AI, Deep Learning, and the Future of Work

Michael Krigsman: Welcome to CXO Talk, episode 860. I'm Michael Krigsman, and we are discussing AI and the future of work with Dr. Terrence Sejnowski, a pioneer in computational neuroscience and deep learning.

He's the Francis Crick chair at the Salk Institute for Biological Studies and is also a distinguished professor in the Department of Neurobiology at UC San Diego. He is also chairman of the NeurIPS Foundation, which runs the largest, most prestigious AI conference in the world. Terry has published over 500 scientific papers and 12 books, including his most recent, ChatGPT and the Future of AI, and has won more awards than I can name.

Terrence Sejnowski: I have a background in physics, which is basically a license to work on anything. But my real focus is on the brain. I've won awards in neuroscience, most recently the Brain Prize, the highest award. And I was a pioneer in the early days of the '80s developing learning algorithms for neural networks with Geoff Hinton and others.

And so I have these two different backgrounds, which I've now reached the point where this is really what's exciting is they're coming together. They're actually synergistic. And so what we learned in the brain is going to help better AI in the future and what we learn in AI is actually helping us in neuroscience. So this is really an exciting time to be doing research.

Michael Krigsman: How do you model AI on the brain? Maybe we should start there.

Terrence Sejnowski: Back in the 20th century, artificial intelligence was really focused on things that ran on puny digital computers, namely symbols, logic, and rules. And there's only so far you can go with that, especially when you're dealing with a complex, very complex world with lots of uncertainties in it.

And so you need much, much more computation. And now we know, in fact, nature has solved all these problems: vision, coordination, motor, and so forth. And so we're learning now, we popped the hood and now we're learning, what do you take from the brain? Massively parallel, highly interconnected parameters called synapses in the brain and learning that you learn the strengths of those synapses with experience.

And now it's big data and that has transformed all of AI. But the point is now we're dealing with the same mathematical structures. These neural networks are very simplified versions of brains, but they have the same mathematical way of analyzing and making progress.

Michael Krigsman: How does the brain organize information differently from machines and neural networks? And I just want to mention that the prominent data scientist, Anthony Scriffignano, has helped me prepare some of these questions.

Terrence Sejnowski: The general principles I've given you take you so far, but actually nature has evolved specialized circuits for specialized problems in different species in order for survival. And there are many, many differences. The deep learning architecture that is used today was inspired by the cortex, cerebral cortex, which is the highest level on top of the brain that humans use for the knowledge base.

And there's a hundred other brain areas that are equally important for being able to handle the complexities of the world and to survive. And a lot of those are ones that are right now being integrated into the next generation of AI.

I'll just give you one example, reinforcement learning. Reinforcement learning is the part of the brain that is found in almost all species and has to do with being able to, first of all, from experience, predict whether a particular action is going to give you a reward or not. And you build that up over many, many trials. If you're wrong, you diddle the weights.

And that's actually used for what's called procedural learning, learning how to play tennis, for example, or learning how to solve problems in law, in medicine, all of that becomes automatic. And now that's being incorporated into AI and it's going to help advance it and make it more broad and much more reliable.

Michael Krigsman: Can you drill into this a little bit? How do you make that transference of understanding the brain into encoding this into a digital system?

Terrence Sejnowski: It's all about architecture. That is to say you have the same units, but you could put them together in different layers and with different connectivity. In Transformers, which is the heart of large language models, there's special things that were added in order to be able to make it a powerful language learner.

And by the way, it was trained completely on the basis of just predicting the next word in a sentence. It wasn't given any special training on any special task, but now that we've trained it up on a huge database, it can solve many natural language problems that before were very difficult and required specialized networks.

And so that's one step closer to the fact that our brains are very general in terms of being able to handle lots of different problems. But reinforcement learning in particular is, I think, a part of the brain that is absolutely essential. It's all about practice.

If you want to learn how to play tennis, there's two options. You can read books about tennis - that's the declarative part, the cortical part, the conscious part. Or you could go out on the tennis court and start hitting them and getting feedback and getting better and better day after day, week after week. And that eventually automatizes it so that you don't have to think about it anymore. But of course, that means you can't explain it to anybody. So if someone asks you, "How did you do this serve?" Well, you put your hand up and you hit the ball, right? And that doesn't really help them.

