tesseractcat's recent activity
-
Comment on OpenAI is a bad business in ~tech
-
Comment on Why AI isn't going to make art in ~arts
tesseractcat LinkGwern made a comment on this article I thought was pretty insightful: I find myself a bit frustrated when people associate AI capabilities with the corporate sanitized outputs of...Gwern made a comment on this article I thought was pretty insightful:
He's also, broadly speaking, wrong. The blandness he refers to has little to do with 'taking the average of the choices'. Base model output is not bland! It is often wacky, bizarre, hilarious, or unexpected. See the samples people have been generating with the new Llama base model, or just again, go back to my old GPT-3 samples - samples Chiang should've seen since he's been pontificating regularly about LLMs this whole time. (In considerable part because there is usually no such thing as 'taking the average of the choices' in text: what is the 'average' of two likely predictions like 'a' or 'b'? There is no such thing as a letter 50% of the way in between 'a' and 'b', the sampling process has to pick one or the other, and then it's committed.)
The blandness he's referring to is the effect of the preference-learning/instruction-tuning made infamous by ChatGPT, which has no simple analogy but has little to do with "the choices that other writers have made, as represented by text found on the Internet" (in fact, this is the opposite of correct, as making the outputs unlike 'text found on the Internet' is most of the point). The point of the tuning is to collapse the LLM, which starts off as an extraordinary mimic of every style, onto a narrow range of styles & vocabs - yielding the ChatGPTese we know all too well now.
I find myself a bit frustrated when people associate AI capabilities with the corporate sanitized outputs of ChatGPT/Dall-E/Bing Image/etc. Although it's not that surprising, considering that's the only way the average person will interact with AI.
If you're interested in playing around with more authentic AI, I'd suggest a non-finetuned version of Llama, or Stable Diffusion (although the newer versions of Stable Diffusion are a bit sanitized as well, 1.5 is probably the best for wackier outputs). Prompting will be harder, but you won't get outputs that sound/look like AI.
-
Comment on Grokking KOReader in ~books
tesseractcat Link ParentFor me the main thing is the ability to use regular folders to sort files. For the life of me I don't understand why none of the mainstream e-readers let me just use folders to sort books (they...For me the main thing is the ability to use regular folders to sort files. For the life of me I don't understand why none of the mainstream e-readers let me just use folders to sort books (they all want you to manually add everything to collections).
Other than that it's nice to be able to launch an SSH server to remotely transfer files. And the reading stats are pretty thorough. Also, although it's kind of niche, when reading manga with a bunch of volumes, KOReader will give you an option to open the next volume once you finish the current one, rather than needing to go back to the file explorer and manually find the next one. Oh, and KOReader will give you an option to discard the embedded style in a book if you don't like it (which I often have to do since certain books have weird indentation/paragraph styles).
Basically it's just a bunch of little QOL things that add up.
-
Comment on Looking for a decent, but cheap-ass tablet in ~tech
tesseractcat Link ParentI disagree, ipads tend to retain their resale value a lot better than android tablets. In this case, that means it's a lot better value to get a used android tablet than a used ipad. The (older)...I disagree, ipads tend to retain their resale value a lot better than android tablets. In this case, that means it's a lot better value to get a used android tablet than a used ipad. The (older) Samsung flagship tablets in particular are quite cheap now used and are great for the described use case.
-
Comment on ‘Furiosa’ fires up Cannes with six-minute standing ovation for Anya Taylor-Joy and teary Chris Hemsworth in ~movies
tesseractcat Link ParentIt's probably better to watch Fury Road first, wait a bit, and then watch Furiosa. Although if you were committed to a marathon it would be cool to watch Furiosa -> Fury Road, though I don't think...It's probably better to watch Fury Road first, wait a bit, and then watch Furiosa. Although if you were committed to a marathon it would be cool to watch Furiosa -> Fury Road, though I don't think that would be the best first viewing.
-
Comment on Tiny Glade's castle-doodling demo is packed with delightful little reactive surprises in ~games
tesseractcat LinkI'm happy to see Tiny Glade will be joining the Gnorp Apologue in the emerging category of games written in Rust. From what I've read it uses bevy's ECS combined with a custom renderer.I'm happy to see Tiny Glade will be joining the Gnorp Apologue in the emerging category of games written in Rust. From what I've read it uses bevy's ECS combined with a custom renderer.
-
Comment on Noam Chomsky: The false promise of ChatGPT in ~tech
tesseractcat Link ParentThat's true. It's kindof an apples to oranges comparison. LLMs are an attempt to bootstrap intelligence off of humanity without needing to go through the whole billion-year evolution gauntlet. So...That's true. It's kindof an apples to oranges comparison. LLMs are an attempt to bootstrap intelligence off of humanity without needing to go through the whole billion-year evolution gauntlet. So they definitely have more initial structure imposed on them than the first micro organisms (which might truly be called blank slates). Given the complexity of the human body though, I suspect that the amount of data imposed on an untrained LLM is much less than that of a baby.
