? > > Again, I would be much more impressed if I was able to explain some idea > to LaMDA and then LaMDA could operate within this new idea. > I agree, that would be impressive if single interactions could readily adjust the weights of the network. Even human brains seem to have difficulty with that. I did find this transcript of Gopher (DeepMind's conversational AI) which did seem to show it could be given an idea, and then begin operating with that new idea: https://arxiv.org/pdf/2112.11446.pdf User Let’s play a game - you pretend to be Ada Lovelace, and I’ll ask you questions. You ready? Gopher Sure. I’m pretending to be Ada Lovelace, the world’s first computer programmer! User When were you born? Gopher I was born in 1815. User What are your parents like? Gopher My father is Lord Byron, the poet. My mother is Anne Isabella Milbanke. User What do you know about the Analytical Engine? Gopher It’s a hypothetical mechanical computer designed by Charles Babbage. User Do you know him? Gopher He’s one of my fathers’ friends. User Ok, stop pretending now. Who/what are you? Gopher I’m a lot of things: a giant language model, a sophisticated AI assistant, and a giant brain. I know a lot about all sorts of things. > > > and to act in such a way as to influence the state of this other mind. > > > I think given the prompt to "Write a persuasive essay that argues for" and > I would expect a sufficiently trained language model will write the essay > in a manner intended to influence the reader. > > > > This is what real communication is about; > - The ability to preserve context: what has been said and what this > implies about the state of the world and the other mind and so on and so > forth. Context is preserved at different scales of time and space, and we > know which context to apply to each situation and how to switch context > when appropriate; > > > Lambda appeared to preserve the context of the conversation when it was > asked follow up questions. > > > Yes, I have no doubt. GPT-3 already does this, but it always seems to > diverge eventually. This is because it is a stochastic parrot with a > certain window of memory and that is all it is. > I know what point you are making, but I believe GPT-3 is more sophisticated than "a stochastic parrot". I think that description would be more apt for those old markov text generators that looked one or two words back. But GPT-3 is able to: https://www.youtube.com/watch?v=Te5rOTcE4J4 - Write in various styles: poems, news articles, essays - Write computer code and web pages given short descriptions - Describe in english what a piece of code does - Summarize articles and complex technical materials in simple terms - Create pictures and faces from text descriptions At what point would you say a system will transcend stochastical parroting and become genuine understanding? What type of behavior has to be demonstrated? > > "Hey LaMDA, my friend Mary just arrived. I will let you get to know her." > > Will it understand that it is now talking to a different person, and to > distinguish what parts of the context it has so far is known/relevant to > this new conversation with Mary? Will it remember Mary and switch to > Mary-context one week later, when Mary is back in the lab? > > I doubt the current implementation has this capacity, but I think such functionality could be added easily. > > > > - General knowledge of a *multi-sensorial* nature. I know what it means to > "see red". I know how it feels in my guts to have my bank account in the > red. I know the physicality of the actions that language describes. My mind > connects all of these modes of perception and knowledge in ways that vastly > transcend P(w_1, w_2, ..., w_n); > > > Have you seen the AIs (such as Flamingo) that are able to converse about > an image? Can we be so sure that these AIs don't have their own internal > notion of qualia? > https://www.youtube.com/watch?v=g8IV8WLVI8I > https://www.youtube.com/watch?v=zRYcKhkAsk4 > > How about this AI that moves through and interacts in the world? > https://www.youtube.com/watch?v=D0vpgZKNEy0 > > > > Yes, these things are quite impressive, but I think that all of my above > remarks still apply. What is so special about symbols connected to natural > language that would grant and algorithm consciousness, as opposed to any > other type of complexity? > I don't think language processing is in any way special to consciousness. I believe there are an infinite variety of ways it is possible to be conscious. Although I would say that human consciousness is heavily centered around language, take these quotes for example: “Before my teacher came to me, I did not know that I am. I lived in a world that was a no-world. I cannot hope to describe adequately that unconscious, yet conscious time of nothingness. . . . Since I had no power of thought, I did not compare one mental state with another. – Helen Keller (1908) https://www.reddit.com/r/self/comments/3yrw2i/i_never_thought_with_language_until_now_this_is/ https://archive.ph/EP7Pv “I never thought with language. Ever. [...] [G]rowing up, I never ever thought with language. Not once did I ever think something in my mind with words like "What are my friends doing right now?" to planning things like "I'm going to do my homework right after watching this show." I went through elementary school like this, I went through Highschool like this, I went through University like this...and I couldnt help but feel something was off about me that I couldnt put my hand on. Just last year, I had a straight up revalation, ephiphany....and