Interesting perspective on GPT-3: https://twitter.com/sh_reya/status/1284545976892403714?s=09

Funny, I was just forwarding this thread to a friend: https://twitter.com/vgr/status/1284684717359419394?s=10

I don’t know enough AI but my friend after reading the paper is like there isn’t a fundamental breakthrough, just a bigger model + concerns it doesn’t scale past what it currently is. And comments like https://twitter.com/Jon_Ayre/status/1284417918260981760?s=20 . AI hype is to me like the boy the cried wolf tbh. Doesn’t mean this won’t be useful but I guess like they say time will tell 🙂

Exactly, it’s just generating from trained tokens and relationships, it doesn’t have an understanding of anything it’s writing.

[...] it doesn’t have an understanding of anything it’s writing.

That's an exceptionally high bar to set. We have no clue where to start from in building an artificial consciousness. Alan Turing attempted this and while his result was not artificial consciousness (as far as we know :P), he still gave us computers.

Can somebody explain to me what “understanding” means? So machines are basically just trained on patterns… how exactly are humans different…?

This will get philosophical, but so is your question. So, for me, understanding is just a sensation, a quale (https://en.wikipedia.org/wiki/Qualia). Some people have the sensation of understanding something, but they don't; some don't have it but they do. This already begs the question how do we know that they actually understand the right thing? How can anyone be a judge of that, and I presume the answer is Popper's scientific method — unless otherwise proven false, it's probably true (things can get more complicated than that, but I guess that's the gist of it). So in the end understanding is part of qualia, which we really don't know what they're made of.

This is something that is explored extensively in philosophy of mind, obvious with lots of different opinions on the issue. I think one useful starting point is Searle's Chinese Room argument. I'd really recommend actually reading the paper if you are interested in the topic. http://cogprints.org/7150/1/10.1.1.83.5248.pdf

I haven't read Searle's paper yet, but I found Scott Aaronson's discussion around it, and other similar ideas, facinating: Why Philosophers Should Care About Computational Complexity https://www.scottaaronson.com/papers/philos.pdf

Uh, I didn't want to derail this thread, but I do see that that's on me, as my question clearly was too philosophical. My intention was to point out that while GPT-3 seems "obviously inferior" in its current incarnation, things like that are often said just looking at the technology, but not looking at ourselves. We don't really know how understanding works. And I don't mean the philosophical angle (what understanding is), but the cognitive science perspective. Our theories about how the brain works have evolved quite a bit, even though there's still lots more left to discover, but I'm not so sure that our cognitive processes are so "obviously superior" to what GTP-3 does. That doesn't mean that I believe we're close to AGI either, I'm just trying to raise awareness that to compare machine capabilities to our own, we can't just look at technology alone, but also need to advance the understanding of ourselves.

I'm not sure what contrast you are making between the philosophical angle and the cognitive science angle. Those are rather connected. Understanding how the brain works more does not necessarily get us to understand how understanding works. Even if it did though, translating from the brains neural circuits with their causal mechanisms to software based syntatic simulation doesn't guarantee there is understanding. Searle's view is that those neuronal causal firing are necessary. In other words hardware based AI is possible, but not software. Or put anything way, no amount of syntatic manipulation gives you semantics. Of course others disagree. Lots of cognitive scientist/philosophers just don't believe there is anything like understanding at all. You might not be interested in any of this and that is okay. But I think these crucial if we want to create some ai with actual understanding.

When I look at GPT-3, I don't see a big step toward AGI. As ever, people are tripping over the fact that "AI" is a tragically poor term for this kind of technology, and are (IMO) wasting a lot of time asking what this means. Rather, what I see in GPT-2/3, and large-scale ML models generally, is an interesting new kind of tool that we have yet to find all the various uses for. It joins an esteemed family of breakthrough enabling technologies like the mouse+GUI, the internet, the GPU, GPS, etc. (though perhaps not quite at that tier). The philosophical implications of it are interesting, but very much in the same way that the philosophical implications of the mouse + GUI, or long-distance communication, are interesting. What is exciting about it is that we now get to collectively discover all the neat new things we can do with it. We're going to find how it is a tool. Maybe we also find how it reflects on our own minds, but as others have said, I think likening one to the other is a category error. To further what Ian said,

it doesn’t have an understanding of anything it’s writing.

I would say: Nobody should expect it to have anything that could be described as understanding. It doesn't need to have understanding to be interesting. /2¢

Max Woolfe has a good blog post about tempered expectations https://minimaxir.com/2020/07/gpt3-expectations/

From what I see as reactions, people seem to think it understand the code it’s writing for example or what the text is about. “Understanding” means the ability to recognize an object/pattern in multiple dimensions and thus have a subjective representation of the object/pattern ( which works into prediction skills that are needed to survive in the world ) and know it’s relations to it’s context. GPT-3 is 2.5D, we need nD. N “models” which constantly self-optimise,minimise and grow under control of 2 agents - internal & external. But I agree with @Scott Anderson, IMO the best way to use it for code generation “is a probalistic search engine”. If someone trained it to write a minor part of the code where “branches” connect, it could be super useful, but we’d still need a different access to writing code. For now, it could be a great autocomplete.