Michael Krigsman: What about explainability? Can you discuss the challenges of explainability that arise from neural networks and machine learning? And this has become very important in business.

Terrence Sejnowski: Let me first point something out that's obvious, which is that everybody has a brain and uses it to solve problems and doesn't know how it does that. The brain is not explainable. It doesn't give us insight to how we solve the problem. And we make mistakes. We hallucinate actually, we make things up, filling in the blanks. We can't remember in detail what happened a couple years ago, but we can kind of fill it in.

So this is not unusual that we're creating a system that is like the brain in some ways that it's also going to have some of the problems of the brains. But now let's look at explainability. Explainability is not just one level of understanding.

Let's take medicine. Well, you go to the doctor, you get a pill and the doctor will tell you, "Oh, it's going to help your immune system." Maybe that's good enough for the patient, but it's actually not good enough if you're trying to do research on medicines and try to understand how it works. And that's another level of explanation. But most people don't have the background to understand that. You have to be an expert.

But even that, okay, there's a lot going on in the brain that even experts don't understand. It's just really complicated, all of the different molecules interacting and signals going back and forth. We have a very crude understanding of how nature put together all the different organs in the body and how they interact with each other.

We're discovering that the gut actually has connectivity and sends signals back and forth with the brain. The gut, the microbiome, right? This is something we didn't know until recently. And we're learning things every day.

Now, but you can go even down deeper. You can go down and understand the actual details of the molecules. And this is a physics explanation, right? This is something that is at the level where you have equations. And that level of explanation, all these explanations have different ways of giving us insights.

And you know, the one that we're looking for is words? No, I think we really want to get an understanding of what's happening at the deepest level and that's beginning to happen because we have a lot of very good mathematicians and people, engineers with training and machine learning are digging down into these networks and figuring out the actual way that they are, how the information flows through the network and how it's able to do the amazing things that it does.

And this is really where we're going to get the real understanding, not just in terms of some explanation that you could, a glib explanation that you might use just to make you think that you've understood something.

Michael Krigsman: As we rely more and more on AI to make decisions, including medical decisions, you just alluded to that. This issue of explainability will become more and more important. So the real question I have is, can we ever really understand how the machine is making these decisions? And then what do we do about the fact that we're asking machines to ultimately decide important aspects of our lives and we don't know fully what's going on.

Terrence Sejnowski: You shouldn't think of AI as being a human, it's not. It has special capabilities. It can take in much more knowledge than any single human. So it's really in some way already super intelligent, although people argue that it's not intelligent, but nonetheless, it's helpful. So AI is a tool. It's going to help you solve problems.

And you talk about medical decisions that have to be made. So a study was done, was published in Nature. Nature is the gold standard for science, all sciences and engineering. And what they did was a study on lesions, skin lesions, some of which are cancerous and some of which are benign. And it's like 2,000.

And what they did was they compared experts, a panel of experts with the best AI solution. AI was given thousands and thousands of examples, images of lesions, and it was taught to classify them. Now it turned out that both groups had a performance of about 90%. They were correct in telling you what kind of lesion or whether or not it was cancerous at about the same level.

However, if they let the doctor use the AI in making the diagnosis, the performance went up to 98%. How could that be? I mean, they started with 90. Well, the reason is that they have different knowledge. The AI had access to much more data than any individual could ever see in their lifetime in terms of different, rare skin conditions. And the doctor had much deeper knowledge about the ones that are the most common.

So it's the doctor who's calling the shots, but the doctor takes advantage of this other partner in order to make better diagnosis. And so this is what's playing out now in every area where it's being used by writers, by people who have to work on ad copy, people who have engineers trying to solve problems.

By the way, by far the most impact that AI's had over the last 10 years has been in science. It has transformed science because science is filled with big data on specific projects that are really very complex, like big particle accelerators, huge amounts of data that's being generated. And it's really transformed the way scientists do science. And so this is really a harbinger of what to expect in other areas.

Michael Krigsman: Please subscribe to our newsletter. Go to cxotalk.com, subscribe to our newsletter. And we do have some questions coming in. So let's jump to LinkedIn. And we have a question from Greg Walters who said, he read somewhere that AGI, LLMs, AI, digital twins, et cetera, visualize actions and processes like athletes and more, a million hours of study in one second. Is this correct? And how does this visualization impact model collapse, and does it?

