Latest / CXOTalk / AI, Deep Learning, and the Future of Work | #860
Transcript
- 0:00Welcome to CXO Talk episode 860. I'm Michael Krigsman, and we are
- 0:07discussing AI and the future of work with Doctor Terry
- 0:11Signowski, a pioneer in computational neuroscience and
- 0:16deep learning. He's the Francis Crick Chair at
- 0:20the Salk Institute for Biological Studies and is also a
- 0:25distinguished professor in the Department of Neurobiology at UC
- 0:30San Diego. He is also Chairman of the Nur
- 0:35IPS Foundation, which runs the largest, most prestigious AI
- 0:41conference in the world. Terry has published over 500
- 0:46scientific papers and 12 books, including his most recent,
- 0:51ChatGPT and The Future of AI, and has won more awards than I
- 0:57can name. I have a background in physics,
- 0:59which is basically a license to work on anything, but my real
- 1:05focus is on the brain. I've won awards in neuroscience,
- 1:08the most recently the Brain prize, which is the highest
- 1:10award. And you know, I was a pioneer in
- 1:14the early days in the 80s developing learning algorithms
- 1:17for neural networks with Jeff Hinton and others.
- 1:20And so I have the, these two different backgrounds, which
- 1:23I've now have reached the point where this is really what's
- 1:27exciting is they're coming together.
- 1:29They're, they're actually synergistic.
- 1:32And So what we learned in the brain is going to help better AI
- 1:36in the future, and what we learned in AI is actually
- 1:40helping us in neuroscience. So this is really an exciting
- 1:42time to be doing research. How do you model AI on the
- 1:47brain? Maybe we should start there.
- 1:49Back in the 20th century, artificial intelligence was
- 1:53really focused on things that ran on puny digital computers,
- 1:58namely symbols, logic and rules. And there's only so far you can
- 2:02go with that, especially when you're dealing with a, a
- 2:04complex, very complex world with lots of uncertainties in it,
- 2:11right? And so you need, you need a
- 2:12much, much more computation and, and, and, and now we know, in
- 2:17fact, you know, that nature has solved all these problems,
- 2:22vision, you know, coordination, motor and so forth.
- 2:27And so we're learning now we popped the hood and now we're
- 2:29learning. What do you take from the brain?
- 2:31Massively parallel, highly interconnected parameters that
- 2:36called synapses in the brain and learning that you learn the
- 2:40strengths of those synapses with experience.
- 2:42And, and now it's big data and that has transformed all of AI
- 2:46And, but the point is though, we're dealing with the same
- 2:49mathematical structures, right? These neural networks are very
- 2:51simplified versions of brains, but they have the same
- 2:53mathematical way of of of analyzing and making progress.
- 2:58How does the brain organize information differently from
- 3:04machines and neural networks? And I just want to mention that
- 3:08the prominent data scientist Anthony Scrifignano has helped
- 3:12me prepare some of these questions.
- 3:14The general principles I've, I've given you take you so far,
- 3:17but actually nature has evolved of specialized circuits for
- 3:24specialized problems in, in different species in order for
- 3:28survival. And you know the, the different,
- 3:32there are many, many differences.
- 3:35There is the, the deep learning architecture that is used today
- 3:41was inspired by the cortex, the cerebral cortex, which is the
- 3:44highest level on top of the brain that humans use for, you
- 3:48know, the knowledge base. And there's a hundred other
- 3:52brain areas that are equally important for her being able to
- 3:56handle the complexities of the world and to survive.
- 3:58And, and a lot of those are ones that are right now are being
- 4:03integrated into the, the, the next generation of AI.
- 4:06And I'll just give you one example, reinforcement learning.
- 4:09So reinforcement learning is, is the part of the brain that is,
- 4:12is found in almost all species and has to do with being able
- 4:16to, first of all, from experience, predict whether a
- 4:22particular action is going to give you a reward or not.
- 4:25And you, you build that up over many, many trials.
- 4:28You know, if you're wrong, you twiddle awaits.
- 4:31And that's actually used for what's called procedural
- 4:34learning, learning how to play tennis, for example, or learning
- 4:37how to do us, you know, solve problems that are, you know, in
- 4:44law, in medicine, all of that is, is done.
- 4:48It becomes automatic. And, and, and now that's being
- 4:51incorporated into AI and it's, it's going to help, you know,
- 4:55advance it and, and make it more, much more broad and much
- 4:58more reliable. Can you drill into this a little
- 5:01bit? How does how do you make that
- 5:05transference of understanding the brain into encoding this
- 5:11into a digital system? It's all about architecture.
- 5:14That is to say, you have the same units, but you could put
- 5:16them together in different layers and with different
- 5:18connectivity. And you know, in Transformers,
- 5:21which is the, the, the, the heart of large language models,
- 5:25there's, there's special things that were added in order to be
- 5:28able to make it a, a, a powerful language learner.
- 5:33And by the way, that it was trained completely on the basis
- 5:36of just predicting the next word in a sentence, right?
- 5:39It, it wasn't given any special training on any special tasks.
- 5:42But now that we trained it up on a huge database that can solve
- 5:45many, many natural language problems that before were very
- 5:48difficult and then required specialized networks.
- 5:52And so that's one step closer to the fact that we know we have a
- 5:56very general, we, our brains are very general in terms of being
- 5:59able to handle lots of different problems.
- 6:02But OK, so reinforcement learning in particular is, is
- 6:06really, I, I think a part of the brain that is absolutely
- 6:12essential. It's all about practice.
- 6:14If you want to learn how to play tennis, there's two options.
- 6:17You can read books about tennis, right?
- 6:19That's the declarative part, the cortical part, the conscious
- 6:22part. Or you could go out on the
- 6:24tennis court and start hitting them and getting feedback and
- 6:27getting better and better day after day, week after week.
- 6:29And, and that eventually automatizes it so that you don't
- 6:34have to think about it anymore. And, and, but of course, that
- 6:37means you can't explain it to anybody.
- 6:40So if someone asks you, how did you do the serve?
- 6:42Well, you put your hand up and you hit the ball right.
- 6:45And that does it really help them?
- 6:47What about explain ability? Can you discuss the challenges
- 6:51of explainability that arise from neural networks and machine
- 6:57learning? And this has become very
- 6:59important and in business and and Justin.
