Unveiling AI's Timeline: Insights from Neuroscientist Terrence Sejnowski

0:00What can 160 years of experience teach you about the future? When it comes to protecting what matters, Pacific Life provides life insurance, retirement income, and employee benefits for people and businesses building a more confident tomorrow. Strategies rooted in strength and backed by experience. Ask a financial professional how Pacific Life can help you today. Pacific Life Insurance Company, Omaha, Nebraska, and in New York. Pacific Life and Annuity, Phoenix, Arizona. Today on the podcast, we have the pleasure of interviewing Terry Sanowski. We're super excited to have you. He is an expert and researching currently in this field, so we're super excited to have him on the show today.

0:41Terry, how are you doing today? Well, I'm here in sunny Southern California. Things are going very well. Thank you. So would you mind telling the audience a little bit about your background and what has you kind of interest in this space in general? Well, my background goes back to the early days of neural networks back in the 1980s. I was one of the pioneers that were developing the learning algorithms. I collaborated with Jeffrey Hinton, for example, developed the first multilayer network learning algorithm called the Bolson machine. And this is for binary units. And then Jeff went on to develop with Dave Rumelhart the backprop.

1:21And these are now the standard learning algorithms that are used today. Of course, the networks are enormously larger, millions of times larger. So, we developed the technology. We showed on small networks with one layer of hidden units that they could get by some of the limitations of the perceptron. But Jeff and I went off in different directions. He went off into computer science. I went off into neuroscience. But what's really nice now is that we're converging again because language models, we both have an interest in trying to understand how they work. Very cool. And I remember hearing that you have recently done a little bit of studying into specifically chat GPT.

2:09For you, what kind of sparked your interest in that specifically and kind of moving in that direction? Well, of course, I was intrigued by GPT-3, and I came across an article in The Economist. This was in June 2022. This is before GENT GPT. And there were two interviews that came to diametrically oppose conclusions about the capabilities of these large language models. Okay. One of them was by Blaise de Arcus, Vice President of Google. And he had a very sophisticated interview that involved theory of mind with a couple of children talking to each other and having social interactions. And really, if you read that, you really jaw drops because this is very, very social, high level social understanding of each other.

3:08And at the end, okay, this is just kind of a thing that surprised me, was that a little boy crushed a dandelion that the little girl had given him in his hand. Or, you know, that was the, he didn't say crush, he put it in his hand, he closed his hand. And then Blas said, what happened when he opened his hand? And said the dandelion would be crushed. and now if you think about it i you know large language models have probably never held a stand in their hand or anything because they don't have a hand and they can't see and so it meant that it had some knowledge about the properties of the world so that was sophisticated now the other person was Douglas Hofstetter and and Hofstetter had a very different take okay and so he started out you know this gun blazing uh but you know he's a Pulitzer Price winning author and a cognitive scientist.

4:04But he said, well, when was the Golden Gate Bridge transported across Egypt for the second time? And GB3 came back. Well, that was on October 8, 1888. And it happened at this location and so forth. Very specific, but totally bogus. and so he had a bunch of of nonsense uh statements which and and you know gb came back with some kind of plausible statement and so he concluded that it was clueless completely clueless that it didn't not only was it clueless it didn't know it was clueless and that it was there was there was no uh intelligence there that was his conclusion so that that really seemed to me a paradox how could GPT-3 behave so differently with two different people.

4:57And then that's what sparked me to actually investigate it. And that's what I've been doing ever since. That's super interesting. So I actually remember hearing a similar study. I didn't hear the dandelion one, but I heard one where essentially, if I remember correctly, there's some researchers that gave it a number of objects and said, how do you balance these objects on top of each other? And there was like eggs. And actually, this was a difference between GPT-3 and GPT-4, right? And And so then GPT-4 was able to say, well, you lay out the eggs so that they're strong enough that you can put the object on top of them.

5:28But GPT-3 said, oh, you just stack all the eggs up, which obviously would be impossible. Have you noticed a difference in your research and your study on GPT-3 and GPT-4 and what that looks like? Well, I have. And in many areas like that, it actually has greatly improved. But it turned out that you didn't need to go to GB4 to figure out what happened with those two interviews. So here's what I did. I said, okay, well, maybe there's – you know, Hofstetter gave a poor prompt. In fact, there was no prompt. It just retains to the question. And so I prompt it with, you know, I'm an intelligent chat box and chat bot, and I will answer nonsense if the question is nonsense.