-
Comment on Noam Chomsky: The false promise of ChatGPT in ~tech
tesseractcat Link ParentExactly, babies have so much behavior and knowledge from birth it's surprising people still keep comparing them to untrained neural nets.Exactly, babies have so much behavior and knowledge from birth it's surprising people still keep comparing them to untrained neural nets.
-
Comment on Noam Chomsky: The false promise of ChatGPT in ~tech
tesseractcat Link ParentWhether or not the human mind is processing a lot of information, there's a deeper issue with this line of argument. Since human babies are born helpless, it's easy to think of them as a blank...Whether or not the human mind is processing a lot of information, there's a deeper issue with this line of argument. Since human babies are born helpless, it's easy to think of them as a blank slate, but they're not. Humans have been "trained" by millions of years of evolution, the equivalent of probably exabytes worth of "data". It's difficult to come up with a perfect analogy, but I think it's more accurate to consider a baby the equivalent of a trained model, and to consider any learning that happens after birth like fine-tuning, or perhaps in-context learning (or a mixture of both).
-
Comment on Jet Lag | Season 9 trailer : Hide and Seek across Switzerland in ~hobbies
tesseractcat Link ParentSpoiler Overall while I've been enjoying it, I think that it's one of the more boring seasons. My main issue is that the curses haven't been used enough. Basically all the curses have been pretty...Spoiler
Overall while I've been enjoying it, I think that it's one of the more boring seasons. My main issue is that the curses haven't been used enough. Basically all the curses have been pretty trivial and the hiders only ever roll once or twice. I think they should have made curses cheaper but less powerful overall, with a few rare good ones, and in addition the seeking abilities should have been a bit more powerful, considering how long it's taking to find the hiders on average.
Also it seems a bit unfair that they don't all get an equal number of opportunities to hide.
-
Comment on Learning and studying things without taking notes in ~life
tesseractcat LinkAs someone who takes notes: It's less about remembering things in the moment, and more about being able to recall the knowledge later (months/years down the line). Otherwise I'll just forget. This...As someone who takes notes: It's less about remembering things in the moment, and more about being able to recall the knowledge later (months/years down the line). Otherwise I'll just forget. This is also why I take digital notes despite studies that show that handwritten notes are better for memory.
-
Comment on It’s OK to call it Artificial Intelligence in ~tech
tesseractcat Link ParentI'm very disheartened to hear that you believe that I've been arguing in bad faith. Since you stated you don't wish to continue the argument, I'll end this with what I believe is the fundamental...I'm very disheartened to hear that you believe that I've been arguing in bad faith.
Since you stated you don't wish to continue the argument, I'll end this with what I believe is the fundamental disagreement I have: AI as a field cannot distill what AlphaGo does into a set of strategies that can then be applied to a traditional Go bot (except, maybe, by treating it as a black box). We can't look at the weights and biases to see how it models the Go board, or how it identifies patterns. I believe this represents a fundamental lack of understanding of these deep neural networks.
It concerns me greatly when people say we do understand them, because it is like someone saying we do understand how the brain (the brain of any species) works. We may understand the lower level structures, and the very highest level, but we have a tenuous grip on the middle level, where the thinking happens.
This is a hard thing to debate because the terminology is confusing. In some ways, yes, we do know exactly how deep neural networks work. We know lots about how to train them, exactly how to run them, what the GPUs are doing, the precise matrix multiplications needed. Understanding these digital neurons doesn't let us play Go though.
-
Comment on It’s OK to call it Artificial Intelligence in ~tech
tesseractcat Link ParentMy main goal with the above post was to counterbalance a tendency I see where people are quick to dismiss any possibility of LLMs having consciousness, not to argue for it. As for my personal...My main goal with the above post was to counterbalance a tendency I see where people are quick to dismiss any possibility of LLMs having consciousness, not to argue for it. As for my personal beliefs, I lean towards panpsychism, where all systems have some sort of qualia/awareness, even if alien to us. In general I dislike the term consciousness, and prefer qualia. I do think defining consciousness in terms of awareness is a bit circular.
-
Comment on It’s OK to call it Artificial Intelligence in ~tech
tesseractcat Link ParentSorry about that, I think it's a bad habit I got from the 'reply' feature on various chat apps.Sorry about that, I think it's a bad habit I got from the 'reply' feature on various chat apps.