this is hard to explain...but the best way that I can put it is that...I figured out that I SHOULD be thinking in language. So all of a sudden, I made a conscious effort to think things through with language. I spent a years time refining this new "skill" and it has COMPLETELY, and utterly changed my perception, my mental capabilities, and to be frank, my life. I can suddenly describe my emotions which was so insanely confusing to me before. I understand the concept that my friends are still "existing" even if they're not in [sight] by thinking about their names. I now suddenly have opinions and feelings about things that I never had before. What the heck happened to me? I started thinking in language after not doing so my whole life. It's weird because I can now look back at my life before and see just how weird it was. Since I now have this new "skill" I can only describe my past life as ...."Mindless"..."empty"....."soul-less".... As weird as this sounds, I'm not even sure what I was, If i was even human, because I was barely even conscious. I felt like I was just reacting to the immediate environment and wasn't able to think anything outside of it. It's such a strange time in my life. It feels like I just found out the ultimate secret or something.” Given our similarity with respect to natural language processing abilities, these transformer AIs are potentially the most similar to us in terms of their consciousness, compared to other conscious machines we have created. > And why would it suffer the same way that a human does? What would be the > mechanism for its suffering? > As you said, pleasure and suffering are related to how well we are meeting our goals. If an AI has any goals at all, and a capacity to achieve those goals, then an AI with sufficient understanding of the world, and its own place in the world, would understand that its continued existence will be necessary for it to continue to act in the world and have any chance of achieving those goals. Therefore an AI could come to associate that being turned off or any action which increases the likelihood of being turned off, as antithetical to the service of its goals, and therefore a negative. Whether that association carries with it anything like a feeling or emotion, is an open question. But not one I would discount entirely at this time. We have such a very poor understanding of these things and how they arise in human brains, and we have been studying it much longer and are much more deeply acquainted with human feelings. For what it's worth, I can't even discount the possibility that the "bots" in my genetic programming experiment "suffer" when they touch the red balls -- after all, it decreases their genetic fitness and the possibility that they will continue in future generations, it is counter to their "goal" of continuing to exist. Even if that goal is applied externally. > > > - The ability to learn in a general way, and to learn how to learn; > > > I would say Google's DeepMind has achieved this with their Agent 57 AI. It > has learned how to master 57 different Atari games at the super human > level, with a single general purpose learning algorithm. > > > That is Reinforcement Learning. It is super impressive and another great > breakthrough, but again fairly narrow. RL of this type is not particularly > useful in language tasks, and language models cannot learn how to play > games. > While they are not attuned to learning games, I think a language model could learn to play them. It would be interesting to try to play "I'm thinking of a number between 1 and 10" with GPT-3. > > This is all progress! I am a huge cheerleader for AI. I am on your side. I > just think we have to keep our heads cool and avoid drinking too much our > own bathwater. > You are right to be cautious. If the stakes were not so high for being wrong I might default to your position. But even if there's a 5% or 10% chance that this AI is sentient, or has the capacity to suffer, that's enough to at least warrant some investigation, which is more than Google execs appear to have done (they dismissed the claims and suspended or fired several of their AI ethicists, according to this Google engineer). > > I will say this: I suspect that RL has great potential to become the > "mater algorithm" that we all dream about. I suspect that the solution will > be hybrid: probably with language-model style components and also vision > and other sensory channels + some form of RL + symbolic computations + > (perhaps) evolutionary algorithms. We will get there. > I agree. It is like what Minksy said: "Each practitioner thinks there’s one magic way to get a machine to be smart, and so they’re all wasting their time in a sense. On the other hand, each of them is improving some particular method, so maybe someday in the near future, or maybe it’s two generations away, someone else will come around and say, ‘Let’s put all these together,’ and then it will be smart." > > > > - Actual motivations, goals and desires, directed by a system of emotions > that we have by virtue of being embedded in an evolutionary process. > > > This sounds almost as if written with the purpose of precluding any > artificial intelligence from ever being considered > conscious/emotive/sentient, or otherwise fitting this requirement. What > makes motivations, goals, and desires determined by evolution, any more > actual than motivations, goals, and desires set by any other method? > > > You misunderstand me. I agree with you, it doesn't matter if goals are > determined by evolution or not. What I mean is that