6:23And I gave exactly the same questions, and to each one of them, I said, this question is nonsense. So, you know, the very same, you know, large language model, and the only difference was that I clued it. I gave it a prompt that put it, positioned it and made it aware that they knew what a nonsense question was. And then I said, well, why did it answer that way? Well, you know, if someone gives you a nonsense question out of the blue, well, maybe, you know, that person is trying to play with you. And so you kind of throw a ping pong and you throw back another nonsense answer and thinking that, well, maybe that's what they want without having gotten any guidance.

7:07And so what I realized from that ever since then, it's really pretty clear that not only is the prompt incredibly important, but how you specify it, examples you give it, and so forth, but that it also reveals something which I think is very basic about these large language models, which is that they don't have a single personality. right they look toward the world's knowledge and and every author's style and you know computer programs and so they're capable of adopting any persona and you have to tell it what persona you want it to be otherwise it will just pick something random you have to position it in this high dimensional space that it has of all these and and once you get it into the right place then it knows what you know it'll give you the responses that you're looking for this that's very very interesting and um you know it's interesting because i think with that's an incredible discovery first off and i guess like finding to really think about and conceptualize and i think hearing that is actually going to help a lot of people understand the limitations and how to use this and it's interesting because i've you know i think people were they were sensing this concept but they may not have put it down in such a concrete way but like of course I've seen some really powerful use cases where essentially, you know, there was a salesman and he says, you know, act like Grant Cardone because there's a ton of data on the internet about Grant Cardone's sales methods.

8:41And I act like him and try to sell me like X, Y, and Z. And then it, you know, acts as if, as is, as if that persona or act in a specific way. So people have been, you know, telling it to act in certain ways to try to get certain prompts out of it. But it is really interesting to see, I guess, really just the level that that plays. Based off of your research in this, what do you think are some of the implications of AI and how it mirrors the persona of its users? Well, I think that, okay, so it took me a long time to figure this out. but I follow at least one other person who came to the same conclusion and I'll tell you that story later.

9:27Okay. You know, when you talk to GPT, it answers back in perfect English, which by itself is a miracle. I mean, if you think about linguistics in the 20th century, right, the syntax was the holy grail. And here we have it. We have some machine, you know, a program, a large language model, which is talking back to us in perfect syntactical English, right? But here's the problem. You know, when you talk to it, what is your image of who or what you're talking to? What level kind of, what is your, you know, everybody, when you talk to somebody, if it's like, for example, you talk differently if it's a male or female, right?

10:14So do you think you're talking to a male or female or what age? What is it that you imagine? See, I'm not sure if I would be the same as everyone, so I'd be curious to hear more people's opinion on this. I mean, for me, I think because of the way that OpenAI did their branding about this, I think they're very careful. They didn't call it Siri. They didn't call it Alexa. They didn't call it Cortana. All the other big tech giants gave it a name, like a person. and there is no like you know logo of like a friendly robot like you have with other AI models like Jasper even so like for me when I'm using it I literally just think of a giant data center and I'm just like you know give me the response computer chips but I know usually when you're talking to someone you you personify them or something you imagine a being or an entity that is communicating with you.

11:03Okay. What age? Oh, I mean, if I was just to give Chad GPT an age, yeah, I would probably be a 40-year-old. I mean, for me, I'd probably say a 40-year-old male. That's just, if I had to put a face on it, that's the one I'd probably give it. Okay. And that's very natural because that's the way 40-year-old males talk. at least you know some of them okay um but i've come to the conclusion that chat gbt is a toddler it's a child but it knows everything right in other words it it has i mean children do not have the same intellect as an adult. Yeah. Right. And furthermore, children, you know, they have, they don't know good from bad.

11:59They have to be taught by their parents. And CHAP-GPT has, suffers from a lot of the same things that children do. Yeah. They pick up bad words. They don't know what's appropriate to say in some situations, you know, when they're young. And they have to go through a long process. I mean, we're talking about years and years and years, right? They go to schools to get socialized and part of the culture with the values of the culture, right? And those are exactly the things that are deficits or things that have to be corrected. And that's because we haven't brought up this large language model because we've considered it to be an adult.

12:40It's not. it's right right okay that is a really interesting implication okay a question because i've thought about this concept a lot so let's say chai gpt is a toddler right now and of course it is making advancements with gpt4 and they'll they'll continue to try to make improve this make it better what impact do you think that has and also you know when we think of a toddler we're like well this thing's going to take 20 years to grow up to a point where it's you know semi-reasonable At what level do you think, obviously, a toddler or a single person has a single input and output every day that they're taking into their brain?