Terrence Sejnowski: This gets to the heart of how you actually create one of these large language models. You start with a huge amount of data, trillions of words, and you go through this process of predicting the next word and gradually getting to the point where you can with very high accuracy.

And the only way you could have gotten to that point is by being able to disambiguate words, have multiple meanings, and you have to look at the context. And so it has to create an internal representation of the semantics, the meaning of the sentence in order to be able to understand just the way you do it. You can predict the next word, but you have to understand what the sentence means before you can do that.

So that takes a huge amount of time and is very expensive. The latest GPT-4, for example, took two months with 20,000 GPUs or graphics processing units and cost $100 million. But the point is that not only now can the large language model ChatGPT answer new questions, but it can do that like in, you press the button and you get the answer in one or two seconds.

And that's shocking. When I first did it, I just couldn't believe this. How could this be? And the reason is it's internalized all this information so that it immediately is able to respond, literally without actually having to think. In fact, these large language models, they don't think. They don't think the way humans think.

And that's one of the things we're going to try to improve. The next generation is going to actually have internal, self-generated activity. Right now, if you stop your dialogue with ChatGPT, it goes blank. It doesn't keep thinking about things the way that you do. If you're on your own without any sensory input, you think about planning, you think about what happened in the past. ChatGPT doesn't do that. It doesn't have an internal life.

Michael Krigsman: This is from Arslan Khan on Twitter, and he says, artificial intelligence is a mirror of human biases. How can AI and our brain become less biased quickly? And I think this leads us directly into a discussion of the impact of AI on our workforce and work displacements and all those implications.

Terrence Sejnowski: The bias could come in because the data are incomplete or skewed. And this came out, for example, when face recognition was being used and was much more accurate with white faces than black faces. And it turns out that there were many fewer black faces that were used in the training set. So you go back to the training set, you have to curate it.

And now we're discovering that the higher the quality of the data, the better the performance in general. But there's other things you can do. And this is a little bit more tricky. Right now, you have to give a goal because deep learning networks don't have intrinsic goals the way that humans have goals.

So you have to give it the goal and the goal typically is I want to optimize performance on a particular class of problems. And so you give it a lot of examples and it chugs away and it gets better and better, and the larger and larger the network, the better and better it gets.

But the point though is that it may not be aligned with a lot of other values that you have, like fairness. And so what do you do? Well, you're going to have to add fairness as one of the goals that it has. And now you have a problem. I want maximum performance, but I also want maximum fairness. I can't have both, right? So I have to weight them. I have to decide, well, half and half, or maybe 90% performance and 10% fairness.

That's a decision has to be made explicitly and somehow we're making it implicitly when we deal with the world, given the cultural biases that we have. Now, here's something about bias though, and I just want to put this to you. I'm going to ask a question of you and your audience. Okay. You know, humans are definitely biased, right? We see this all the time. And LLMs reflect that to some extent because they're trained on human data. Here's the question. Which do you think is going to be easier to fix? The LLM or the human?

Michael Krigsman: I would assume the human is easier to fix because with the human you have control, in quotes, over one, whereas to fix bias in the LLM, you need to adjust a very large data set.

Terrence Sejnowski: Actually, I'd be curious to know what your audience thinks.

Michael Krigsman: Folks, you heard the question.

Terrence Sejnowski: In my experience, I've known a lot of different people and very rare that they change their minds about they're really fixed in their ways, you know, as they get older. Maybe younger children, okay, there's some hope there of changing their biases.

But you come into the world, you don't know what the language is, you don't know what the values are of the culture, and you learn through reinforcement learning, by the way, you know, what's things that you should, how to deal with relationships, how to value people's advice, things that are different in different cultures, how you deal with truth and so forth. It's different. It's not the same. And you learn.

Michael Krigsman: And Greg Walters on LinkedIn agrees with you. He comes back and he said, "It's the LLM that will be easier to."

Terrence Sejnowski: That's because we engineered it, we can fix it.

Michael Krigsman: But how? How do you fix that bias given there are so many levels of insidious?

Terrence Sejnowski: Okay, okay, okay. So in my book, ChatGPT and the Future of AI, I actually laid out because there's a lot in the book that is really inspired by the brain. And by the way, there's a revolution going on in neuroscience that is equally important as what's happening in AI. And this has to do with the Brain Initiative, Obama's Brain Initiative. It's been 10 years on, and we have made enormous strides in understanding the mechanisms underlying your brain.

