All right. Great to be here again in a conversation among friends. So it is great to be here and to be in conversation, and I'm looking forward to talking with you, Peter. I have my notes in case I get stage fright.
Yeah, I also have notes, but I'm not sure whether we need them, actually. But yeah, so we have four different topics that we wanted to touch upon. I think the sort of high-level topic is the miseducation of AI.
Yeah. And obviously the first topic, just to give a high-level overview: AI paradigms, where are we today, AGI, superintelligence, where do we go from there, what are the risks, agentic AI, other risks, and then eventually looking at US and Europe, or US versus Europe, and then a bonus topic at the very end. So why don't we kick off with AI paradigms: where are we, and when will we see AGI and superintelligence?
Oh, I love this question because it requires me to do like a one-hour history lesson if you want to actually dig in. Yeah. No. I think the struggle with this type of question is that almost all of the terms we invoked are not technical terms of art. They are, you know, marketing terms or aspirational terms, but we're not talking about something specific. Even when we talk about AI, let alone AGI. Superintelligence is a sort of sci-fi concept, not augured in a technical reality. And so we can think of these almost as unmoored or floating signifiers that get, you know, over the 80-year history of the term AI, have been filled with very different meanings, have referred to very different technical approaches, all of which were aspiring to create something that kind of did things that humans do, at least according to various assessments. And so right now we are in a very particular moment where the term AI, which was coined to mean something very different, refers almost completely in the general technical and public discourse to large-scale neural networks, to the deep learning approach, which became very, very popular in the early 2010s, largely because the platform business model had the economies of scale and the data via its platforms and network effects to reanimate deep learning as a going concern, to reanimate and kind of bring back from the archives approaches that were created in the late 1980s, early 1990s, then dropped because they didn't deliver on their promises. That's what's called an AI winter, and then came back into vogue because, oh wow, we can use the parallelization capabilities of GPUs. We can use all of the surveillance data collected by the platform companies, which were the first to recognize how useful these could be, and we can make these do things. We can make them dance in ways that they weren't able to dance before. So we're now living in that age. You know, every step in AI, all the progress we talk about is still in the deep learning paradigm. It's a, you know, scaling laws are sort of one of the big buzzwords around this. But that's one paradigm of AI. That is one approach, and my critique, you know, one of my primary critiques of this approach is it ultimately relies on the type of concentrated power that is in the hands of tech monopolies that are largely based in the US and to some extent in China due to historical factors that sort of, you know, allowed them to grow their reach, allowed them to grow their networks and economies of scale, which is a business model that is very hard to interrupt once it's established. So that's, you know, there are many other approaches is the good news here. There are many other approaches to AI to fill that signifier that don't require that scale and that have been largely ignored because that has sort of taken up all of the room under the AI umbrella.
So the boring part here is that they put on stage two people that actually agree. Yeah. So well, you guys can take a nap if you want. So no, but so I've said that, you know, these are just A-to-B mappings, right? And, you know, obviously neural networks are universal function approximators, they will approximate any function out there, that's been proven. But we are teaching the machines to approximate very specific functions, and so that's, you end up with that conclusion. Two things: you did not answer the question, when AGI or when superintelligence?
I'm so good at media training I can just avoid it.
But if we look at different modalities, I mean, you know, if you have A-to-B mappings, tokens, pixels, whatever, what we're seeing now is these chatbots and assistants creating value in the world. But then we have different modalities like life sciences, say automotive, material sciences, gaming, visual media, entertainment, and so forth. How do you look at sort of what's coming? What's the next paradigm? How does this paradigm evolve, and then when do we get to AGI and superintelligence?
Yeah. Yeah. I want your answer to that as well. But I think, you know, I reject inevitability as just a political stance. And so I think what's coming is what we create in a large sense, and kind of allowing the narratives of the large companies to sort of, you know, drive us toward creations that are convenient for them and their business model, and assuming that's a synonym for scientific progress doesn't, you know, doesn't serve us. So I think part of what I hope slash want to help create as a future is a much more heterogeneous field of AI development. So when we're looking at, you know, automotive, look at the problems that need to be solved for that industry. Look at issues of data integrity. Look at issues of, you know, local context. I've heard a number of my friends just anecdotally who are in the UK, where, you know, they drive on the wrong side of the road, where driving norms are a little bit different, talking about their Teslas driving like an American, right? That's a very shallow example of, you know, a much deeper problem. Like if your data isn't rigorous, if your distribution isn't robust as a proxy for the reality in which you want to deploy, your system will be janky, to use a technical term. And so that's, you know, what I'm describing there is actually a market that could be opened up if we take those types of problems seriously and if we're willing to experiment with approaches to AI that aren't, again, just, you know, white-labeling a large-scale, you know, a foundation model API and applying that in one or another context.
Did that, and when AGI back, right? You know, yeah. And I mean, what is your definition of AGI, or how do you view that kind of aspirational north star slash sci-fi concept?
So you're doing exactly the same that I'm doing. I usually give it back and then I ask for the definition. So I don't think... well, I think we both spent time in academia, so that's...
But okay, let's move on. I know you've been talking about risks with AI, very specific, but let's broaden up first. And, you know, what I've oftentimes said is that, you know, while, you know, I don't know how this technology would get us to AGI and superintelligence and risks related to that, I think at the same time we have for the first time in human history in mass distribution technology that fulfills the Turing test. So we as humans cannot distinguish between machines and humans, and, you know, you can do bad things with such technology. There are certain risks with it, and of course, you know, risk as such is a broader question. So why don't you sort of open up, how do you look at risks with AI today?