- 7:03Let me first point something out that's obvious, which is that
- 7:07everybody has a brain and uses it to solve problems and doesn't
- 7:13know how it does that right. The brain is not explainable is
- 7:17that it doesn't give us insight to how we solve the problem,
- 7:20right? But, and we make mistakes, we
- 7:23hallucinate. Actually, we we make things up,
- 7:26filling in the blanks, right? We can't remember in detail what
- 7:29happened, you know, a couple years ago, but we can kind of
- 7:31fill it in. So, you know, this is not, you
- 7:33know, it's unusual that we're creating a system that is like
- 7:39the brain in some ways that it's also going to have some of the
- 7:42problems of the brain. So OK, but now let's look at
- 7:43explainability. Explainability is not just one
- 7:47level of understanding and, and let's take medicine, right?
- 7:51Well, you go to the doctor, you get a pill and the doctor will
- 7:53tell you, oh, it's going to help your immune system, you know,
- 7:57and maybe that's good enough for the patient, but it's actually
- 7:59not good enough if you're, you know, trying to do research on
- 8:03medicines and try to understand, you know, how it works.
- 8:07And that's another level explanation.
- 8:09But, you know, most, most people would be, you know, you know,
- 8:13don't have the background to understand that you have to
- 8:15have, you know, be an expert. But even that, OK, there's a lot
- 8:19of going on in the brain that that even experts don't
- 8:23understand. It's just really complicated,
- 8:26You know, all of the different molecules interacting and
- 8:29signals going back and forth. You know, we, we have a very
- 8:32crude understanding of, of really how, how nature put
- 8:37together the, all the different organs in the body and how they
- 8:41interact with each other. We're discovering that the, the
- 8:44gut actually has a connectivity, you know, and send signals back
- 8:48and forth with the brain, the gut, you know, the microbiome,
- 8:51right? This is something we didn't know
- 8:53until recently and we're learning things every day now.
- 8:56But you can go even down deeper. You can go down and
- 8:59understanding, you know, the actual details of the molecules
- 9:01and, and this is physics explanation, right?
- 9:04This is something that is at the level where you have equations
- 9:07and, and that level of explanation.
- 9:09All these explanations, you know, are, are, have different
- 9:13ways of, of giving us insights and, and you know, the, the one
- 9:17that we're looking for. Is it words?
- 9:19No, I think we're, we're really, we really want to get an
- 9:21understanding of what's happening at that deepest level.
- 9:24And that's beginning to happen because we have a lot of very,
- 9:28very good mathematicians and people, engineers with the
- 9:32training and machine learning are digging down into these
- 9:35networks and figuring out this the actual way that they are,
- 9:40you know, how the information flows through the network and
- 9:43how how it's able to do the amazing things that it does.
- 9:46And and, you know, this is really where we're going to get
- 9:50the real understanding, not just in terms of some explanation
- 9:53that, you know, you could live, the explanation that you might
- 9:57use just to make you think that you've understood something.
- 10:00As we rely more and more on AI to make decisions, including
- 10:07medical decisions, you, you just alluded to that this issue of
- 10:13explainability will become more and more important.
- 10:17So the real question for from that I have is can we ever
- 10:23really understand how the machine is making these
- 10:27decisions? And then what do we do about the
- 10:30fact that we that we're asking machines to ultimately decide
- 10:35important aspects of our lives and we don't know fully what's
- 10:40going on? You shouldn't think of AI as
- 10:42being a human. It's not right.
- 10:45It's it has special capabilities.
- 10:47It can, it can take in much more knowledge than any single human,
- 10:50right. So it's, it's really in some way
- 10:52already super intelligent, although people argue that it's
- 10:55not intelligent, but nonetheless it's helpful.
- 10:58So AI is a tool. It's going to help you solve
- 11:01problems and, and you talk about medical of decisions that have
- 11:05to be made. So a study was done.
- 11:07It was published in nature. Nature is the gold standard for
- 11:11science, all sciences and engineering.
- 11:16And, and what they did was a study on lesion skin lesions,
- 11:20some of which are cancerous and some of which are benign, right?
- 11:23And it's like 2000. And what they did was they
- 11:25compared experts or panel of experts with the best AI
- 11:31solution. AI was given thousands and
- 11:34thousands of examples, images of of lesions and it was taught to
- 11:38classify them. Now it turned out that both
- 11:41groups had a performance of about 90%.
- 11:44They were correct in telling you what kind of lesion or whether
- 11:47or not it was cancerous at about the same level.
- 11:50OK, so that's interesting. However, if they let the doctor
- 11:54use the AI in making the diagnosis, the performance went
- 11:58up to 98%. How could that be?
- 12:01I mean, they, they started with 90.
- 12:03Well, the reason is that they have different knowledge.
- 12:06The AI had access to much, much more data than any individual
- 12:11could ever see in their lifetime in terms of different rare skin
- 12:14conditions. And the doctor had much deeper
- 12:17knowledge about the the ones that are the most common.
- 12:19So, you know, and and it's a Doctor Who's calling the shots,
- 12:22but the doctor takes advantage of this other partner in order
- 12:28to make better diagnosis. And so this is what what's
- 12:31laying out now in every area where it's being used by
- 12:35writers, by people who have to, you know, work on ad copy,
- 12:42people who have, you know, engineers are trying to solve
- 12:45problem. By the way, the by far the most
- 12:47impact that a is had over the last 10 years has been in
- 12:51science has transformed science, because science is filled with
- 12:55big data on specific projects that are really very complex,
- 13:00you know, big particle accelerators, huge, huge amounts
- 13:04of data that's being generated. And, and it's it's really
- 13:07transformed the way scientists do science.
- 13:10And so this is really a harbinger of what to expect in
- 13:12other areas. Please subscribe to our
- 13:16newsletter, go to cxotalk.com, subscribe to our newsletter.
- 13:20And we do have some questions coming in.
- 13:22So let's jump to LinkedIn and we have a question from Greg
- 13:28Walters who said he read some somewhere that AGILLMSAI,
- 13:37digital twins, etcetera visualize actions and processes
- 13:43like athletes and and more a million hours of study in one
- 13:48second. Is this correct?