And so we're going to actually be able to really get, come up again with a much better explanation for the decisions that we make. But here is, when we're dealing with, as I told you, with humans, you use reinforcement learning.

So here's what has to happen in AI. Instead of just training the network at the beginning and then that's it, what you're going to have to do is use reinforcement learning all along. We need lifelong learning where we have reinforcement learning going hand in hand with the declarative. It's called the declarative knowledge system.

And so we have to basically treat the LLM like a child, right? And punish the child when it does say something wrong, does something wrong. And how else is it going to learn that? We have to align it with our values. And that's beginning to happen.

And I know a couple of companies have already used it not for this particular purpose, but for example, something called chain of thought. When you want to solve a problem, you break into pieces and you solve each one separately. Mathematicians do this all the time when they're proving a theorem.

But LLMs, if you ask them, they just jump to the answer and sometimes they're right, sometimes they're wrong. But now you can actually train using reinforcement learning this habit of breaking into pieces and solving each one separately. And now it's gone, the performance has gone on these mathematical problems from 20% to 80%.

So this is an example of using reinforcement learning to actually impart a way of solving problems, and actually a method that could help not just solve mathematical problems, but any problem that requires in manufacturing, in figuring out how to reconfigure your office, all of that should, we should be able to help. We should be able to explain through a sequence of actions how you're going to get to some goal.

Michael Krigsman: And in answer to your question to the audience of which will be easier to remove or address bias in the LLM or in people, Isaac Sacolick comes back and he says, "It's easier to fix AI than humans. People don't change belief systems easily and data and weights can change in an AI. The real issue is understanding what values the data scientists embedded in their training."

So he's in agreement with you. And Arslan Khan is also in agreement. Everybody's against what I said. Arslan Khan says, "LLMs might be easier to change since the input is from multiple people, whereas people rarely change," exactly your point. But now let's talk about the implications of all of this on work, the workforce, jobs. Thoughts on that?

Terrence Sejnowski: We are at the stage that the Wright brothers were at the beginning of aviation 100 years ago. The very first flight by the Wright brothers was 10 feet up and 100 feet forward. That was it. They were inspired by how birds glide, by the airfoil. They were inspired by the lightness of the wings and that was why they used canvas.

But the most difficult part of making aviation safe was figuring out how to regulate and how to control the direction you're going and regulate it so that you don't crash. That was the difficult problem. Well, we're living through that right now, except that maybe it won't take 100 years, it'll only take 10, 20 years.

So we're just at the very beginning of that process. And what's going to happen over the next 10 years near term is incremental advances. And that's already happening. You see all these language models, there's now dozens of them out there with different capabilities and focused on different problems that have to be solved.

So we're just exploring that. And one of the directions that we're going is actually not bigger and bigger because we've run out of data, but smaller and smaller with quality data, with data that's more specific to your particular company, to your particular profession.

And that now is going to shift the balance away from these big high-tech companies toward the enterprise companies, the ones that are actually doing by far the most important job in society, which is making things and getting things to work and dealing with the complexity of the world.

And that could be done. This is something that is not going to take just five years, it's not going to take 10 years because it takes, you're going to have to re complete reorganization of the way that your office and your company is organized.

So right now, for example, I'll just give one example, you know, you have databases all over the place, especially if you're an international company. And how do you integrate across databases? Well, I mean, you have an IT person who has ways of doing that, but to answer a question that requires a lot of comparison and accretion of different sources. That's going to take a long time.

But what are the things that LLMs are really good at because it's already done that for the database of the world. Why not for your company? All of those databases could be integrated. And now when you want to answer a question, you don't go to the IT guy, you ask your chat GPT. Well, you'd call it something else, right? It would be called chat IBM.

And that way, you're going to be able to have access to instantly to all kinds of important questions that you can answer that may help you make decisions. This is if you're head of a company.

What about if you're in the middle of a company? Now, one of the things you keep in mind is that these tools require training. ChatGPT out of the box, you can ask it questions, but the answers you get back are often not the most useful. That's because like any tool, you have to learn how to use the tool.

And there's a whole group of people now that call themselves prompt engineers. What does that mean? It means that they know how to use the prompt in order to be able to more quickly get useful answers out. And in my book, I go into this in quite some detail that may help them.

