Latest / Technology Now / Chatbots, agents & LLMs: the future of AI where bigger isn’t always better | Kirk Bresniker
Transcript
- MICHAEL BIRD Sam, we talk about AI quite a lot on this show. but I wanna sort of ask you a question I don't think I've asked you before, which is, when do you first remember hearing about AI in your everyday life? SAM JARRELL Oh gosh, probably, within the last five to six years obviously, but, I think probably if you remember when there were lots of those fun, poorly done at the time edits of famous actors doing things like being on the runway or eating spaghetti. MICHAEL BIRD The eating spaghetti one haunts my nightmares. what do they say? That was a horrible day to have eyes when I first saw that. mine is slightly more nerdy. I remember the first time I could access one of those image generation tools, and I can't remember what I asked it to generate, but it looked nothing like it. And then, fast-forward like five years on and last year we did an interview with somebody, and we said "How can you tell a deep fake?" And you said to get people to show your hands on camera. I don't think that's the case anymore.Now, the difference in AI back then to where it is today, is frankly astronomical. but I think what we wanna know is what is AI going to look like in the future? I'm Michael Bird SAM JARRELL I'm Sam Jarrell And welcome to Technology Now from HPE KIRK BRESNIKER We can go back decades. You can go back to when people were tuning a neural network by having a stepper motor turn the knob on a potentiometer. Clink, clink, clink, clink, clink. we do all of that now digitally and at incredible volume thanks to semiconductor processing, thanks to GPUs, thanks to supercomputing and the networking of storage that enables us to process literally every byte we've ever recorded and encode it into a model. But even before that, we had breakthroughs in AI. I remember the first time my first printed circuit board was being assembled in the factory that was only steps away from my engineering desk, and even then it was computer vision, right? Those parts were all image, adjusted, and put down with, millimeter precision because of computer vision, a branch of artificial intelligence and generally, artificial intelligence is how do we get a computer to do some capability that we recognize in our own intelligence? So we built up models. We built up that ability to make predictions. Now, when we sort of got to the watershed moment, uh, was with the large language models because language is at the heart of how we think, how we feel. And so now to have systems that can use language, can begin to repeat language, to create language, to understand , and do knowledge work based on language, that's sort of that watershed event, and that's only, November of '22. and, suddenly all the things that we do are becoming that much more accessible for acceleration with artificial intelligence. MICHAEL BIRD think most people when they think of AI are, I think they would say 2022. But, uh, to your point, it's been around for decades KIRK BRESNIKER Decades. This goes all the way back to Alan Turing and the beginning of computing. It's also when he began to think about thinking machines. So yes, the history of AI is almost as long as the history of electronic, advanced computing. MICHAEL BIRD 'Cause there's the Turing test, isn't there? KIRK BRESNIKER the Turing test, has been solved, right? So you can talk to a chatbot now, and it's impossible to discern whether or not that is a human. So we actually need to go beyond that because it seemed like it was so insurmountable for so long, and then it was done. the power of these tools to mimic language and behavior. can be a hazard for us because it can be an echo chamber. It can be reinforcing, something that may not be true. It can be, we say, hallucinating. all those are processes of language that we employ. We do that creative spark, but then we know how to reason over it, and that can be part of the challenge. These language models are trained on every element of eloquent sophistry that's ever been recorded. It's all in there, and it's all coming back to us, and so maybe it's going to produce an invalid result that is so well articulated, we're going to believe It. MICHAEL BIRD when, again, most people think of AI, I guess what they're probably thinking of is, is an LLM. KIRK BRESNIKER Now they are. I mean, now that it's, it has gained the faculty of language, that's what everyone is thinking about. and even those things that we used to do in another way, like going through a search, engine and looking at the search results, now we want to have a conversation because We're drawn in as humans. We're drawn into conversation just like the one we're having right now and that, can be the challenge, MICHAEL BIRD Yeah, 'cause it, feels like we briefly went into an environment where most organizations were trying to solve every problem with LLMs when they weren't necessarily the right solution to the problem. But it was 'cause everyone else was doing it, so they wanted to catch up. KIRK BRESNIKER it is always the question of the right tool for the right job. but that's part of that enchanting behavior. It is so well conversant and so fluent that we presuppose other depths of capability that aren't really in the purpose or, supported by the way that the systems have been designed. And so that's part of our own education, is knowing when to use the appropriate tool for the job. And certainly also within