- 13:50And how does this visualization impact model collapse and does
- 13:56it? This gets to the heart of of how
- 13:58you actually create one of these large language models.
- 14:01You start with a huge amount of data, trillions of words, and
- 14:04you go through this process of, of, of predicting the next word
- 14:10and gradually getting to the point where you can with very
- 14:13high accuracy. And the only way you could have
- 14:15gotten that where that point is by being able to disambiguate.
- 14:19Words have multiple meanings and you have to look at the context.
- 14:22And so it has, it has to create an internal representation of
- 14:25the semantics, the meaning of the sentence, in order to be
- 14:28able to understand just the way you do it, right?
- 14:30You can predict the next word, but you know, you have to
- 14:33understand what the sentence means before you can do that.
- 14:36So but that takes a huge amount of time and and very expensive.
- 14:40The the the latest GPT 4, for example, took two months with
- 14:4420,000 GP US of graphics processing units and cost $100
- 14:49million, right. So, but the point is that not
- 14:52only now can the, can the, the large language model chap GPT
- 14:58answer new questions, but it can do that like in you press the
- 15:03button and you get the answer in, you know, one or two
- 15:05seconds, right? And, and that's shocking.
- 15:07I, when I first did it, I just couldn't believe this is how
- 15:09could this be? And the reason is it's
- 15:12internalized all this information so that it
- 15:14immediately is able to respond literally, you know, without
- 15:19actually having to think. In fact, these large, they don't
- 15:22think, they don't think the way humans think.
- 15:25And that's one of the things we're going to try to improve.
- 15:27The next generation is going to actually have internal self
- 15:32generated activity. Right now, if you stop your
- 15:35dialogue with ChatGPT, it goes blank.
- 15:38It doesn't keep thinking about things the way that you do.
- 15:41If you're on your own without any sensory input, You think
- 15:44about planning, you think about what happened, you know, in the
- 15:47past. Chat GDP doesn't do that.
- 15:50It doesn't have an internal life.
- 15:52This is from Arsalan Khan and on Twitter.
- 15:57And he says artificial intelligence is a mirror of
- 16:01human biases. How can AI and our brain become
- 16:08less biased quickly? And I I think this leads us
- 16:12directly into a discussion of the impact of AI on our
- 16:16workforce and work displacements and all those implications.
- 16:21The bias could come in because the data are, are incomplete or
- 16:26skewed. And, and, and, and this came
- 16:28out, for example, when face recognition was being used.
- 16:31It was much, much, much more accurate with white faces and
- 16:35black faces. And it turns out that there were
- 16:36many fewer black faces that were used in a training set.
- 16:39So you go back to the training set, you have to curate it.
- 16:43You have to put, make sure there's, and now we're
- 16:46discovering it. The higher the quality of the
- 16:47data, the better the performance in general.
- 16:51But there's other things you can do.
- 16:52And this is, this is a little bit more tricky right now.
- 16:56You know, you give, you have to give a goal because it is, you
- 16:59know, deep learning networks don't have intrinsic goals the
- 17:03way that humans have goals. So you have to give it the goal.
- 17:08And the goal typically is I want to optimize performance on a
- 17:11particular class of problems. And so you give it a lot of
- 17:14examples and it, it chugs away and, and it gets better and
- 17:18better and better and better. And the larger and larger the
- 17:20network, the, the better and better it gets.
- 17:23But the point though, is that it, it may not be aligned with a
- 17:26lot of other values that you have, like fairness, right?
- 17:30And So what do you do? Well, you're going to have to
- 17:34add fairness as one of the goals that it has.
- 17:38And now you have a problem. OK, I want maximum performance,
- 17:42but I also want maximum fairness.
- 17:44I can't have both, right? So I have to wait them.
- 17:48I have to decide, well, half and half or maybe 90% performance
- 17:52and 10% fairness, you know, that that's a decision has to be made
- 17:56explicitly. And you know, we're somehow
- 17:58we're making it implicitly. And when we deal with the world,
- 18:02you know, given the cultural biases that we have now, here's
- 18:06something about bias, though, and, and I just want to put this
- 18:09to you, OK, I'm going to ask a question of you and your
- 18:12audience, OK? You know, humans are definitely
- 18:16biased, right? We see this all time and LO, Ms.
- 18:21reflect that to some extent because they're trained on human
- 18:23data. Here's the question, which do
- 18:26you think are going to be easier to fix, the LLM or the human?
- 18:30I would. Assume the human is easier to
- 18:33fix because with the human you have control in quotes over one,
- 18:40whereas to fix bias in the LLM you need to adjust a very large
- 18:45data set. Actually, I'd be curious to know
- 18:47what your audience thinks. Folks, you heard the question.
- 18:51In my experience, I've known a lot of different people and very
- 18:55rare that they change their minds about it.
- 18:57They're really fixed in their ways.
- 19:00You know, as you get older, I mean, maybe younger children,
- 19:04OK, there, there, there's some hope there of, of changing their
- 19:06by. But you come into the world, you
- 19:08don't know what the language is. You don't know what the values
- 19:10are of the culture. And you learn through
- 19:13reinforcement learning, by the way, you know, what's, what's
- 19:16things that you, that you should, how, how to deal with
- 19:19relationships, how, how to value, you know, you know,
- 19:23people's advice, think things that you know, that are
- 19:27different in different cultures, how you deal with, you know,
- 19:32truth and so forth. It's it's different, it's not
- 19:35the same and and you learn. And Greg Walters on LinkedIn
- 19:41agrees with you. He comes back and he said it's
- 19:44the LLM that will be easier to. That's because we engineered it.
- 19:48We can fix it. But how?
- 19:50How do you fix that bias given there are so many levels of
- 19:56insidious? OK, OK, OK, so he he in my book,
- 20:00you know, chat GDP and the future of I, I actually laid out
- 20:04because I there's a there's a lot in the book that is is is
- 20:07really inspired by the brain. And by the way, there's a
- 20:11revolution going on in neuroscience that is equally
- 20:14important as what's happening in AI.