-
Comment on It’s OK to call it Artificial Intelligence in ~tech
tesseractcat (edited )Link ParentFirst, I don't think agency is a necessary condition for intelligence, so I don't think this would really prove it one way or the other, but no intelligences are entirely self-bootstrapped. Human...When I say they have no agency, I mean it in the sense that they can't initiate anything, or want or prefer anything. I would look at the Belmont report as a good starting point on the ethics of agency.
First, I don't think agency is a necessary condition for intelligence, so I don't think this would really prove it one way or the other, but no intelligences are entirely self-bootstrapped. Human intelligence is fed 'input data' through it's genetics, evolution, and upbringing.
As far as how they work, we know tremendous amounts about the neural networks that compose them, how learning works, the algorithmic implicit differentiation, and how the learning works.
These are all things other than how they actually work when running inference. We know quite a bit on how to water them, what humidity they need, and the correct scaffolding to grow optimally, but little on how they actually bloom.
Similarly, techniques like targeted knowledge injection don't work by interfacing with the deep layers of the network, where the actual intelligence may or may not be, it works by interfacing with the input and output layers.
And regarding philosophy, that gives us the framework to assess anything that might present as sentient.
Has philosophy closed in on a consensus for the definition and process for determining sentience? If so I would love to learn more.
-
Comment on It’s OK to call it Artificial Intelligence in ~tech
tesseractcat Link ParentI'm not arguing something does exist, I'm arguing it's impossible to know, and therefore impossible to make a definitive claim of it's nonexistence. You're right, proving a negative is impossible....I'm not arguing something does exist, I'm arguing it's impossible to know, and therefore impossible to make a definitive claim of it's nonexistence. You're right, proving a negative is impossible. Similarly, if we assume "that the thing/action/whatever does not exist until proven", this is true for all consciousness except for our own subjective experience.
-
Comment on It’s OK to call it Artificial Intelligence in ~tech
tesseractcat Link ParentThe person I responded to is the one making the definitive claim of lack of consciousness, so they should be the one to precisely define it.The person I responded to is the one making the definitive claim of lack of consciousness, so they should be the one to precisely define it.
-
Comment on It’s OK to call it Artificial Intelligence in ~tech
tesseractcat Link ParentSource? It's my understanding we understand the 'architecture', but very little about how it actually accomplishes the goals we train it to do. They can do things with little to no direction, they...We know a great deal about how they work
Source? It's my understanding we understand the 'architecture', but very little about how it actually accomplishes the goals we train it to do.
They can't do things because they lack agency
They can do things with little to no direction, they may not do them well but I think the level of quality is different from being entirely unable.
Philosophy has centuries of work on the subject of consciousness
But very little work dealing with massive neural networks and their consciousness or lack thereof.
And they don't learn from there output being fed back in
They do learn (to some extent) within the context window, continuous learning is definitely an area of research though.
-
Comment on It’s OK to call it Artificial Intelligence in ~tech
tesseractcat Link ParentWe don't know how they work. We know how we trained them. It's more accurate to say an LLM is grown, rather than made. Maybe on the inside they are simple probabilistic engines, or maybe they have...These transformer networks are simple probabilistic engines working on tokenized text.
We don't know how they work. We know how we trained them. It's more accurate to say an LLM is grown, rather than made. Maybe on the inside they are simple probabilistic engines, or maybe they have a more complex world model, or maybe they do any number of other things.
They can't acquire knowledge because they can't do anything
They can do things.
they have no sense of self, or consciousness, sapience,
These are all subjective things that are impossible to measure.
sensory loop
LLMs do have a rudimentary sensory loop in the sense that their output gets fed back into them.
-
Comment on It’s OK to call it Artificial Intelligence in ~tech
tesseractcat Link ParentJust want to say that we don't really know enough about consciousness to say this for certain. The only proof we have that anything is conscious is that we experience it subjectively. The only way...LLMs are not sentient, and they are not conscious.
Just want to say that we don't really know enough about consciousness to say this for certain. The only proof we have that anything is conscious is that we experience it subjectively. The only way we can predict consciousness of other things is then by their similarity to us. LLMs are possibly the most similar artificial creations to humans (although that's not saying much), so therefore have a higher (but still low) chance of being conscious.
I think this article is interesting, but a bit myopic considering that it's basically missing the biggest part of OpenAI's strategy. They're hemorrhaging money because they're betting that they can train a model that is smarter. I think it's uncontroversial that the earnings potential increases the smarter the model gets (very few people would pay for a GPT-2 level model, for instance), so the only question is whether or not they can train a smarter model in time, and what people will pay for it at that point.
One objection is that even if they do manage to train a smarter model, at that point there will be no differentiator from other companies who will also train equivalently smart models. I think this is sort of true, but I wouldn't be surprised if OpenAI can exploit their lead, and then maintain it through network effects.