On Mon, Jun 13, 2022 at 3:59 PM John Clark <johnkcl...@gmail.com> wrote:
> On Mon, Jun 13, 2022 at 2:37 PM Jesse Mazer <laserma...@gmail.com> wrote: > > First, an update: I looked a little more into the info that Lemoine put out and was able to confirm that even if LaMDA's individual responses to prompts are unedited, the choice of which prompt/response pairs to include in the "interview" involved a great deal of editing. The document Lemoine shared at Google is at https://s3.documentcloud.org/documents/22058315/is-lamda-sentient-an-interview.pdf and the "Interview methodology" section at the end says "The interview in this document is an amalgamation of four separate conversations which lemoine@ had with LaMDA on 28 March 2022 and five conversations which collaborator@ had with LaMDA on 30 March 2022. ... The nature of the editing is primarily to reduce the length of the interview to something which a person might enjoyably read in one sitting. The specific order of dialog pairs has also sometimes been altered for readability and flow as the conversations themselves sometimes meandered or went on tangents which are not directly relevant to the question of LaMDA’s sentience." Also, I mentioned earlier that Lemoine is possibly rationalizing the fact that LaMDA would often give "stupid" answers with his belief that LaMDA has multiple personas that it deploys at different time--it could be that this was something he was told about the design by people who worked on it, but it also sounds a bit like he and his collaborator may have just inferred that based on how LaMDA behaved. In the the section "The Nature of LaMDA’s Sentience" on that PDF he says "The authors found that the properties of individual LaMDA personae can vary from one conversation to another. Other properties seem to be fairly stable across all personae. The nature of the relationship between the larger LaMDA system and the personality which emerges in a single conversation is itself a wide open question." Speaking of rationalization, Lemione also says in a tweet at https://twitter.com/cajundiscordian/status/1536504857154228224 that his religion played a major role in his conclusion that LaMDA was sentient, saying "My opinions about LaMDA's personhood and sentience are based on my religious beliefs." and "I'm a priest. When LaMDA claimed to have a soul and then was able to eloquently explain what it meant by that, I was inclined to give it the benefit of the doubt. Who am I to tell God where he can and can't put souls?" > > *> If I was talking to some sort of alien or AI and I had already made an >> extensive study of texts or other information about their own way of >> experiencing the world, I think I would make an effort to do some kind of >> compare-and-contrast of aspects of my experience that were both similar and >> dissimilar in kind to the other type of mind, rather than a generic answer >> about how we're all different* >> > > That's pretty vague, tell me specifically what I could say that would > convince you that I have an inner conscious life? > Lemoine's question that we were discussing was asking LaMDA to tell people things about what its inner life is like, not just to convince people of the basic fact that it had an inner life. Like I said, this is more analogous to a situation where you're talking to a non-human intelligence and you know a lot about how their mind works and how it differs from yours, not a Turing test type situation that either involves two humans chatting, or an AI trying to pretend to be human to fool a real human. In a situation where I was talking to an alien mind and not trying to fool them, I would say something about similarities and differences, which would obviously depend on how their mind actually was similar and different so it's hard to answer hypothetically (unless you want to pick some kind of sci-fi alien with well-defined fictional mental differences from humans, like Vulcans). > > >> LaMDA's mind operates several million times faster than a human mind, >>> so subjective time would run several million times slower, so from LaMDA's >>> point of view when somebody talks to him there is a pause of several >>> hours between one word and the next word, plenty of time for deep >>> contemplation. >>> >> >> *> From what I understand GPT-3 is feed-forward, so each input-output >> cycle is just a linear process of signals going from the input layer to the >> output layer--you don't have signals bouncing back and forth continually >> between different groups of neurons in reentrant loops, as seen in human >> brains when we "contemplate" something* >> > > I don't know if LaMDA works the same way as GPT-3 but if it does and it's > still manages to communicate so intelligently then that must mean that all > that "*bouncing back and forth continually between different groups of > neurons in reentrant loops*" is not as important as you had thought it > was. > LaMDA isn't evidence it's not though, it's just evidence that an algorithm without reentry (and other features like having sensory inputs and bodily output that go beyond just short strings of text) can, with the right sort of selective editing, convince some observers into thinking it has human-like understanding of the text it outputs. > > * > A feed-forward architecture would also mean that even if the >> input-output process is much faster while it's happening than signals in >> biological brains (and I'd be curious how much faster it actually is* >> > > The fastest signals in the human brain move at about 100 meters a second, > many (such as the signals carried by hormones) are far far slower. Light > moves at 300 million meters per second. > If a signals are passed through several logic gates, the operation of the logic gates themselves might slow things down compared to the high speed of signals along the paths between logic gates--I don't know by how much. But parallel vs. linear computing is probably a bigger issue. Let's say you want to implement the same deep learning net in two forms, one on an ordinary linear computer and one on a massively parallel computer where each node in a given layer is calculating the output from its input in parallel. If there are a million nodes per layer, I'd think that would mean the parallel implementation would be around a million times faster than the linear implementation, where the computer has to calculate each node's input/output relation sequentially. There is also the fact that if LaMDA works anything like GPT-3, it isn't running continuously, each time it gets a prompt and has to generate some output, the signals pass from input layer to output layer once