As AI chatbots learn from online data and interactions, artificial intelligence companies are working to make their content safer for users. Duke University professor and former FDIC Chief Innovation Officer Sultan Megzi describes a flaw in AI that engages more users, saying that large-scale language models may actually learn from “the wrong side of reinforcement learning.”
video transcript
– AI, of course, is the buzzword of 2023, and it made headlines again this week. Sergey Brin is back at his Google, just to name a few. The White House said OpenAI and Google have pledged to watermark AI content to ensure its safety. And Apple is developing its own artificial intelligence system. And now, according to the White House, some of the biggest companies, including Amazon, Google and Meta, have voluntarily signed new contracts to manage risks related to AI. So there’s a lot going on.
But is the hype and positive chatter dying down? I think the next guest might be. Let’s invite Duke University professor and former FDIC Chief Innovation Officer Sultan Megzi. Sultan, nice to see you again. As you know, this year has been quite a big thing with his AI. And the market certainly sees a lot of enthusiasm in this area. Where do you think we are in the AI hype cycle?
Sultan Moezari Meghzi: Well, I’m happy to be back. This is a great question. Because much of today’s AI is only driven by marketing teams, not really by technology. So I think it’s not at all surprising that we’re talking about hype cycles. A lot has been announced. As you know, social media and elsewhere are full of glittery logos and great press releases and discussions.
But we haven’t really seen much work done. And the current generation of AI will start to make a real impact in the coming years, streamlining back-office processes and more. But that’s an entirely separate set of activities from those featured in the news.
– Sultan, that’s fine. Yes, AI seems to be taking on more roles within more companies, such as efficiency and news assistants. whatever that means. However, there are some reports that the AI itself is not as smart as it used to be by layering things like machine learning. And this kind of feeding frenzy itself makes it even more silly. Have you seen such things in your research?
Sultan Moezari Meghzi: absolutely. As you know, almost every large language model out there exhibits the following symptoms: [INAUDIBLE] Wrong side of reinforcement learning. AI therefore needs to be very careful about the data it uses to train. And at some point, the positive returns stop. You stop getting smarter. And at that point you need to stop and take a step back and focus more on the type of data you’re analyzing. For now this is the best we can do. And you know, we’re starting to see that.
I think almost all of the big tech companies are hitting a plateau in terms of their published LLM capabilities.
– Part of it – the Sultan, at least according to some reports I’ve seen, is not only stagnating, it’s getting stupider, I can’t find a better word for it.
Sultan Moezari Meghzi: yes.
– I mean, I feel bad for calling something stupid, but I think it doesn’t live up to it. that’s ok. But I do understand that there will be stagnation. But why retreat?
Sultan Moezari Meghzi: As you know, many of these systems are basically based on what they have learned from what is published on the Internet. Let’s be honest, a lot of what’s on the internet isn’t that great. And when you run out of data to train on, you’re essentially drinking water from the same well that you’re supplying because all you can do is look across the internet and see the data your system generates.
So, you know, it’s not at all surprising to me that one of the LLMs, in one report, was doing the math very well until a few months ago, and now suddenly can’t even do 2 plus 2 reliably. This speaks to the fact that the data used for training has not grown as much as the system itself has grown.
– Given the context of Sultan, especially ChatGPT, there has been some criticism of where Apple stands on AI, but Microsoft is now the frontrunner. And now Apple is apparently creating its own – according to some reports, Apple GPT or Ajax. I think that’s the name they call it. What are they working on? Who do you think will be the winner in this area?
Sultan Moezari Meghzi: Well, please don’t tell me to bet on Apple, but that sounds dangerous. But Apple has a long history. It’s part of the organizational culture. They weren’t the first personal computers. IBM did. They weren’t the first smartphones. Before that, everyone had a BlackBerry.
I wouldn’t be at all surprised, perhaps, that they’re taking a slightly slower, more engineering-focused approach than many other organizations. But they also have a huge consumer base as one of the multi-trillion dollar tech companies. And the moment you decide to release something like that, within 24 hours millions or even hundreds of millions of people will be using it.
And, as you know, with this scale, we have to be very careful about introducing new backend supported features. One thing most people may not realize is that the growth of artificial intelligence is limited by the number of CPUs and GPUs that can actually perform the computations that make it work.
And if you suddenly had to add 100 million users to your AI system, you would need a significant amount of infrastructure behind it. And it’s also possible that they simply don’t have it.
– yes. For example, this explains the enthusiasm and demand for NVIDIA chips and stocks. But again, maybe we should take lessons from Apple when it comes to talking about where we are in the cycle of being a little more cautious. So I’m thinking about all the hype cycles we’ve been through in the last decade, from cryptocurrencies to the metaverse to even fully autonomous driving. The talk was big there, and the revenue either never materialized, or it came much later than supporters expected.
Earlier this year, I kept hearing that “AI is different.” It’s a big deal and can have many implications. Was it just fundamentally wrong, or was it just another extension of the schedule?
Sultan Moezari Meghzi: Well, it’s a really nice framework. I think it’s the timeline. You know, the metaphor I use is the cloud. He was in 2007 or 2008 when the word cloud was actually coined. And it took him until 2023 for most organizations to develop a cloud strategy.
So, you know, the US government, where I work, couldn’t move completely to the cloud. Not yet. And when talking about the impact of large-scale technology, there are time considerations that most people are unaware of. So, going back from 2008 to 2023, we haven’t fully moved to the cloud yet, and AI is an even more influential technology in my opinion, so I wouldn’t be at all surprised if we’re talking about a journey that takes decades to really realize the benefits of artificial intelligence.
– And Sultan, you just talked about the data to train on, you know, stupid, stupid, but where is the responsibility when it comes to AI training, especially when you think about the possibilities and use cases in law enforcement and the legal field in general to prevent discriminatory prejudice?
Sultan Moezari Meghzi: That is, discrimination occurs not only in law enforcement, but also in basic matters such as credit and loan decisions. It carries a lot of risk. So, you know, the Biden administration is approaching artificial intelligence as risky, and that’s what they’re doing, and they show that they’re front and center about this. This task, like the Internet itself, cannot be met by governments alone. Private companies cannot do it alone. Academia alone cannot do it.
All these organizations, including the average American, the average human, need to be on board with this discussion. Because current laws, current regulations do not consider this kind of technology. In my opinion, we need new systems, new regulations, new ways of thinking about this.
and it should be built. And it most likely doesn’t exist yet. So while it’s a nice step to hear that some companies will watermark there or choose to work with the U.S. government in an informal capacity, it doesn’t really solve the broader problem.