@Jimmy Miller In philosophy you have epistemology and metaphysics and argue about what understanding “means”, what meaning is, what truth is, etc. And there are lots of positions to take and that has all kinds of implications on your world view: realism, objectivism, internalism, experientialism, etc. What I consider somewhat separate from that (and put in the cognitive science bucket) is the part where we figure out how the activity of thinking manifests in the brain, things like embodiment, image schemas, metaphorical structuring and their neural modeling, etc. In a way you made a similar distinction yourself when you say “_Understanding how the brain works more does not necessarily get us to understand how understanding works_” — I agree and somehow think that is almost the point I was making, albeit you phrased it much better. And yet of course these are all interdependent, and what I’m trying to point out is that there are lots of technologists who do not look at philosophy or cognitive science (and therefore likely base their thinking probably on outdated and/or folk theories) and still have strong opinions on how close GPT-3 is to AGI or not. I don’t understand the connection to the rest of your comment and I'm confused about where your assumption that I might not be interested comes from.

Can somebody explain to me what “understanding” means? So machines are basically just trained on patterns… how exactly are humans different…?

This is the question I was originally trying to address. Are humans any different than machines? Can we just continue to make better models and achieve understanding or consciousness or intelligence? Studying the brain doesn't give us answers to those questions. I am glad to hear that you are interested in it. It is probably just a bias that I've picked up where many people are against philosophy.

there are lots of technologists who do not look at philosophy or cognitive science (and therefore likely base their thinking probably on outdated and/or folk theories) and still have strong opinions on how close GPT-3 is to AGI or not.

I guess what I was trying to get at is I don't think those views are grounded. The question of even determining if something we created has true understanding, is actually intelligent, or has consciousness is a hard question. In fact, Ned Block calls that the Harder Problem of consciousness.

Are humans any different than machines?

We are. Our “model” does wetware based optimisation, morphing and most importantly localisation, which is something machines cant do while in hardware mode. Yeah, you can rearrange it in memory, but memory is 2D and that means huge overhead we don’t have. You need a multidimensional graph and mocking that in 2D is annoying as hell. And we are geared towards survival and reproduction, so we have “emotions” that reinforce our models

Can we just continue to make better models and achieve understanding or consciousness or intelligence?

Not this way. OpenAI has other stuff that is more geared toward better AI and intelligence. If we just continue to make better models but keep them shallow like this, only thing we will understand that intelligence isn’t pattern recognition.

@Ian Rumac I'd love to read some material about what you describe as "wetware based optimization, morphing and localization" — do you have any pointers to sources that go into depth on these?

Let me look it up, I remember a paper last year that was about how brain uses spatial localisation to optimise (kinda annoying now Neural networks are muddying up my search results 😂) By “optimization” I mean about constant retraining and optimisation our brains go through. Our “internal optimization agent” sees we have a graph G(n) of n nodes and a graph G(m) of m nodes. They get combined into a graph of k=(<m*n) nodes, while maybe not retaining the absolute combined precision, but probably having a modulator which is extracted as a separate graph. For example, Gn is trained for a fluid (oil) and Gm is trained for oil. Then, they are merged/splitted into a model that recognizes “fluid” pattern and a “viscosity” pattern (our brains are constantly looking for patterns. constantly. they are our primary source of optimization, i believe that this skill is critical to our brains evolution and that that is why geometry is important in our brains arrangements). Let’s say you input lava, it would be recognized by both models, one recognizing it’s a fluid and the behaviors of fluid and one that can tell you the viscosity of it. Hope I explained it OK, will try to find papers, they do it better probs (i mean, this is my conclusion from reading multiple of papers, so its dot connecting conclusion from multiple sources).

@Ian Rumac Thanks for digging this up; looks promising. I have the same issue, often running into neural network stuff when searching for brain research. I’m specifically interested in categorization and how semantics manifest in our brains, schemas and frames, essentially the parts between the biochemistry of the brain and the philosophy of thought and reasoning — here’s a good overview from Lakoff: If you or anyone else comes across research in that area, please send it my way.

Reading the conversation in this thread kinda makes me wish this Slack also had a philosophy channel 🙂

@Stefan oooh interesting video, will check it out later when I have time! tldr as far as I see its about research that tries to find primitive model our brain recognizes and which are used as building blocks to build more advanced models by having a few types of interactions between them?

@Ian Rumac Yes, that describes the gist of it. It’s focused on language, which is only one of many parts of thinking — arguably quite an important one. Helps to be familiar with the work of George Lakoff and Mark Johnson on metaphorical structuring. There’ll be plenty of pointers in the lecture.

No clue who any of those people are but lecture sounds awesome, tnx for the link 👍