industry, there's certain elements. We think about programming, we think about, legal documentation. All those are language, but very specific language, and they're repeated motifs and, that's why we call things like boilerplate, whether it's legal or programming, "Oh, we got the boilerplate code in." Fantastic. That is a great opportunity for us to automate some of those laborious, repetitive tasks, to have them always be compliant. It's always, being coded against the style guide because now the style guide has been incorporated into the large language model. Now, it's not necessarily though where I want to do more complex reasoning and complex opinions out of those systems, although we're doing some of that as well. and that's where we still have to ask ourselves: Is this the right tool for the job? MICHAEL BIRD Yeah. So, what's next after LLMs? Like is there the next big thing that's gonna land and is gonna, have as big a cultural impact as LLMs had back in '22? KIRK BRESNIKER certainly at Labs we believe that there is more that's beyond the LLM. First, there's... The LLMs aren't going away. we will continue to see them increase in sophistication, but, we also have to caution ourselves because one of the things about a large language models, it has something, we've talked about this before, quadratic scaling. You want the model to be 10 times bigger? Well, expect to invest 100 times the resources into creating and utilizing that model. The converse is also true. If I make a model 10 times smaller, I'm gonna use 1/100 of the resources. So understanding how we collect groups of models together and specialize the model, so not just one general purpose model that can do everything. How about I have expert models? In the same way I hire expertise in a team and actually what we're doing research at Labs is, how do you assemble these model experts, tune each one according to their individual capabilities, and then set them about working in concert, with an orchestration agent? So one agent is really good at keeping all the other agents in line and so understanding how we assemble those models and then, begin to reason with those models. We can take all this data, and create a inductive reasoning model. Huge amount of data, and it's been distilled into the correlations. And so now I can take new data in and boop, really fast, very quick decision-making. Now, the challenge is, that kind of model isn't very good at extrapolating to go beyond, or when two answers might be equivalent, "I don't know which one is the right one," the model's not gonna help us. And so new styles of models are going to emerge, new capabilities to add more of that deductive reasoning, If the premise is true, uh, the conclusion is necessarily true. So adding that True first order logical reasoning capability with new styles of models is what we'll see complement the LLM. So today we reason by repetitive LLM usage. Tomorrow we'll complement the LLM with new styles of models that will have that deductive capability. MICHAEL BIRD so it sort of sounds like going from, these massive large language models that uses quite a lot of resources to maybe answer a simple question because you're asking it to, look at all the world's knowledge to understand, what time does the local shop down the road close, or where should I go on holiday? whereas actually, maybe what we're going towards is lots of different components. a model,that's specialized in booking flights and a model that's good at understanding particle physicsand then having something that can orchestrate them and say, "Right, I need you to answer that question. I need you to answer that question. I need you to answer that question," and then we're gonna combine it all together and do something with it. KIRK BRESNIKER and especially when these models are now becoming tool users. I wanna control this complex cyber physical system. Maybe you'll have a model that knows how to run the simulator. Maybe you'll have a model that understands the digital twin and say, "Okay, well, uh, I think first you should do an analysis of this idea. Okay, here's the simulation results." I can have an agent that reads the simulation results and then interprets that and then turns it into the next plan of action. So augmenting LLMs with additional tools, with these agents that can perform additional actions in the same way that we would populate a team. We would construct an organization that is capable of, of taking on a complex task by breaking it down into smaller pieces, so understanding how we will bring in new tools, new capabilities, new types of models to complement the LLM is where we see us, you know, really looking beyond. MICHAEL BIRD and the presumption is that that means that we can be a bit more energy efficient using less energy, using less resources to be able to achieve maybe a similar or the same answer. KIRK BRESNIKER Agreed. And as for that energy efficiency, part of this is also including more of our knowledge as, resources for the model. So rather than just work on incorporating the knowledge that we have into the weights of the individual model, let's also have, comprehensive, understanding and representation of those models, knowledge graphs that are available for the tool to examine, to guide the tool. So simulation and modeling plus AI and LLMs plus knowledge graphs is the way that all these things might come together. Some agents are simulating, some agents are running the robotic laboratory, some agents are then