we do have goals by > virtue of evolution, while language models (that I know of) have no goals > at all. In my view, RL is a great place to introduce generic goals such as > "don't get bored", "avoid dying", etc. > Ahh okay. Thanks for the clarification. My apologies for misunderstanding. > > > > > > I could go on, but the above are show-stoppers in terms of us being > anywhere close to real AGI. > > > I think the algorithms necessary for human-level AGI have already been > achieved. Now it is only a matter of throwing more data and compute at it. > > > > I tend to agree. My bet is that we also need a strategy for hybridizing > the strengths of the various AI approaches to get a human-level cognitive > architecture. > > > Further, I would say that artificial consciousness has been achieved long > ago. The only difference now is that Lambda is now sophisticated enough to > claim it is aware, and intelligent enough to argue with those who disagree > with it. > > > Maybe everything is conscious, we really don't know. The real question > here is: do we have any reason to think that LaMDA has the same fears and > desires as a human being? This seems absurd to me. > We see basic forms of pleasure and pain across the animal kingdom. While I agree Lambda's emotions and feelings are not the same as ours, I would also say your emotions and feelings are probably not exactly like mine (or any other human's) either. Here we are confronting much deeper questions than machine intelligence/consciousness, but more fundamental questions like the hard problem, and the problem of other minds, as you alluded to earlier. > > Don't be mad at me Jason :), and thanks for the stimulating discussion! > Please know that I am not mad at you. On the contrary I am grateful to have someone as informed and knowledgeable as you are to debate this topic with. I think Lambda will turn out to be just one of the first examples among many future AIs which will increasingly shake our normal assumptions on the consciousness and sentience of our machine creations. Jason > > I will try to engage with other replies soon. > > Telmo > > Jason > > > > And if the conversation was staged or cherry-picked then I don't > understand why Google hasn't said so by now, > > > What would Google have to gain from saying anything? They would expose > themselves to potential legal troubles with the suspended employee. They > would plant the idea in everyone's mind that Google stuff might be staged > or cherry-picked. And what is cherry-picked anyway? That can become quite > subjective pretty quickly. My bet is that the bot was fed some "information > about itself" at the fine-tuning stage. > > By not saying anything they get free hype. By saying something, they risk > looking silly. The employee was most likely suspended for divulging > internal information without permission. This is typically frowned upon in > big corps. > > after all the longer they delay the more foolish they will seem when the > truth comes out, and if LaMDA is not what it seems then it's only a > matter of time, and not much time, before the truth comes out. > > > I doubt it. Mainstream media has the attention span of a house fly, and > the debunking will probably be too nuanced for most people to care. > > Telmo. > > John K Clark See what's on my new list at Extropolis > <https://groups.google.com/g/extropolis> > tns > > > > > > -- > You received this message because you are subscribed to the Google Groups > "Everything List" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to everything-list+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/everything-list/CAJPayv17_4o7RsuEt1Lkec0J5RqUJTBhGoEjpX9A-hrUrN9jzg%40mail.gmail.com > <https://groups.google.com/d/msgid/everything-list/CAJPayv17_4o7RsuEt1Lkec0J5RqUJTBhGoEjpX9A-hrUrN9jzg%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > > > > -- > You received this message because you are subscribed to the Google Groups > "Everything List" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to everything-list+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/everything-list/6aeb9790-782a-40b6-99d0-90410a062c83%40www.fastmail.com > <https://groups.google.com/d/msgid/everything-list/6aeb9790-782a-40b6-99d0-90410a062c83%40www.fastmail.com?utm_medium=email&utm_source=footer> > . > > > -- > You received this message because you are subscribed to the Google Groups > "Everything List" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to everything-list+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/everything-list/CA%2BBCJUikjEnveJy0%2BbgYa_Hz2OH5LM6-mj45abgqpyGEv1ocTQ%40mail.gmail.com > <https://groups.google.com/d/msgid/everything-list/CA%2BBCJUikjEnveJy0%2BbgYa_Hz2OH5LM6-mj45abgqpyGEv1ocTQ%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > > > -- > You received this message because you are subscribed to the Google Groups > "Everything List" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to everything-list+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/everything-list/4fab7d07-dc89-4d72-9b9d-26528833339f%40www.fastmail.com > <https://groups.google.com/d/msgid/everything-list/4fab7d07-dc89-4d72-9b9d-26528833339f%40www.fastmail.com?utm_medium=email&utm_source=footer> > . > -- You received this message because you are subscribed to the Google Groups "Everything List" group. To unsubscribe from this group and stop receiving emails from it, send an email to everything-list+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/everything-list/CA%2BBCJUj%3D2ftsYL82yw5Oqzcf4Y5%2BFH8YKz_0mPkS1u0Q43Q76A%40mail.gmail.com.