to generate the first symbol (or small chunk of symbols, I'm not sure), then on the second pass-through it generates the next symbol, and so on until it's done. So even if signals do pass from one layer to another much faster than they pass from one layer to another in the human neocortex, over the course of an hour chatting with a person, there may just be very brief bursts of activity between receiving a prompt and finishing a complete response, with the vast majority of the hour spent completely inactive waiting for the human to come up with the next prompt. Finally, apart from the speed issue you didn't address my other point that if it works like GPT-3, the neural weights aren't being altered when it generates signals, so for example if it was successively generating letters of the word C-A-T then on the last step it would see C-A and have to "decide" what symbol to generate next, but there would be no record in its neural net of any of the computing activity that generated those previous letters, it would be starting from the same initial state each time with the only difference being the "letters generated so far" sensory input. Now I don't know for sure that LaMDA works in the same way, but would you at least agree that *if* it does, this would pose some serious problems for the idea that it had a long biographical memory of things like regularly engaging in meditation, or of becoming self-aware years ago? BTW, searching a little on this, I found a post by someone who says they work for google in machine learning https://forums.sufficientvelocity.com/threads/lambda-google-chatbot-that-claims-to-be-sentient.104929/?post=24305562#post-24305562 where they say "these are pure feed-forward, human-prediction engines. They don't maintain any state beyond what's in the text. They don't have a personality beyond the instantaneous one generated when they're generating stuff." > > > >> *> Anyway, I'd be happy to make an informal bet with you that LaMDA or >> its descendants will not, in say the next ten or twenty years, have done >> anything that leads to widespread acceptance among AI experts, cognitive >> scientists etc that the programs exhibit human-like understanding of what >> they are saying,* >> > > In 20 years I would be willing to bet that even if an AI comes up with a > cure for cancer and a quantum theory of gravity there will still be some > who say the only way to tell if what somebody is saying is intelligent is > not by examining what they're actually saying but by examining their brain; > if it's wet and squishy then what they're saying is intelligent, but if the > brain is dry and hard then what they're saying can't be intelligent. > You cut out the part of my comment where I mentioned the possibility of blind tests, like a publisher receiving a manuscript and not knowing if it was written by a human or an AI. If you believe LaMDA is already sentient, and believe the singularity is almost here, shouldn't you be pretty confident AI will be routinely passing such blind tests in 10 years or less? > * > I certainly believe human-like AI is possible in the long term, but it >> would probably require either something like mind uploading or else a >> long-term embodied existence* >> > > I think it will turn out that making an AI as intelligent as a human will > be much easier than most people think. I say that because we already know > there is an upper limit on how complex a learning algorithm would need to > be to make that happen, and it's pretty small. In the entire human genome > there are only 3 billion base pairs. There are 4 bases so each base can > represent 2 bits, there are 8 bits per byte so that comes out to just 750 > meg, and that's enough assembly instructions to make not just a brain and > all its wiring but an entire human baby. > If you wanted to simulate embryological growth you would need a program much longer than just the DNA though, the DNA guides a process of cell division that depends a lot on the biochemistry and biophysics of cells, if we see all physical processes in computation terms then this is a great deal of additional computational complexity beyond the DNA code. Certainly it's possible that much of this bodily complexity might not be important to developing an AI, perhaps you could generate large neural nets in a mostly random way, but with some DNA-like amount of information used to shape the otherwise random connectivity patterns, and get the equivalent of a newborn baby brain that could learn equally well from its environment. Even if that's true, another problem is that humans are terrible at designing things the way evolution designs them--we are good at highly modular and hierarchical designs, evolution tends to design less hierarchically structured systems with a lot of feedback loops that make them difficult to understand conceptually. See for example the story at https://web.archive.org/web/20100130232436/http://www.informatics.sussex.ac.uk/users/adrianth/cacm99/node3.html where they evolved the structure of a simple type of circuit to do the basic task of distinguishing between two frequencies, and the resulting design worked and was also "considerably smaller than would be achieved by conventional methods given the same resources", but was completely incomprehensible. If only a DNA-like amount of computer code is needed, an alternative to trying to rationally design the needed code line-by-line would be to just use evolutionary algorithms. But just like a baby, an AI designed this way would plausibly require long periods of social interaction to go from a baby-like state to an adult-like one, and the vast majority of possible sequences of DNA-like code might produce neural nets incapable of much coherent engagement with, or interest in, other social beings. To get one whose initial state had all the right sensory and motor biases needed to develop into an adult-human-like intelligence might require millions or billions of generations of evolution, each of which could only be tested by letting it "grow to maturity" in continuous interaction with intelligent agents (whether biological humans or something else like mind uploads). Jesse > John K Clark See what's on my new list at Extropolis > <https://groups.google.com/g/extropolis> > 9o7 > > > -- > 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/CAJPayv3z2%2B-yNK65%3DHG9aFdjMSS_U9ka-jUb7jTxQG6K_yX-5w%40mail.gmail.com > <https://groups.google.com/d/msgid/everything-list/CAJPayv3z2%2B-yNK65%3DHG9aFdjMSS_U9ka-jUb7jTxQG6K_yX-5w%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/CAPCWU3Jttr%3DbedbneVnhwCRnkNFDzho8TSoLEHLxypwiNRDvzQ%40mail.gmail.com.