collecting up and correlating that data and helping the scientist understand and interpret the results, and then starts this cycle of hypothesis, experimentation, and conclusion over and over again. MICHAEL BIRD I guess there probably are, small or medium language models are they generally just more efficient? KIRK BRESNIKER Well, certainly they have the advantage of the quadratic scaling. And so then the question is, how can we efficiently utilize that mixture of experts? That's the other term you'll hear about people who have already using this for very comprehensive models, not because they have a single model that is comprehensive, but they have expertise. Again, all of this is seeming like an echo of how we organize ourselves for success as well. So we're seeing that repeated motif of drawing upon individual strengths across a range of diverse, skills in order to efficiently arrive at a solution. that process of continuous refinement of how we partition the knowledge into the models,I don't wanna hope that I've trained my model in understanding that V equals IR, Ohm's law. I want it to understand and utilize those basic formulas of engineering. That's a world model. That's something we know about the world. and certainly while it is reflected, hopefully, in all the data I've used to train, why do I try and achieve it efficiently that way when it's much more efficient to have the simulation capabilities at the beck and call of the model so that it can run the simulation, it can look up V equals IR, it can do the multiplication in a supercomputer rather than, trying to repetitively, talk itself into a solution, with language. MICHAEL BIRD So Kirk, cast your mind forward to maybe twenty thirty, twenty thirty-five, twenty forty even. Mm-hmm. What will AI look like in a decade, 15 years, two decades? KIRK BRESNIKER Yeah, I think, AI will look like an expanded set of faculties. We think of our own, philosophical examination of our own human intelligence, and it's multipartite. There are skills and behaviors and, our ability to react quickly and fight or flight. Well, that's something you have to do fast. You do it on scant information. You've done it, uh, because you have those dedicated capabilities that have evolved in your brain. But then there's the deep thought, right? So what is the model equivalent of that deep thinking, sometimes called system two thinking, and part of the research that's very promising in this is something that's called an energy-based model. So rather than train all the data and create all the correlations, let's make some assumptions. Let's think about how we decide how the world works. We associate, how things fit into that world model with energy, and we can evaluate efficient circuits from our research and accelerators at Labs is how to efficiently optimize, energy profiles and to get the best way to run a distribution network, or to satisfy a bunch of constraints. That's also how you do logical reasoning, is constraint solutions. And so seeing some additional capabilities, additional faculties of reason brought into and complementing the incredible generative language capabilities of the LLM. So we have the LLM, we have the EBM, we have the world model. It's an increasing set of faculties in the artificial intelligence portfolio, and then that's really what's gonna build up that next set of rich experiences, whether talking about science, engineering, or public or private leadership, having now additional capabilities that we know we have as well, that ability to reason, that ability to logically think out, to plan, to hypothesize, and then prove. I think we'll see all those come out, and I don't know that we'll have to wait till 2040. I think maybe in 2030, we'll be all talking about how we've added to the capacity of the enterprise to generate AI outcomes with the LLM extended, and complemented with novel technologies. MICHAEL BIRD Kirk, as always, absolute pleasure chatting to you. thank you so much for your time. Real pleasure having you on the show. KIRK BRESNIKER Absolutely. Absolutely Happy to do it. MICHAEL BIRD I always love having Kirk on the show. I always feel like I learn something new every single time I speak to him. SAM JARRELL I agree. I think he also does a very good job of bringing things down to the appropriate level depending on who the audience is. I always feel a bit like concepts that go way over my head, I always wanna grab his ear and be like, "Hey, can you explain? Like, how does AI work? How does quantum cryptography work?" And he is the best person I've found to bring it to my level. MICHAEL BIRD I would agree. Now, I sort of took away two themes as to, the future of what the future of AI is going to look like. maybe you disagree or agree with me, but, basically more efficient models and, agentic orchestration. SAM JARRELL I think that that's largely correct. I felt like quite a bit too, he was emphasizing the idea that AI is going to extend human thinking and had a little bit of a, like a caution around like AGI. it goes back to that conversation we have around will AI replace humans? Or what does it look like in an, agentic world where it's us and the AI? it seems a lot like in order to do that extension, we'll have specialized agents that then operate sort of multiple LLMs and then can also themselves consult with like whatever compute power they need. MICHAEL BIRD Yeah, it feels like the, big large language models that we have at the moment are sort of sledgehammers. And actually, if