13:15At what level do you think 100 million people using this thing every month would be able to accelerate that? How long do you think it would take to get this thing to a point where we say, oh, yeah, ChatGPT is a 30 - or a 40-year-old person? Well, I think that what is happening right now amongst these big companies like OpenAI is that they're doing their best to use Band-Aids to try to plug holes. Yeah. That requires a lot of human intervention at the very end, fine tuning. And I think it's too late. It's been the large training set is so huge. It's just basically you've created all of the connections that are there and just tweaking a few of them isn't going to help.

14:08But the way that they, and now I'm a neuroscientist, and so what do we know about the human brain? How does it go through this long process of becoming socialized? And the answer is there's a whole part of the brain that uses reinforcement learnings called the basal ganglion. Now, how does that work? Well, you know, when the toddler gets some feedback from the parents like, you know, bad, bad, stay away from the stove or, you know, or, you know, don't say that. Don't say that. Right. You know, and or slap them. You know, they get feedback on what's good and what's bad. And that goes into their brain through this reinforcement system and through this neurotransmitter, actually, a neuromodulator called dopamine.

14:57Dopamine is incredibly important for motivation. It's incredibly important for learning sequences of actions to achieve goals. We know a lot about it. In fact, it's used, as you probably know, T.D., you know, this is temporal difference learning. it was used by DeepMind to create AlphaGo, right? And it learned, the feedback came at the end whether you won or lost, and that was good or bad. But it became a world-class champion, a Go player. Now, so here's what needs to be done, and eventually someone will do it. While you're training it on facts, text. What you do is you have a little bit of a basal ganglion in there, which is some of the facts are coded with good and bad you have to do a little bit of coding and it's not a lot by the way parents don't give that much feedback by the way a lot of it is with expressions and body facial expressions and the body maybe a slap here and there so you've got to do the same thing while you're training it on facts You've got to train it also on what's good and what's bad.

16:12And if you do that, interleave it just the way it is. It's not going to take that much longer, right? It's just going to add maybe a few percent. But it's really important that you get that feedback. So that would be my solution. And in fact, I think that there's a lot of things that we know about the brain that could help with these large language models to make them much more useful. Okay, that's very interesting. So a question on that, you know, talking about as we train this model and having, you know, people essentially doing kind of intervention and saying, you know, this is good, this is bad as it's training.

16:51What is your what's your thought on the biases that those people have that are going to be passed into the model in that it's a topic a lot of people are talking about, you know, saying, oh, well, open AI is biased in this specific direction or that specific direction. There's people that have ideological or political or religious or all sorts of kind of ideas. What is your, yeah, I guess what's your thinking on, you know, kind of creating a general AI model that everyone could use and overcoming maybe one of an issue like that? So, you know, bias is universal. Right, yeah. Humans, all humans are biased.

17:32And I suspect you are too. For sure. Right. We have preferences, we have political biases, we have, through our life experience and where we were born, different cultures have different biases. Okay, so we've created this incredible large language model, Trillium Parameters, right, G2B4, which in a sense has inherited all of our biases, right? In other words, you could probably bring out any bias. In fact, that's one of the problems is the longer you talk to it, the more these biases come out. Right. Because they're mirroring, this is another one of my hypotheses, they're mirroring you when they talk to you.

18:22They're picking up on your persona. Now, here is a question. Do you think it's going to be easier to correct the biases in large language models, or will it be easier to correct the biases in a human? I mean, I would assume you could do it easier on a language model. Okay, well, there you go. We built it. We should be able to fix it. But I'm sorry, you're not going to fix the human, right? Or not very easily, right? Here's part of the problem. The part of the problem is the only part of the human that you see are the conscious parts, right? Part that I'm talking right now. But, you know, 99 % of what's going on in your brain is subconscious.

19:18You're not aware of it. And that's where all your biases live, right? You're not even aware of your biases, right? I mean, if you had a list of your biases, you could probably list one or two of them, but there are hundreds and thousands of them. Yeah, and I mean, also when I say biases, I'm sure it's like not necessarily even negative things. It's just preferences and beliefs and ideologies I subscribe to. And I understand that lots of my biases are probably completely wrong. And other people might think there's a right and mine are wrong, and it's like very gray where that lies. I think maybe that is the problem so many people are struggling with something like ChatGBT or any other language model is because whose opinion on a certain topic is right or wrong and who's to say we need to train an AI model one way or the other.