- 20:16And this has to do with the brain initiative, Obama's brain
- 20:19initiative. It's been 10 years on and we
- 20:22have made enormous strides in understanding the mechanisms
- 20:27underlying your brain. And so we're, we're going to
- 20:31actually be able to really, you know, get, get, come up again
- 20:35with a much better explanation for the decisions that we make.
- 20:39But, but, but here, here's, you know, when, when we're dealing
- 20:43with, as I told you with for humans, you use reinforcement
- 20:50learning. So here's what has to happen in
- 20:51AI instead of just training the network at the beginning.
- 20:55And then that's it. You, what you're going to have
- 20:57to do is use reinforcement learning all along.
- 21:00We need lifelong learning where we have reinforcement learning
- 21:02going hand in hand with the, the, the declarative, it's
- 21:06called the declarative knowledge system.
- 21:09And, and so we have to basically treat the LLM like a child,
- 21:13right, You know, and, and punish the child when it does say
- 21:16something wrong, does something wrong, and that how else is it
- 21:19going to learn that right? We have to align it with our
- 21:21values. And, and, and that's beginning
- 21:24to happen. And I know a couple of companies
- 21:25already used it, not for this particular purpose, but for
- 21:29example, something called chain of thought.
- 21:30When you want to solve a problem, you break into pieces
- 21:33and you solve each one separately.
- 21:34Mathematicians do this all time when they're proving a theorem.
- 21:37But the LLMS, if you ask them, they just jump to the answer and
- 21:41sometimes they're right, sometimes they're wrong.
- 21:43But now you can actually train using reinforcement learning,
- 21:46this habit of breaking into pieces and solving each one
- 21:50separately. And now it's gone.
- 21:51The performance has gone on these mathematical problems from
- 21:5420% to 80%. So this is an example, you know,
- 21:57you're using reinforcement learning.
- 21:59It's actually in part a, a way of a way of solving problems and
- 22:04actually a method that could help not not just solve
- 22:07mathematical problems, but any problem that requires, you know,
- 22:10in, in in manufacturing, in figuring out how to reconfigure
- 22:15your office, all of that. Should we should be able to
- 22:19help. We should be able to explain,
- 22:21you know, through a sequence of actions, how you're going to get
- 22:24to some goal. And in answer to your question
- 22:27to the audience of which will be easier to remove or address bias
- 22:32in the LLM or in people, Isaac Sacolic comes back and he says
- 22:38it's easier to fix AI than humans.
- 22:40People don't change belief systems easily, and data and
- 22:44weights can change in an AI. The real issue is understanding
- 22:47what values the data scientists embedded in their training.
- 22:51So he's in agreement with you. And Arsalan Khan is also in
- 22:57agreement. Everybody's against what I said.
- 22:59Arsalan Khan says LLMS might be easier to change since the input
- 23:06is from multiple people, whereas people rarely changed.
- 23:10Exactly your point. But now let's talk about the
- 23:15implications of all of this on work, the workforce, jobs.
- 23:23Thoughts on that? We are at the stage that the
- 23:26Wright brothers were at the beginning of aviation 100 years
- 23:30ago. The very first flight by the
- 23:32Wright brothers was 10 feet up and 100 feet forward.
- 23:36That was it. They were inspired by how birds
- 23:39glide by the airfoil. They were inspired by the
- 23:42lightness of the wings, and that was why they use canvas.
- 23:47But the most difficult part of making aviation safe was
- 23:51figuring out how to regulate, you know, and how to control the
- 23:56direction you're going and, and regulate it so that you don't
- 23:58crash. That was the difficult problem.
- 24:02Well, we're, we're living through that right now, except
- 24:04that maybe it won't take 100 years.
- 24:06It'll only take, you know, 1020 years.
- 24:09So we're, we're, we're, we're just at the very beginning of
- 24:12that process. And what's going to happen over
- 24:13the next 10 years near term is incremental advances and that's
- 24:18already happening. You see all these language
- 24:20models, there's now a dozens of them out there with different
- 24:24capabilities and, you know, focused on different problems
- 24:28that have to be solved. So, you know, we're, we're just
- 24:30exploring that and, and one of the directions that we're going
- 24:34is actually not bigger and bigger because we've run out of
- 24:37data, but smaller and smaller with quality data, with data
- 24:43that's more specific to your particular company, to your
- 24:47particular profession. And that now is, is going to
- 24:51shift the balance away from these big high tech companies
- 24:54toward, you know, the enterprise, the companies, the
- 24:58ones that are actually doing, you know, the, the, by far the
- 25:01most important job in society, which was, you know, making
- 25:06things and getting things to work and, and dealing with the
- 25:10complexity of the world. And that, that could be done.
- 25:14This, this is something that is, is, is, is not going to take
- 25:18just five years. It's not going to take 10 years
- 25:21because it takes is you have to complete reorganization of the
- 25:25way that the, the, your office and your company is organized.
- 25:30So right now, for example, I'll just give one example.
- 25:35You know, you have databases all over the place, especially if
- 25:38you're an international company and you know, how do you
- 25:40integrate across databases? Well, I mean, you have an IT
- 25:43person who has ways of doing that and, but you know, to
- 25:46answer your question, that requires a lot of comparison
- 25:49and, and, you know, accretion of, of, of, of different
- 25:54sources. That's, that's going to take a
- 25:57long time. But it one of the things that
- 25:59LMS are really good at it because it's already done that
- 26:01for the database of the world. Why not for your company?
- 26:04All of those databases could be integrated.
- 26:07And now when you want to answer a question, you don't go to the
- 26:09IT guy. You ask your, your ChatGPT.
- 26:13Well, you'd call it something else, right?
- 26:14You know, it would be called chat IBM, you know, and, and
- 26:19that way you're, you're going to be able to have access to
- 26:22instantly to all kinds of important questions that you can
- 26:25answer that may help you make decisions.
- 26:28This is if you're, you know, of course, a head of a company.
- 26:31What about if you're in the middle of the company?
- 26:33Now, one of the things to keep in mind is that these tools
- 26:37require training. You know, chat GDP
- 26:41out-of-the-box. You can ask the questions, but
- 26:43the answers you get back are often not the most useful.
- 26:46That's because like any tool, you have to learn how to use the
- 26:49tool. And you know, there's a, a whole
- 26:52group of people now that call themselves prompt engineers.