And the other thing is, it's really interesting, if you really want to incentivize people, what you have to do, dealing with ChatGPT is fun. People do it on their own. People are just doing it because it's a lot of fun, right? Well, once you get trained on doing this, this is going to help with productivity. It's already doing that in many areas, like I gave the example of medicine, but it's going to help people in ways that nobody ever even imagined.

I mean, this is something that like, what was the killer app for personal computers? It wasn't simply typing and letters. It was Lotus 1-2-3. Do you remember that?

Michael Krigsman: Okay.

Terrence Sejnowski: It was a spreadsheet. And that suddenly, that was something very useful that could be used by not just individuals, but by companies and organizing all the data and so forth. And that has become central to the way that now we all deal with large data sets. We have them in these spreadsheets now that are much better designed than Lotus 1-2-3.

But the same thing is going on right now is that we're beginning to develop the tools that are specific and surprising ways. Let me give you one example of something that surprised me. It's in my book.

So a lot of people have mental problems, they have anxiety, they have phobias and so forth and worse, mental disorders. But what do they do? They go to a psychiatrist or a cognitive therapist and they have to make their appointment weeks ahead of time and then they go and then they have an hour and then they get a big bill, right?

Now, this is, it helps. It really turns out cognitive therapy is as good as taking a pill. It really is. And it's complimentary, by the way. Now, they did a test. This is a scientific study. If people are given a choice between an AI psychiatrist or a human psychiatrist, what would they choose? What do you think, Michael? What do you think?

Michael Krigsman: I will say that they choose the AI over the psychiatrist.

Terrence Sejnowski: You figured this out from my first question. Yeah, you're absolutely right. By the majority of humans, they'll feel more comfortable talking about personal things that it might be a little embarrassing, right? If they're talking to another human, who are very judgmental, right?

But the AI, it's not going to judge you. It doesn't have any internal care. It's going to give you respond in a way that is objective and so forth. So this already happened back at MIT in the early days of AI when someone put a very simple program that did very simple things, just repeated the question.

Michael Krigsman: Eliza.

Terrence Sejnowski: Eliza. And Weizenbaum discovered that his secretary is using it during lunch breaks. It's like, and she didn't care that it was a machine and a very simple-minded machine. She did it because it kind of helped, it got her to think about things explicitly.

So there you go. I mean, that's unexpected, right? There's going to be a lot of unintended consequences. And by the way, not all good. There's going to be bad ones too. And so that's why we have to really be careful about what to expect.

Michael Krigsman: You have just been describing what I would call incremental changes of efficiency, right? We now have a critical thinking or an information partner to help us improve our writing, to research, and so forth. But what about the more, the deeper implications and more foundational changes that this may drive in society and in the workplace? Any thoughts on that?

Terrence Sejnowski: In the short run, you're not going to lose your job. Your job is going to change, and it's going to change because you have these new tools. You're going to have to learn new skills.

And that's why it's really important that we have within companies, and it turns out that we have ways of doing that that can help companies. This is called massive open online courses, and they're free, and there's a lot of them out there now. There's like 10,000.

And I have one myself called Learning How to Learn. And this is with Barbara Oakley, and what we realized was a lot of people, especially 25 to 35 half from a college educated are in the workforce and now they need new skills and it's harder to learn when you're in that position because you can't go back to school and you have mortgage and children.

So you want to be able to be more efficient. We teach you how to be more efficient through what we know about the brain, how the brain learns. In any case, there are all kinds of books out there about how to use the new tools.

And the idea, I think, is for people now who have a job are going to be using the tools more efficiently, but gradually the job is going to morph. It's going to be, the boundaries between jobs may change in terms of what people can do, integrating more often rather than having stacked layers and so forth.

You know, it's impossible to predict the future. Let me ask you, okay, Michael, you were around in the '90s, right?

Michael Krigsman: Yes.

Terrence Sejnowski: Okay. You remember when the internet went public, right?

Michael Krigsman: Yeah.

Terrence Sejnowski: Could you have predicted what influence the internet would have on every aspect of your life today?

Michael Krigsman: Absolutely not. Not at that time.

Terrence Sejnowski: Okay, well we have another technology that's going to have an equally important impact up the road that we can't predict. It's just, it has so many different effects, so many different parts of society.