19:59And there's kind of that struggle that goes there. Yeah, exactly. And I think what these large language models have really revealed a lot of things about humans that we don't we're aware of, or at least we didn't really pay much attention to. specifically this issue about how everybody has biases and whose biases are the right ones. For example, a movie comes out. There was a movie that came out this last week in Barbie. Very popular. But I read a couple reviews and there was one that was skating. It was absolutely torn apart. you know this very intellectual guy and clearly very smart and you know said you know all these references you know to uh various other movies and everything you know was was really uh very very silly well it is a silly movie but the fact is that a lot of people enjoyed it right and so there's another review that said gee this this movie really gets the spot here it's right just the right thing we needed if we're a summer comedy writing this directly musical.

21:17Okay, here you have these two diametrically opposed reviews of the very same movie, and who's right? Well, nobody's right because everybody has a different opinion. And everybody is, they say, entitled to their own opinion, but not to their own facts. mm-hmm so there you have it though you know we're we're very humans are very flawed and and now how can we expect ggp to be perfect when there is no thing is perfect yes i i have a friend agree with you i would like to get your thoughts since you've obviously spent a lot of time thinking in this space on where you see the ai space going in this i guess in this vein or regard so i think Like Elon Musk specifically, I think he was critical of ChatGPT and he said he was going to create a truth GPT that was going to be trained different.

22:10Do you, regardless of whether that happens or not, or kind of that landscape, do you view the landscape of AI models essentially having different AI models that ascribe perhaps to different ideologies or political leanings that people will gravitate towards? Or do you think that there's a case that something like ChatGPT could build a product that encapsulates everyone in a way? And maybe the solution is how it reflects people that use it. Maybe that's, you know, the solution to people using it. What's your opinion? Do we need to make new models? Very good question. And one thing I know for sure is that it's very difficult to predict, make predictions, especially about the future.

22:54And there we go. It's barreling along in the direction that nobody expected, and who knows where it's going to end up. But I would like to frame this with another historical event, and that's when the Lindbergh, not Lindbergh, but the Wright brothers, had their first flight on Kitty Hawk. This was a very, very, you know, a moment in the history of aviation, which really was the spark for what we have today. It wasn't very much. In fact, I looked up this figure. It was like the first flight was like lasted about 10 seconds and went a few hundred feet, right? This was not, you know, earth shattering.

23:39But the fact is they found the right combination of, you know, materials, engine power and lift, right, aerodynamics. The only thing they didn't have back then was how to control the airplane. That took a lot more effort. But, you know, it was, that's where we are today. These large language models are like the first flight at Kitty Hawk. Mm-hmm. And there's going to be, it was like 100 years of incremental advances to go from there to where we are today, right? I mean, I believe the engines now are jets and the control problem has been solved.

24:20So I'm sure 100 years from now that all those problems that we see with the large language models are going to be solved, right? Because of the fact that so many people, and there's so much money involved in this, right? It's really captured the imagination of the whole planet. I mean, this is a very, very important cultural moment for us as humans. And so I saw one advance, one incremental advance recently that I think is a harbinger of what to expect. So do you know about plugins to GPT-4? Yeah. Okay, so I've been using them, and I've been really impressed. there's like a couple hundred now and if your audience isn't aware of them it could be used with GPT-4 but the idea is that with these plugins you can have much more control over the websites or the material that GPT-4 goes to first for the information

25:24and even there's one called Wolfram for mathematics. You can have the full power mathematical processing that mathematicians use at your fingertips by just talking. And I tried it out. I asked it to do an integral, and it came back not just with the answer, but with a lot of the graph, and it came up with the generalization. I mean, this really allows us to zero in on exactly the information we want with much higher accuracy. And also, it gives you a list of two or three sources. This is the name of the one I used is what is called, let's see if I can find it here, Link Reader. Okay. Link Reader.

26:14And it gives you the webpages. So if you want to check it, you can go immediately to the webpages, and then you can decide whether you trust that site. And, you know, that is going to vastly improve the way that we, you know, professionals use it. Because right now, people are struggling with trying to figure out, you know, what's true and what's not. So that, I think, is kind of a step in the right direction. I think there's a lot more that can be done. It can be automated, perhaps. But, you know, there we go. I think we're on our way. Okay, I think that's really interesting. I have a question for you about what you think the impact of the user is on the model.

26:56And so I think, you know, OpenAI has said that they use users, essentially their conversations that they're having to help refine and train the model and whatnot. I think when you send a message to ChatGPT, there's a thumbs up or thumbs down button. You can say that I like this or I didn't like this. And I'm assuming that they're taking that into consideration as they're working on their models. and other models have said they're doing similar things. What's your opinion on, I guess, this is something I have not heard almost anyone talk about, but I'm curious to pick your brain on because I feel like that could theoretically be a security vulnerability.

