I want to pass on to MindBlog readers some clips I have made for my own use from the transcript of a podcast interview
of Gary Marcus by Ezra Klein. These abstractings help me absorb the material better, and make it easier for me to revisit and recall the arguments at a later date. Marcus is an emeritus professor of psychology and neural science at N.Y.U. who has become a leading voice of not quite A.I. skepticism, but skepticism about the A.I. path we’re on. He has founded multiple A.I. companies himself. He thinks artificial intelligence is possible. He thinks it is desirable. But he doesn’t think that what we are doing now — making these systems that do not understand what they are telling us — is going to work out the way we are hoping it will. Here are the clips:
Marcus: the system underneath ChatGPT is the king of pastiche…to a first approximation, it is cutting and pasting things…There’s also a kind of template aspect to it. So it cuts and pastes things, but it can do substitutions, things that paraphrase. So you have A and B in a sequence, it finds something else that looks like A, something else that looks like B, and it puts them together. And its brilliance comes from that when it writes a cool poem. And also its errors come from that because it doesn’t really fully understand what connects A and B.
Klein: But … aren’t human beings also kings of pastiche? On some level I know very, very little about the world directly. If you ask me about, say, the Buddhist concept of emptiness, which I don’t really understand, isn’t my answer also mostly an averaging out of things that I’ve read and heard on the topic, just recast into my own language?
Marcus: Averaging is not actually the same as pastiche. And the real difference is for many of the things you talk about, not all of them, you’re not just mimicking. You have some internal model in your brain of something out there in the world…I have a model of you. I’m talking to you right now, getting to know you, know a little bit about your interests — don’t know everything, but I’m trying to constantly update that internal model. What the pastiche machine is doing is it’s just putting together pieces of text. It doesn’t know what those texts mean.
Klein: Sam Altman, C.E.O. of OpenAI, said “my belief is that you are energy flowing through a neural network.” That’s it. And he means by that a certain kind of learning system.
Marcus: …there’s both mysticism and confusion in what Sam is saying..it’s true that you are, in some sense, just this flow through a neural network. But that doesn’t mean that the neural network in you works anything like the neural networks that OpenAI has built..neural networks that OpenAI has built, first of all, are relatively unstructured. You have, like, 150 different brain areas that, in light of evolution and your genome, are very carefully structured together. It’s a much more sophisticated system than they’re using…
I think it’s mysticism to think that if we just make the systems that we have now bigger with more data, that we’re actually going to get to general intelligence. There’s an idea called, “scale is all you need.”..There’s no law of the universe that says as you make a neural network larger, that you’re inherently going to make it more and more humanlike. There’s some things that you get, so you get better and better approximations to the sound of language, to the sequence of words. But we’re not actually making that much progress on truth…these neural network models that we have right now are not reliable and they’re not truthful…just because you make them bigger doesn’t mean you solve that problem.
Some things get better as we make these neural network models, and some don’t. The reason that some don’t, in particular reliability and truthfulness, is because these systems don’t have those models of the world. They’re just looking, basically, at autocomplete. They’re just trying to autocomplete our sentences. And that’s not the depth that we need to actually get to what people call A.G.I., or artificial general intelligence.
Klein: from Harry Frankfurt paper called “On Bullshit”…“The essence of bullshit is not that it is false but that it is phony. In order to appreciate this distinction, one must recognize that a fake or a phony need not be in any respect, apart from authenticity itself, inferior to the real thing. What is not genuine may not also be defective in some other way. It may be, after all, an exact copy. What is wrong with a counterfeit is not what it is like, but how it was made.”…his point is that what’s different between bullshit and a lie is that a lie knows what the truth is and has had to move in the other direction. ..bullshit just has no relationship, really, to the truth…what unnerves me a bit about ChatGPT is the sense that we are going to drive the cost of bullshit to zero when we have not driven the cost of truthful or accurate or knowledge advancing information lower at all.
…systems like these pose a real and imminent threat to the fabric of society…You have a news story that looks like, for all intents and purposes, like it was written by a human being. It’ll have all the style and form and so forth, making up its sources and making up the data. And humans might catch one of these, but what if there are 10 of these or 100 of these or 1,000 or 10,000 of these? Then it becomes very difficult to monitor them.
We might be able to build new kinds of A.I., and I’m personally interested in doing that, to try to detect them. But we have no existing technology that really protects us from the onslaught, the incredible tidal wave of potential misinformation like this.
Russian trolls spent something like a million dollars a month during the 2016 election… they can now buy their own version of GPT-3 to do it all the time. They pay less than $500,000, and they can do it in limitless quantity instead of bound by the human hours.
…if everything comes back in the form of a paragraph that always looks essentially like a Wikipedia page and always feels authoritative, people aren’t going to even know how to judge it. And I think they’re going to judge it as all being true, default true, or kind of flip a switch and decide it’s all false and take none of it seriously, in which case that’s actually threatens the websites themselves, the search engines themselves.
The Klein/Marcu conversation then moves through several further areas. How large language models can be used to craft responses that nudge users towards clicking on advertising links, the declining returns of bigger models that are not helping in comprehending larger pieces of text, the use of ‘woke’ guardrails that yield pablum as answers to reasonable questions, lack of progress in determining trustworthiness of neural network responses, the eventual possible fusion of neural network, symbol processing and rule generating systems, the numerous hurdles to be overcome before an artificial general intelligence remotely equivalent to ours is constructed.
Post a Comment