- 26:56And what, what, what does that mean?
- 26:57It means that they know how, how to use the prompt in order to be
- 27:03able to more quickly get useful answers out.
- 27:06And, and in my book, I go into this in quite some detail that
- 27:11may help, that may help them. And, and the other thing is, is
- 27:15it's, it's really interesting. It's the, you know, if you
- 27:18really want to incentivize people, you, you know, you what
- 27:23you have to do, dealing with chatty P is fine.
- 27:27People do it on their own. People just doing it because
- 27:29it's a lot of fun, right? Well, once you, once they get
- 27:32trained on doing this, this is going to help with productivity.
- 27:35It's, it's, it's already doing that in the many areas, like I
- 27:38gave the example of medicine, but it, it's, it's going to help
- 27:42people in ways that nobody ever even imagined, right?
- 27:45I mean, this is, it's, it's, this is something that like, you
- 27:50know, what was the killer app for personal computers, right?
- 27:53It, it wasn't, you know, simply typing and, and you know, the
- 27:58letters, it was Lotus 123. Do you remember that?
- 28:02It was a spreadsheet and, and that suddenly that was something
- 28:06that's very useful that could be used by not just individuals,
- 28:10but by companies and organizing all the data and so forth.
- 28:14And, and that has become central to the, to the way that now we
- 28:18all deal with large data sets as we, we have them in these super,
- 28:24you know, you know, the spreadsheets now are, are, are
- 28:27much better design than Louis 123.
- 28:29But the same thing is going on right now is that we're, we're,
- 28:32we're beginning to develop the tools that are specific and in
- 28:37surprising ways. Let me give you one example of
- 28:39something that surprised me. It's in my, and again, it's in
- 28:42my book. So a lot of people have mental
- 28:46problems, you know, they have anxiety, they have phobias and
- 28:50so forth. And, and you know, and, and
- 28:52worse and mental disorders, but you know that what do they do?
- 28:55They go to a psychiatrist or a cognitive therapist and you
- 29:00know, they, they have to make their appointment weeks ahead of
- 29:04time. And then, you know, they go and
- 29:07then, you know, they have an hour and then they get a big
- 29:10bill right now. You know, this is, this is, you
- 29:14know, it, it helps, it really, it turns out cognitive therapy
- 29:17is as good as taking a pill. It really is.
- 29:21It's and, and it's complementary, by the way.
- 29:24Now they did a test. This is a scientific study.
- 29:29If people are given the choice between an AI psychiatrist or a
- 29:34human psychiatrist, what would they choose?
- 29:36What do you think, Michael? What do you think?
- 29:38I will say that they choose the AI over the psychiatrist you'd.
- 29:44Figure this out from my first trick question.
- 29:48Yeah you're absolutely right. But majority of, of humans feel
- 29:51more comfortable talking about, you know, personal things that
- 29:56it might be a little embarrassing, right?
- 29:58If they're talking to another human who are, you know, humans
- 30:00are very judgmental, right? I mean, but the AI, you know,
- 30:03it, it's not going to judge you away.
- 30:06It doesn't have any internal care, but you know, it's going
- 30:09to give you respond in a way that is objective and so forth.
- 30:14So you know and you know this is this already happened back at
- 30:19MIT in the early days of AI when someone put a very simple
- 30:23program that it did very simple thing that just repeated the
- 30:25question. Eliza.
- 30:27Eliza And, you know, a Weisbaum Weisenbaum discovered that his
- 30:32secretary is using it during the lunch break.
- 30:35You know, it's like, and you know, it didn't, she didn't care
- 30:39that it was a machine as in a very simple minded machine.
- 30:41You know, she did it because it kind of helped.
- 30:44It got her to think about things, you know, explicitly.
- 30:46So there you go. I mean, that's that's
- 30:48unexpected, right? There's a lot of unintended
- 30:51consequences. And by the way, not all good.
- 30:52There's going to be bad ones too.
- 30:54And so that's why we have to really be careful about, you
- 30:57know, what, what to expect. You have just been describing
- 31:02what I would call incremental changes of efficiency, right?
- 31:08We now have a critical thinking or an information partner to
- 31:14help us improve our writing to research and so forth.
- 31:17But what about the more that the deeper implications and more
- 31:25foundational changes that this may drive in society and in in
- 31:30the workplace? Any any thoughts on that?
- 31:33In the short room, you're not going to lose your job.
- 31:35Your job is going to change and it's going to change because you
- 31:39have these new tools. You're going to have to learn
- 31:41new skills. And that's why it's really
- 31:44important that we have within companies and, and it turns out
- 31:48that we have a, a ways of doing that that could help companies.
- 31:52This is called massive Open online courses and they're free
- 31:56and there's a lot of out there now there's like, you know,
- 31:5810,000 and I have one myself called learning how to learn.
- 32:02And this is with Barbara Oakley. And what we realized was a lot
- 32:05of people, especially 25 to 35, half of them are college
- 32:10educated, are in the workforce and now they need new skills.
- 32:13And it's harder to learn when you're in that position because
- 32:16you can't go back to school and you have mortgage and children.
- 32:20And so you know, you want to be able to be more efficient.
- 32:23We teach you how to be more efficient through what we know
- 32:26about the brain, how the brain learns.
- 32:28In any case, they're there, but there are all kinds of mooks out
- 32:31there about how to use the new tools of and the idea I think is
- 32:36for people now who have a job are going to be using your tools
- 32:42more efficiently. But gradually the job is going
- 32:45to morph. It's it's going to be the
- 32:47boundaries between jobs may change in terms of, you know,
- 32:50what people can do and integrating more often rather
- 32:54than having stacked layers and so forth.
- 32:57You know, it's, it's impossible to predict the future.
- 32:59Let me ask you, OK, Michael, you were around in the 90s, right?
- 33:04Yes. OK.
- 33:05You remember when the Internet went public, right?
- 33:08Yeah. Could you have predicted what
- 33:11influence the Internet would have on every aspect of your
- 33:13life today? Absolutely not.
- 33:15Not at that time. OK.
- 33:17Well, we have another technology that's that's going to have a
- 33:19equally important impact up the road that we can't predict.