you're trying to hammer in a nail, you probably can do it with a sledgehammer, but it's massively overkill, and maybe a better way of doing it is using multiple specialized tools to achieve the same thing. SAM JARRELL in some ways we kind of do this on like a small scale currently, right? Where in different AI tools, that you may leverage, specifically in like LLMs, some tools do offer the ability for you to make your own LLM-based agents, right? And you can sort of define the characteristics of them, what you want them to know, what documentation you want them to reference, and how you want them to go about reasoning through different things. So I think we do a little bit of this now, so it didn't entirely surprise me when Kirk said for some of these things, he doesn't think that we're gonna be waiting until 2040. He thinks it'll be as soon as 2030, which is very, very soon. MICHAEL BIRD Yeah. I found that quite fascinating too. I think you've mentioned on the show before, but the quadratic scaling, I, I actually sort of took time to, to write it down and sort of look at it on paper. So if you want your model to be 10 times bigger, you need 100 times the amount of resources. If it's 10 times smaller, you need one hundredth of the resources. So again, those smaller, more efficient, more knowledgeable, more specialized agents with an orchestration agent feels like that's potentially a way more efficient way of doing things. And again, I think we talked about this on the show before, but like modern computers with their modern processors, they will often have chips that do particularly- particular tasks really well, like a chip for video encoding or a chip for security, rather than having one big beefy general purpose processor. So it sort of echoes of those lessons that we've learned a- as to how we design computers is maybe how we're designing, an AI or an AI agent or set of agents to solve a particular problem. SAM JARRELL It goes back to the conversation around the right tool for the right job. Same thing with like if you want your AI to be specialized in the right way, you want the right setup for it. you guys had the conversation around for a while it just kind of seemed like, "Let's put AI in everything. Let's put it in the fridge. let's try to have it solve all of these crazy problems and have it do things that it's maybe not even intended to do yet." I think that similarly with whether or not an agent is going to be successful is going to be whether or not it is set up correctly. And there's a lot of work that still has to be done around making the right setup to do the right job with the right tools. MICHAEL BIRD Now, Sam, you mentioned this at the top of this discussion, artificial general intelligence SAM JARRELL I did. mainly I was kind of getting a drift from Kirk around like this is an extension of our capabilities and I think that, I get the feeling he's thinking that we're not there yet. MICHAEL BIRD Yeah. Well, I'm so glad you asked the question 'cause it's gonna lead us very nicely onto our final thought because the final question I asked Kirk was, if he thought we would ever get to a true AGI KIRK BRESNIKER we have one example of fully conscious, fully self-aware, intelligence in the universe. and unfortunately we have unreliable narration on how it came about. so we'll have to see. But, when we talk about, every aspect of human capabilities done better or faster by an artificial general intelligence, I don't know. I'm hesitant. and I actually wonder sometimes if it is, not the way we should be thinking about the capabilities. not that we want to have something that does exactly what we do, but just better and faster, but to, again, complement and extend our thinking. right now we know of no natural intelligence that didn't develop in a gravity well. so everything that we think of, the entire spectrum of intelligence on, on this planet still knows about up and down. what would an intelligence be like if it evolved in the physics of free space? If there was no up and down, how would their conceptions, how would their thought process, how would, their intelligence work differ from our own? what would evolve? And so I think when I think about these intelligent systems that aren't dependent upon how we think and where we evolve, I want to know what that intelligence is like as a way to complement, rather than, compete with our own human intelligence. SAM JARRELL Okay that brings us to the end of Technology Now for this week. Thank you to our guest, Kirk Bresniker And of course, to our listeners. Thank you so much for joining us. MICHAEL BIRD If you’ve enjoyed this episode, please do let us know – rate and review us wherever you listen to episodes and if you want to get in contact with us, send us an email to technology now AT hpe.com, subject line: Gravity Well, and don’t forget to subscribe so you can listen first every week. Technology Now is hosted by Sam Jarrell and myself, Michael Bird This episode was produced by Harry Lampert and Eva Higginbotham with production support from Alysha Kempson-Taylor, Nik Damarell Beckie Bird, Alissa Mitry, and Jenessa Ayache. Our theme music was composed by Greg Hooper. SAM JARRELL Our social editorial team is Rebecca Wissinger, Judy-Anne Goldman and Jacqueline Green and our social media designers are Alejandra Garcia, and Ambar Maldonado. MICHAEL BIRD Technology Now is a Fresh Air Production for Hewlett Packard Enterprise. (and) we’ll see you next week. Cheers! SAM JARRELL Bye y’all