00:00So let's talk about these LLMs, these models, and how we can interact with them
00:04to actually benefit our health, because so much of what we've heard about has been
00:08the detrimental effects on our health and well-being.
00:13Yeah, pleasure to be here. So basically, LLMs today, of course, are everywhere. They're being
00:20pushed to us in our text editors, Google search, all of that. But we still don't know what the
00:28impact is of using these chatbots for assistance in our daily lives. And one of the things that
00:36we are looking at here at the MIT Media Laboratory is exactly the impact of using these technologies
00:45day in, day out on our cognitive functioning, on our ability to think critically, our ability to
00:52learn, our ability to solve problems, our socializing, our loneliness, and so on.
01:03Yeah, you know, that's the thing. One of the things I think about, Dr. Moss, is we did so well
01:10in keeping
01:10on top of social media, so it did no harm.
01:13She's being sarcastic.
01:14Thank you so much. He's right, though. So my fear is, how do we keep on top of this? So
01:19tell us about
01:19the work you are doing to try and get a handle on how it might impact our brain, our psychological
01:26health, and how that might inform the companies that are in the thick of it.
01:32Mm-hmm. Yeah, so we've been doing studies here at the MIT Media Lab showing that when you use these
01:39technologies to help you with a particular work task or with writing an essay, et cetera, it may
01:48actually improve your performance in the moment when you have access to these tools, but it can
01:54actually erode some of those same skills in the long term if you rely on these technologies for too much
02:03of your work, basically. So it's really important that we think about how these technologies should
02:11be integrated in our daily life and in our work life in such a way that they don't erode skills
02:18that
02:18we don't want to lose. And that's exactly what we try to focus on with our work.
02:25So how do we do that when we are encouraged to use these? I mean, you have people come on
02:32our
02:32program and say, if you're not using these tools, then you are essentially erasing yourself from the
02:37workforce in the future. But there's that balance that we want to make sure we continue to be able
02:41to think. Yeah, well, the emphasis of LLM development has been especially on increasing their capacity,
02:51their efficiency, things like that. But it hasn't exactly been on the design of these systems and
02:58designing them in a way so that ultimately the benefit for people is a net positive, especially
03:06in the long term. So we believe that we actually have to increase the friction in these systems and
03:14that they shouldn't too readily just do all the work for us, but they should play a more augmentative
03:22role, basically, and support us in also developing our skills and keeping our skills sort of up to
03:31a certain level. So what do we do? Is it create guardrails? Like, or when we ask a question, like,
03:38that's pretty easy. Don't you want to do that on your own? And, you know, then there are those of
03:42us
03:42who are going to be like, yeah, it might be easy, but I don't have the time. Like, I just
03:45wonder,
03:45how do we actually manage this? Yeah. And, you know, I was going to say, and we're in a world
03:54where the government and you have, you know, is it open AI, right? That has been pushing back about
04:01concerns about weapons that can ultimately, you know, or systems that can think on their own.
04:06So, you know, anthropic, forgive me. Anthropic is who we're thinking about, but like this pushback
04:11and the concern about what this means for weapons. But I do think about these systems that
04:16can maybe almost have a life personality of their own and what damage they can create,
04:20especially for a younger population, but not just younger. Yeah. Yeah. Well, these systems are not
04:26necessarily a hundred percent correct yet. All the time, they still hallucinate. They don't really
04:32deeply understand our world. And so it's important that we still are able to understand things,
04:40solve problems, and that we can supervise these systems so we know when they may be incorrect
04:46or when we maybe shouldn't be trusting them. So one of the things we do here at MIT is we've
04:53been
04:54developing a set of benchmarks that actually evaluate the impact of LLM use on people, on their ability
05:04to think critically, for example, and to think for themselves. So it's very important that LLMs don't
05:12just always give you the answer depending on the context and who's the user and what they're using
05:20the LLM for. It may be better for that chatbot to just ask a question first of the user to
05:27engage them
05:28in thinking because our tests show that when an LLM just gives you the answer readily, people stop
05:36thinking for themselves. They're naturally a little bit lazy and they stop.
05:42Dr. Moss, does that align with the incentives of these LLMs?
05:47No, it doesn't exactly align with the incentives, which is another reason why we are developing these
05:53benchmarks, the benchmark for the human impact of AI. Because if we have a benchmark like that,
06:00we can put pressure on these companies to actually do well on that particular human impact benchmark.
Comments