- 33:23It's, it's just, it has, it has so many different effects, so
- 33:28many different parts of society. But Terry, I did not invent or
- 33:33discover important aspects, parts of the that underlie the
- 33:38Internet, but you did when it comes to deep learning.
- 33:42And so therefore I look to you as the the the prophet and the
- 33:47seer of what may come. The reality is that I'm actually
- 33:53probably no more capable than you are.
- 33:56In fact, maybe less capable of of predicting the future.
- 34:00You know, the Internet pioneers back then.
- 34:05Let me give you an example, OK. You know, Google, when it
- 34:10started had a motto, do no evil. Do you remember that?
- 34:14Yeah, absolutely. And, and, you know, that they
- 34:16were very idealistic. They said this is going to
- 34:19democratize information. Everybody is going to have
- 34:22access to information and their voices will be heard.
- 34:25OK, well, could they have predicted how that will play
- 34:31out? And here we are now we're we're
- 34:32dealing with, you know, misinformation, fake news, echo
- 34:37chambers. I mean, this is like, you know,
- 34:39who could have predicted that, right?
- 34:40I mean, this is something that even the experts didn't predict.
- 34:43The enthusiasm of young people before they became billionaires.
- 34:47That's right. That's right.
- 34:49So I'm a young, I was once a good person in any case.
- 34:54Yeah, we, we, we, we are, you know, I'm, I'm actually an
- 34:58optimist. I, I actually think that we will
- 35:00get it sorted out, but I know it's not going to be easy and
- 35:03it's going to take a long time. And and that's true of all
- 35:06technologies. It's not like it, this is
- 35:07different. It's just, it's just that it's
- 35:10so new that we really are just beginning to, you know, you
- 35:17know, trying to beginning to understand the capabilities it's
- 35:21moving target, it's incrementally improving there.
- 35:25There may be a breakthrough and it may not be, it may be, you
- 35:28know, I don't know, 10 years, maybe 20 years.
- 35:32And I think it's going to be the breakthrough, I think is going
- 35:34to be when you put something which call is called System 2
- 35:38into these large language models.
- 35:39System 1 is what we have right now, which is basically just
- 35:43trained to, to produce, you know, generate the generative AI
- 35:46one word after the next. But this, this system too, is
- 35:50the one I alluded to earlier, which is cell generated
- 35:52activity. And, and we, we actually know
- 35:55how that's done in the human brain.
- 35:57And so it's just going to be a matter of time before that's
- 36:00transplanted into AI. You know it, but it's not going
- 36:04to take place in, in, in 5 or 10 years.
- 36:07We're talking here for probably, you know, decades out.
- 36:10Chris Peterson on Twitter says the current hype based AI takes
- 36:17vastly more resource inputs per question than a human brain
- 36:21takes for 24 hours of doing everything.
- 36:24Can you talk about the sustainability, the power
- 36:28requirements, the resource requirements?
- 36:31You alluded to this earlier in your discussion of the trade off
- 36:35between completeness versus rooting out bias and the input
- 36:41cost and time associated with that.
- 36:43We have to set priorities. It can't things have been the
- 36:48cost of things have is gone through the roof and with that
- 36:51cannot continue. And I also alluded to the fact
- 36:54that the next generation they're going to be smaller language
- 36:57models with higher quality data. That's One Direction.
- 37:00The other direction is hardware and I have old chapter on this
- 37:04in my book, right? If you're interested go, you
- 37:06should go there. So the human brain is able to
- 37:11function at very high levels with 20 watts of power, some a
- 37:16little brighter than others, right?
- 37:18But you know that's much less than the the gigawatts that are
- 37:23out there being used right now to answer your questions.
- 37:27So there's clearly a huge, huge imbalance.
- 37:30And that's because the technology that nature uses is,
- 37:33is based, is gone down to the molecular level.
- 37:36It's really, really advanced. But you know, we're, we're
- 37:40actually, there's a hundreds of companies out there that are
- 37:43building more efficient technology that is going to make
- 37:47lower power much more the, the architecture that you need for
- 37:54these deep learning networks is completely different from the
- 37:57von Neumann architecture that is in your PC.
- 38:00That architecture does 1 instruction at a time and it's a
- 38:06has to fetch information from memory, which is separate.
- 38:08Well, in your brain, the memory and the processors are one in
- 38:14the same. They're, they're integrated and
- 38:16there are a lot more processors, right?
- 38:18This is all the neurons. You have 100 billion neurons in
- 38:20your brain. And now we're up to, you know,
- 38:23it, it, you know, literally 100 trillion of synapses in your
- 38:29brain, we're up to a trillion weights in these networks.
- 38:33So we're still like only 1% of the brain even just in raw
- 38:37power. But here's the beauty is that we
- 38:40can special purpose chips that could be much more efficient
- 38:44and, and and that will take us down by a couple orders of
- 38:47magnitude over the next 10 years already going on.
- 38:49And this has happened to GP us are a good example.
- 38:51That is A2 orders of magnitude when they put GPT, when they
- 38:55transplanted these original deep learning networks onto GPUs, it
- 39:01was tours of magnitude more efficient.
- 39:03And the reason is that they have thousands of processing units
- 39:06just like all the neurons working in parallel exchanging
- 39:10information. So that was that was a big step.
- 39:12That was a huge jump. And that there's an inflection.
- 39:15The the Moore's law went from doubling every two years to
- 39:18doubling every two months, literally.
- 39:20I mean, this was a huge, huge change.
- 39:22Now that's that's going to happen again.
- 39:24And it's, it's, it's, it's going to become probably another two
- 39:28orders of magnitude or three, possibly with a, a new class of
- 39:33computing called neuromorphic engineering.
- 39:36What is that? Well, it turns out that if if
- 39:39you take your digital chip, which are basically 0 ones,
- 39:42right, and run them near 0 where things are analog, by the way,
- 39:46digital chips are actually analog when you look at the
- 39:48actual transistors, but they run into the rail.