Michael Krigsman: But Terry, I did not invent or discover important aspects, parts of the internet, but you did when it comes to deep learning. And so therefore, I look to you as the prophet and the seer of what may come.

Terrence Sejnowski: The reality is that I'm actually probably no more capable than you are. In fact, maybe less capable of predicting the future. The internet pioneers back then, let me give you an example, okay? You know, Google, when it started, had a motto, "Do no evil." Do you remember that?

Michael Krigsman: Yeah, absolutely.

Terrence Sejnowski: And you know that they were very idealistic. They said this is going to democratize information. Everybody is going to have access to information and their voices will be heard, okay?

Well, could they have predicted how that will play out? And here we are now. We're dealing with misinformation, fake news, echo chambers. I mean, this is like, who could have predicted that? I mean, this is something that even the experts didn't predict.

Michael Krigsman: The enthusiasm of young people before they became billionaires.

Terrence Sejnowski: That's right. That's right. So I'm a young, I was once a young person. In any case, yeah, we are, I'm actually an optimist. I actually think that we will get it sorted out, but I know it's not going to be easy and it's going to take a long time. And that's true of all technologies. It's not like this is different.

It's just that it's so new that we really are just beginning to try to understand the capabilities. It's a moving target. It's incrementally improving. There may be a breakthrough and it may not be, it may be 10 years, it may be 20 years.

And I think it's going to be, the breakthrough I think is going to be when you put something which called is called system two into these large language models. System one is what we have right now, which is basically just trained to produce, generate, generative AI, one word after the next.

But this system two is the one I alluded to earlier, which is self-generated activity. And we actually know how that's done in the human brain. And so it's just going to be a matter of time before that's transplanted into AI. But it's not going to take place in five or 10 years. We're talking here for probably decades out.

Michael Krigsman: Chris Peterson on Twitter says, "The current hype-based AI takes vastly more resource inputs per question than a human brain takes for 24 hours of doing everything." Can you talk about the sustainability, the power requirements, the resource requirements? You alluded to this earlier in your discussion of the trade-off between completeness versus rooting out bias and the input costs and time associated with that.

Terrence Sejnowski: We have to set priorities. You can't, things have been, the cost of things have gone through the roof and that cannot continue. And I also alluded to the fact that the next generation are going to be smaller language models with higher quality data. That's one direction.

The other direction is hardware. And I have a whole chapter on this in my book. If you're interested, go there. So the human brain is able to function at very high levels with 20 watts of power. Some, a little brighter than others, right? But that's much less than the gigawatts that are out there being used right now to answer your questions.

So there's clearly a huge, huge imbalance. And that's because the technology that nature uses is based, is gone down to the molecular level. It's really, really advanced. But we're actually, there's hundreds of companies out there that are building more efficient technology that is going to make lower power, much more, the architecture that you need for these deep learning networks is completely different from the von Neumann architecture that is in your PC.

That architecture does one instruction at a time, and it has to fetch information from memory, which is separate. Well, in your brain, the memory and the processors are one and the same. They're integrated, and there are a lot more processors, right? This is all the neurons. You have 100 billion neurons in your brain.

And now we're up to literally 100 trillion synapses in your brain, and we're up to a trillion weights in these networks. So we're still like only 1% of the brain, even just in raw power. But here's the beauty, is that we can special purpose chips that can be much more efficient.

And that will take us down by a couple of orders of magnitude over the next 10 years. It's already going on, and this is happening. GPUs are a good example. That's two orders of magnitude when they put GPT, when they transplanted these original deep learning networks onto GPUs, it was two orders of magnitude more efficient.

And the reason is that they have thousands of processing units, just like all the neurons, working in parallel, exchanging information. So that was a big step. That was a huge jump, and there's an inflection. Moore's Law went from doubling every two years to doubling every two months, literally. I mean, this was a huge, huge change.

Now that's going to happen again, and it's going to become probably another two orders of magnitude or three, possibly, with a new class of computing called neuromorphic engineering. What is that? Well, it turns out that if you take your digital chip, which are basically zero ones, right? And run them near zero where things are analog, by the way, digital chips are actually analog when you look at the actual transistors, but they run into the rail.