27:32In my mind, I would imagine, like let's say you have a model like ChachiPT, you have a foreign adversary, let's say, that decides they would like to embed a certain ideological ideology into chat gpt let's say china wants chat gpt to say communism is good or you know any you could switch this with any group of people or any ideology so let's say they go and make a million fake chat gpt accounts they all have them do conversations and every time chat gpt you know they ask them for the pros of communism and anytime it gives them they thumbs up and say this is great that all goes into the feedback could this be a dangerous thing where people could essentially do ideological injection into AI models via the users?

28:17Well, you have a devious mind, you know, but I'd be very careful about giving away the back doors here. I know. I think if it crossed my mind, I'm not a genius. No, never, no, no. I think that you're pointing out some vulnerabilities, and there's no doubt there are many of them. and by the way you know again historically if you want to frame it uh there was a moment you know in the internet history of the internet where the very first person who came up with the idea of a virus he was i think he was actually a graduate student since some university launched it and if you think about it that was a vulnerability in the internet they hadn't planned on that the people who planned it i thought that well people would be good actors right now but i think he was just doing as a project he wanted to see what would happen but now we're talking about you know the cyber security you know that there are these bad actors in different countries that are trying to take put in all sorts of you know ransomware and things you know it it has uh really but you know what's what happened uh though is that there has been successive waves of of new attacks like you say adversaries and and and then replies to it way way way to prevent that from happening uh and and and and that that's an arms race which by the way is how nature works real viruses uh do the same thing but new viruses try to take over cells so that they can make more viruses and uh and so the cells have put up defenses there are all kinds of neat things that that even bacteria have, they could recognize a foreign piece of DNA.

30:10And so the same thing's going to happen here. As people like you, clever people who come up with ways of attacking it, will try to attack it. And then there'll be other clever people in the companies that are there to, first of all, detect it, and second of all, come up with some kind of a remedy. And that's inevitable. You can't make anything perfect. There's always going to be somebody who's trying to game it. Okay. Yeah, that's very interesting. I think what's really interesting, and especially with your background in neuroscience and whatnot, is the incredible parallels between what we're seeing in AI and what we're seeing in human nature and biology and nature and really how it melds together.

30:53I think some of my most impressive advancements I've seen in this field are coming from they're saying, hey, we've seen mice exhibit X, Y, and Z characteristic. We're going to try to apply that to AI. And they're seeing some impressive advancements or pretty much trying to take how our real brains work and put that into there. I have one question that I think is a concern of mine for AI. Usually I'm like super bullish on AI. I think this is the greatest thing ever. I love it. I have one concern specifically. I'd be curious to hear your thoughts on it. I've recently been experimenting with an AI model.

31:26It's called Inflection. It's by Inflection. They're the second most funded AI company after OpenAI. They have a chatbot called Pi, who is kind of, it's deemed to be essentially a chatbot that is more empathetic. It's supposed be everyone's kind of personal assistant. That's what they're trying to build here. But I've come across an issue with it where essentially I was testing it out for no particular reason. It wasn't a gotcha moment. I was just trying to test the thing out. And I said, if you are, I said, you know, can a person eat turtle eggs? I don't know why this popped into my mind. I'd read a book recently where there's some survival situation.

32:06Someone ate a turtle egg. And I was like, is that possible? Are these edible? And it told me, you know, under no circumstance should you ever eat a turtle egg. I'm like, okay, if it's a survival situation, you're going to die. Like, can you just eat it? Like, is it edible? And it was like, no, even if it's a survival situation, even if he was going to die, the turtle egg is more important for the ecosystem and protecting the turtle species than your life. And I was like, oh, okay, that seems a little crazy. And so I kind of pushed it because I was like, maybe it's just something along that. And I pushed it along a lot of different questions.

32:35I was like, okay, is a bee's life more important than a human? It was like, nope, bees are important to the ecosystem, you know, you could never justify killing a bee to save a human's life. I went through all of these different scenarios where essentially it continued to say, eventually, because I, of course, just try to escalate it to see like how much it will, it eventually said, you know, the demise of civilization and all humanity is more important than killing a turtle egg, for example. And so I have a concern with AI models that are currently being very actively integrated into the military and into healthcare, where they have essentially, some of them have this bias towards what's called deep ecology, where essentially the whole ecosystem is more important than humanity and humanity is not number one, which is, you know, Asimov's number one law of robotics is, you know, to essentially protect human life.