- 39:52But if you work down there, and this is Carver B, that Caltech
- 39:55actually understood this, that you can actually take advantage
- 39:58of that processing at that low level for doing crude multiplies
- 40:02and adds much more efficiently by by like two or three orders
- 40:06of magnitude. And now that's mature
- 40:07technology. And so that's going to come
- 40:09online over the next 10 years. So see, what's going to happen
- 40:12is just like, you know, Moore's laws, that the technology gets
- 40:15more and more efficient. And, and we're, we're just at
- 40:17the beginning of that because we, you know, we realize now
- 40:21that this architecture may be more important than, than the
- 40:24von Neumann architecture for solving many, many kinds of
- 40:27problems that before, you know, we couldn't.
- 40:30This is from Isaac Sukolic and Isaac is a big time CIO
- 40:36influencers. A lot of CIOs listen to him and
- 40:39he says he liked your healthcare example.
- 40:43It may be hard to benchmark human versus human plus AI
- 40:49versus AI and other diagnostic areas.
- 40:53What are options to validate accuracy and build trust with AI
- 41:00recommendations? I think that's a really
- 41:02important question. That's productivity and, and,
- 41:05and that's the bottom line actually.
- 41:06You know, for a long time, computers were you the, the, the
- 41:10whole business world and, and, and, and public are investing in
- 41:14digital computers and productivity didn't change very
- 41:17much. It wasn't until you connected
- 41:18them together that they could exchange information.
- 41:21That's when things took off. Let me go back to just answer
- 41:26the question in a concrete way. Let's go back to healthcare.
- 41:30So, you know, you go to a doctor for, you know, you have some
- 41:34medical problem and you know, you walk into the office, you
- 41:37have 20 minutes doctor sitting there next to a computer and is
- 41:40asking you questions, you know, about your symptoms, about your
- 41:45previous medical history. What is he doing?
- 41:46He's looking at the computer he's typing in because you got
- 41:49to get all that in the computer because you have to have a
- 41:51record, right? And then, you know, he's not
- 41:53looking at you. And then at the end, you know,
- 41:56he, he, he, he gives you a list of, of drugs that you should
- 41:59take or some advice and so forth.
- 42:01And, and off you go. And you know, you, maybe you
- 42:04get, you remember half of it if you're lucky, right?
- 42:07So here's what's happening right now.
- 42:08There are 10 companies out there that have realized I could put
- 42:12together a lot of different AI systems to make that process of
- 42:15not just a lot more efficient, but actually much more likely to
- 42:20be of help to the to the patient.
- 42:23So first of all, speech recognition, the doctor doesn't
- 42:26have to look at the computer. He can talk to the patient.
- 42:28It turns out you get a lot of information by looking at the
- 42:30patient, not just, you know, what they're saying.
- 42:33You can, you can you can tell from their facial expressions,
- 42:35you know, how serious something is.
- 42:36You can look at their face, you know the color of the face.
- 42:40I mean, a good diagnostician use can use that information right
- 42:44and, and jump to, you know, to conclusions that you can't just
- 42:48with a couple of numbers that they are that that you're
- 42:51getting, you know, typing in and OK, now what happens?
- 42:55So you press a button at the end and much more human interaction
- 43:00and now chat GDP does the heavy lifting.
- 43:03It gives you a summary. It's very good at that doctor
- 43:06looks at it and checks to see what's you know, if there's any
- 43:09problem in terms of the recommendations and so forth.
- 43:12And now you know that goes back and, and the the patient now
- 43:16walks out with actually very detailed summary.
- 43:19And so now the patient is going to be able to understand much
- 43:21better what needs to be done over the next couple of months,
- 43:24you know, in order for the patient to be able to, you know,
- 43:29improve their, their health. And, and you know, this, this is
- 43:32just you kind of one example of how we have a system right now,
- 43:36which is really disadvantages patients, right?
- 43:40And doctors, by the way, doctors don't just stop because they
- 43:44have to write up notes afterwards and often late into
- 43:46the night. And now all of that is being
- 43:48done. They could, they can go and
- 43:50they, you know, this is something that it's going to
- 43:52take decades to permeate the medical system.
- 43:55You know, they're very conservative.
- 43:56So it's going to take a long time and it has it'll have flaws
- 43:58and it'll there'll be problems, but ultimately I think that
- 44:02it'll greatly improve healthcare.
- 44:03Chris Peterson has a question, and very quickly please how does
- 44:08the enormous cost of retraining the big models play into GDPR
- 44:13style right to be forgotten laws?
- 44:17Will AI systems need an exception because it's just not
- 44:21practical to retrain for each request and very quickly?
- 44:25In addition to the initial training, which is the costly
- 44:28part, that is also opportunities later to do something called
- 44:31fine tuning. And what's that mean?
- 44:33Fine tuning is basically giving it extra data and trying to
- 44:41train it in a, in a gentle way such that it now has access to
- 44:46new information without interfering with the existing
- 44:50database. Now, it, it's not perfect.
- 44:52What happens is that the, the, the large language model gets
- 44:55dumbed down to some extent because it's taking on this
- 44:58extra burden. But that's not nearly as
- 45:01expensive. And that can be done in house,
- 45:03right? That could be done in individual
- 45:04businesses. You, you have to hire somebody
- 45:06who, who can do that for you. You know, that these are machine
- 45:10learning people. And by the way, the, the, you
- 45:13know, I, when I started, we were doing neural networks, right?
- 45:16And that was in the 80s. And then it'd be overtime
- 45:20Neurips became a machine learning conference.
- 45:23But now it, it, it's come back, you know, full circle.
- 45:26And now it's, it's not called neural networks anywhere.
- 45:28It's called AI, right? You know, and it wasn't because
- 45:32we called it that. It was because the world was
- 45:34calling it that. But it's really about machine
- 45:36learning. That's the heart of AI today.
- 45:39And it's all about data. Whoever has more data wins in
- 45:44any area. And I'll tell you your company
- 45:47is sitting on enormous amounts of data that's really important
- 45:50for you. And if you can get that into an
- 45:53AI, you will win. Can you talk about how the
- 45:56assumptions made during data curation influence AI outcomes
- 46:03usually? This is really going to become a
- 46:07very large business data, as I said earlier, data is hard to
- 46:13come by quality data even harder.
- 46:18And and so, you know, how do, how do you, how, how is it that
- 46:26we're going to be able to overcome a lot of the problems?
- 46:33You know, hallucination is a good example.
- 46:37You know, hallucination actually is, is has has advantages.