But if you work down there, and this is Carver Mead at Caltech actually understood this, that you can actually take advantage of that processing at that low level for doing crude multiplies and adds much more efficiently by like two or three orders of magnitude. And now that's a mature technology, and so that's going to come online over the next 10 years.

So see, what's going to happen is just like Moore's Law is that the technology gets more and more efficient. And we're just at the beginning of that because we realize now that this architecture may be more important than the von Neumann architecture for solving many, many kinds of problems that before we couldn't.

Michael Krigsman: This is from Isaac Sacolick, and Isaac is a big-time CIO influencer. A lot of CIOs listen to him. And he says, he liked your healthcare example. "It may be hard to benchmark human versus human plus AI versus AI and other diagnostic areas. What are options to validate accuracy and build trust with AI recommendations?"

Terrence Sejnowski: That's productivity, and that's the bottom line. Actually, you know, for a long time, computers were, the whole business world and public were investing in digital computers and productivity didn't change very much. It wasn't until they connected them together that they could exchange information. That's when things took off.

Let me go back to, to answer the question in a concrete way. Let's go back to healthcare. So you go to a doctor for some medical problem and you walk into the office, you have 20 minutes, doctor is sitting there next to a computer and is asking you questions about your symptoms, about your previous medical history. What is he doing? He's looking at the computer, he's typing in because he's got to get all that in the computer because he has to have a record, right? And you know, he's not looking at you. And then at the end, you know, he gives you a list of drugs that you should take or some advice and so forth. And off you go, and you maybe you get, you remember half of it if you're lucky, right?

So here's what's happening. Right now, there are 10 companies out there that have realized I could put together a lot of different AI systems to make that process not just a lot more efficient, but actually much more likely to be of help to the patient. So, first of all, speech recognition. The doctor doesn't have to look at the computer. He can talk to the patient. It turns out you get a lot of information by looking at the patient, not just what they're saying. You can tell from their facial expressions how serious something is. You can look at their face, the color of the face. I mean, a good diagnostician uses, can use that information, right? And jump to conclusions that you can't just with a couple of numbers that they're typing in.

And okay, now what happens? So you press a button at the end and much more human interaction, and now ChatGPT does the heavy lifting. It gives you a summary, it's very good at that. The doctor looks at it and checks to see what's, if there's any problem in terms of the recommendations and so forth. And now that goes back and the patient now walks out with actually a very detailed summary. And so now the patient is going to be able to understand much better what needs to be done over the next couple months in order for the patient to be able to improve their health.

And this is just kind of one example of how we have a system right now, which really disadvantages patients, right? And doctors, by the way, doctors don't just stop because they have to write up notes afterwards and often late into the night. And now all of that is being done. They can go and this is something that it's going to take decades to permeate the medical system. You know, they're very conservative, so it's going to take a long time and it has flaws and there'll be problems. But ultimately, I think that it will greatly improve healthcare.

Michael Krigsman: Chris Peterson on Twitter says, "The current hype-based AI takes vastly more resource inputs per question than a human brain takes for 24 hours of doing everything." Can you talk about the enormous cost of retraining the big models? How does this play into GDPR style right to be forgotten laws? Will AI systems need an exception because it's just not practical to retrain for each request?

Terrence Sejnowski: In addition to the initial training, which is the costly part, there's also opportunities later to do something called fine tuning. And what's that mean? Fine tuning is basically giving it extra data and trying to train it in a gentle way such that it now has access to new information without interfering with the existing database. Now it's not perfect. What happens is that the large language model gets dumbed down to some extent because it's taking on this extra burden. But that's not nearly as expensive, and that can be done in house, right? That can be done in individual businesses. You have to hire somebody who can do that for you, these are machine learning people.

And by the way, when I started, we were doing neural networks, right? And that was in the '80s. And then over time, NeurIPS became a machine learning conference. But now it's come back full circle, and now it's not called neural networks anymore, it's called AI, right? You know, it wasn't because we called it that, it was because the world was calling it that. But it's really about machine learning. That's the heart of AI today, and it's all about data. Whoever has more data wins in any area. And I'll tell you, your company is sitting on enormous amounts of data that's really important for you. And if you can get that into an AI, you will win. As for the right to be forgotten, it's a tricky issue. Fine-tuning helps, but doesn't completely erase data. New techniques are needed, and research is ongoing. It's a legal and ethical challenge, not just a technical one.