- 46:42If you're a writer or, you know, ad copy, it's really good at
- 46:46that or poems, but not so good if you're asking for a fact
- 46:52hallucinates, you're in trouble. So I think what we're going to,
- 46:54we're going to need is another layer on top of the, of the
- 46:57language models, which is, is, is a fact checker basically.
- 47:01And, and will, will take what is coming out and we'll learn how
- 47:05to then look through the database world's database and,
- 47:08and come up with, you know, sources for that particular
- 47:13fact. So again, it may be this isn't
- 47:18really the exciting thing is that right now humans are doing
- 47:20that. You know, that all these big
- 47:21companies have thousands and thousands of humans that are
- 47:24going through and sorting through getting rid of racist
- 47:28data, trying to, you know, make prevent of these large language
- 47:33models from saying from doing anything that is going to hurt
- 47:35people. You know, that that is being
- 47:37done on a individual basis. And that's very labor intensive.
- 47:41But you know, we, we should be able to train another generation
- 47:46of large language models to actually do that part of the job
- 47:49of, of, of, of sorting through and flagging things.
- 47:53And, and, you know, even in the earliest days of, for example,
- 47:56neural networks in the 80s, when I started, they were using these
- 48:00with very, very, at that time, they were very shallow with one
- 48:02layer of pin units to take a slide with cancerous, it was
- 48:08called a Pap smear, you know, the cervical cancer and, and
- 48:12humans would have to look through the every cell and it
- 48:14would take hours and hours and very costly.
- 48:16Well, you know, neural networks could do that much more quickly,
- 48:19reduce it to 100. And then the humans were going
- 48:22to, you know, use their much more powerful visual system to
- 48:27look at the, the debris and, and, and all the different types
- 48:30of cells and, and, and come up with a much actually thorough
- 48:36and faster way of analyzing the data.
- 48:38And so the idea is we have layers and, and humans are one
- 48:42of the layers. They could be the layer at the
- 48:44top or they could be in the middle or, you know, anywhere.
- 48:46But the idea though, is that we have different systems that are
- 48:49there for spotting different problems in the data.
- 48:53And, and this is, again, this is where we're at the very
- 48:55beginning of this, this is going to take decades.
- 48:59You know, it, it, it took it literally.
- 49:01Just think of the Wright brothers, right?
- 49:02Look how long it took to get to jet engines, right?
- 49:05And and jet planes now you know you can go across the world now,
- 49:09but back then, right, you could be be lucky to just go across
- 49:13the countryside. So, on the subject of humans and
- 49:17datas and data, how can we detect and understand the impact
- 49:22of adversaries manipulating data, for example, with
- 49:27misinformation and disinformation?
- 49:30Well, that's already happening. You know, we're social media.
- 49:32This is filled with that, you know, that that's, that's a, a
- 49:36very, very difficult problem. I don't think it's going to be
- 49:39any easy solution. I, I outlined just now what I
- 49:42think is going to happen with all these layers of, of, of, of
- 49:45filtering. But you know, there's, I think
- 49:51one thing that we should be open to is the fact that every once
- 49:56in a while there's a, a massive a breakthrough.
- 49:58The last one occurred in 2022 when chat GDP was opened up to
- 50:05the world and suddenly everybody realized, oh, my God.
- 50:09Yeah. What hath God wrought?
- 50:12That was the that was the first message sent over the Telegraph.
- 50:15And we're going to have it. I'm sure we'll have another
- 50:20moment like that. I don't know when, I don't know
- 50:22how, but you know, it's going to happen.
- 50:24It's going to happen, and it's already happened.
- 50:27And I told you in science, it's happened multiple times.
- 50:30I'll just give you one example. Protein folding, how does it a
- 50:33string of amino acids get folded up to become an enzyme that has
- 50:36a function and that turns out to be a computationally
- 50:40intractable. You can't do it.
- 50:41It just takes too much for computing power and, and most
- 50:44biologists thought it would never be solved, but we've
- 50:45solved it. Alpha fold DeepMind created a
- 50:49transformer that could solve that problem and and it's
- 50:52transformed biology. This is like a huge why because
- 50:55you can now design enzymes that have much more specific actions
- 51:00and better drugs, better ways of, of being able to make
- 51:04predictions about how different interactions are going to occur
- 51:08between proteins. And this is really a
- 51:12transformation in the whole business of, of creating better
- 51:16healthcare, better drugs. So it's going to happen in many,
- 51:19many areas in ways you can't predict.
- 51:22What advice do you have for business people given everything
- 51:27we've discussed? Number one, read my book, not my
- 51:33lips. Read the book.
- 51:36Because I talk a lot about these issues that, I mean, I don't
- 51:38really touch the surface of what is in the book.
- 51:40OK #2 the most important thing you should be thinking about is
- 51:44training your workforce with this new technology and take it
- 51:47seriously. It's, it's going to, it's it,
- 51:50this is just the beginning, right?
- 51:51And, and unless you start now, right, helping your workforce,
- 51:55we use these tools, the new tools and, and it really is that
- 51:59they're tools that if you can use them badly and, and, and
- 52:02very well. And you know, people are trained
- 52:04in that. And it does take time.
- 52:06It takes, you know, maybe 100 hours.
- 52:10That's a lot less than a 10,000 hours it takes to become an
- 52:12expert, right? And #3 download your company
- 52:16into ALLM. And I'll tell you, this is
- 52:21already being done with the brains of flies and zebrafish.
- 52:26These are brains that have 100,000 neurons.
- 52:28We can download them now into AI and and they have similar
- 52:33behaviors, they have similar activity patterns.
- 52:36You know that's the future. My mind is blown from this
- 52:40conversation. Terry Signowski, thank you so
- 52:43much for taking time to be with us.
- 52:46Oh, I'm, I'm very pleased. Thank you so much for making
- 52:49this opportunity possible for me.
- 52:51And a huge thank you to everybody who watched.
- 52:55You guys are amazing. You asked such excellent
- 52:58questions. Before you go, please subscribe
- 53:02to our newsletter, go to cxotalk.com, subscribe to our
- 53:05newsletter, subscribe to our YouTube channel, and we have
- 53:10amazing shows coming up. Thank you so much everybody, and
- 53:13I hope you have a great day. Take care.