Michael Krigsman: Can you talk about how the assumptions made during data curation influence AI outcomes?

Terrence Sejnowski: Hugely. This is really going to become a very large business, data. As I said earlier, data is hard to come by. Quality data, even harder. And so how is it that we're going to be able to overcome a lot of the problems? You know, hallucination is a good example. Hallucination actually has advantages. If you're a writer or ad copy, it's really good at that, or poems, but not so good if you're asking for a fact. If it hallucinates, you're in trouble.

So, I think what we're going to need is another layer on top of the language models, which is a fact checker, basically. And it will take what is coming out and will learn how to then look through the database, world's database, and come up with sources for that particular fact. So again, it may be, this is really the exciting thing, is that right now, humans are doing that. You know, all these big companies have thousands and thousands of humans that are going through and sorting through, getting rid of racist data, trying to prevent these large language models from saying, from doing anything that is going to hurt people. You know, that is being done on a individual basis, and that's very labor intensive. But we should be able to train another generation of large language models to actually do that part of the job of sorting through and flagging things.

And even in the earliest days of, for example, neural networks in the '80s when I started, they were using these very, at that time, they were very shallow with one layer of hidden units to take a slide with cancerous, it was called a Pap smear, cervical cancer. And humans would have to look through every cell, and it would take hours and hours and very costly. Well, neural networks could do that much more quickly, reduce it to 100, and then the humans were going to use their much more powerful visual system to look at the debris and all the different types of cells and come up with a much, actually thorough and faster way of analyzing the data. The assumptions made during curation are critical. If your data is biased, your AI will be biased. Garbage in, garbage out. It's about careful selection, cleaning, and labeling, understanding the limitations of your dataset and constantly evaluating for unintended consequences.

Michael Krigsman: So on the subject of humans and data, how can we detect and understand the impact of adversaries manipulating data, for example, with misinformation and disinformation?

Terrence Sejnowski: Well, that's already happening, you know, where social media is filled with that. That's a very, very difficult problem. I don't think there's going to be any easy solution. I outlined just now what I think is going to happen with all these layers of filtering, but I think one thing that we should be open to is the fact that every once in a while, there's a massive breakthrough. The last one occurred in 2022, when ChatGPT was opened up to the world, and suddenly everybody realized, "Oh my God." You know, "What hath God wrought?" That was the first message sent over the telegraph, right?

And we're going to have it. I'm sure we'll have another moment like that. I don't know when, I don't know how, but it's going to happen. It's going to happen, and it's already happened in, I told you, in science. It's happened multiple times. Detecting adversarial manipulation is an arms race. We need better techniques for provenance tracking, anomaly detection, and understanding how information spreads. We also need media literacy and critical thinking skills to be more widespread. It's a societal problem, not just a technical one. Think of it like cybersecurity for data.

Michael Krigsman: What advice do you have for business people, given everything we've discussed?

Terrence Sejnowski: Number one, read my book, not my lips. Read the book because I talk a lot about these issues. I mean, I only touched the surface of what is in the book, okay?

Number two, the most important thing you should be thinking about is training your workforce with this new technology and take it seriously. It's going to, this is just the beginning, right? And unless you start now, right? Helping your workforce use these tools, new tools, and it really is, they're tools that, if you can use them badly and very well, and people are trained in that, and it does take time, it takes maybe 100 hours. That's a lot less than the 10,000 hours it takes to become an expert, right?

And number three, download your company into an LLM. And I'll tell you, this is already being done with the brains of flies and zebrafish. These are brains that have 100,000 neurons. We can download them now into AI. And they have similar behaviors, they have similar activity patterns. You know, that's the future.

Experiment, but be cautious. Understand the limitations of current AI. Focus on augmenting human capabilities, not replacing them. Data strategy is key. And think about the ethical implications of everything you do.

Michael Krigsman: My mind is blown from this conversation. Terry Sejnowski, thank you so much for taking time to be with us.

Terrence Sejnowski: Oh, very pleased. Thank you so much for making this opportunity possible for me.

Michael Krigsman: And a huge thank you to everybody who watched. You guys are amazing. You ask such excellent questions. Before you go, please subscribe to our newsletter. Go to cxotalk.com, subscribe to our newsletter, subscribe to our YouTube channel. And we have amazing shows coming up. Thank you so much, everybody, and I hope you have a great day. Take care.