 Despite the sun here you are, this is wonderful. Welcome to a new term of Osher Lifelong Learning Institute in central Vermont. We think we've got a good program and nothing better than Tamas than to kick us off today. Next week we have another interesting thing that I can get Tony Caving talking about the Death and Dying Act in Vermont. So be sure to tune in for that. Let me get this over so I don't have to be stationary. I've known Tom and his wife Mary Ebslin for almost 50 years. In fact it was over 40 years ago that they gave my then 16-year-old son his first programming job at their company Solutions which they ran and they founded and operated right in Montpelier. And Solutions in case you didn't know was a company that created some of the earliest software for the then brand new Apple PCs which came, unlike today, with practically no applications on it. There was very little you could do. My son programmed his own games to play on this Apple PC. In the same way Solutions produced some of the first ETF electronic funds transfer software which we all did for granted. Later on Solutions was sold to Microsoft in town, worked for a time with a parent company introducing a lot of things that we take for granted today. We also had a spell with AT&T and was responsible for launching their first internet service provider which popularized the flat rate for internet service that most of us enjoy today. Now while his wife calls him her nerd, I think of Tom really as a renaissance man. He's a serial entrepreneur after Solutions going on to found ITXHC which was a listed company, one of the most successful and earliest voiceover internet companies. And more recently he started a company here in Vermont called NG Advantage which provides gas, liquefied gas to big commercial customers. He's done also his share of public service serving as Vermont's transportation secretary in the 80s. He was stimulus desired during our big recession about 15 years ago and he has consulted with several governors on technical issues. Add to all this, and you'll see why I call him a renaissance man, a novel and short story both available on Amazon you can check it out. And his current blog entitled Fractals of Change in which he discusses many things. One wonders when he had time to do, with all that, how he had time to become also listed as an inventor on eight US patents and he also pilots planes. So if I made any attempt to give you more detail about the many things Tom is going to do, we'd have no time for him to speak to us. So he's been playing with AI systems since early on and he closely follows this emerging technology which is both intriguing and scary to me. So let's see what he thinks about its enormous challenges and opportunities. Thank you very much Edie, and to Edie of course is the renaissance woman, I really am a nerd and the reason my wife calls me that is it's much more useful than somebody who can't plumb and doesn't do a very good job at carpentry. But I can get her computer and her phone working and thank God that's necessary these days so I have a role. Thank you all for coming. This presentation is from a mature audience as well over 65 aren't we? AI seems to have come along just in time and it's like self-driving, I'm getting to the age when I took the keys away from my parents and I'm hoping that soon I'll have a car smart enough so maybe I won't be allowed to drive it myself but at least I can go places and artificial intelligence may supplement some of the memory that's slipping away. But let's talk about artificial intelligence. Like any new technology it can be scary. I don't want to break the suspense but my hope is that it'll be much more of a benefit than it is a drawback. It has good uses and has bad uses like any technology there's no question about it and people will use it well and people will use it badly and there's no question about that either. But if we understand it then we can encourage the good uses and we can avoid if not discourage the bad uses. So let's just switch gears a little bit and go over to chat GPT. How many of you have used chat GPT or used it with somebody else? Okay great. For the rest of you this will be your first introduction. Chat GPT was introduced in November of 2022. You may remember it caused a great deal of stir for a while that was the most downloaded new application on the internet and ever since there's been a lot of discussion about it. One of the reasons or perhaps the reason why artificial intelligence has been in development for a long, long time, 50 years actually or some people say it's what Oz gave to the straw man so maybe even longer than that but when chat GPT came out artificial intelligence suddenly broke into the public consciousness and part of the reason for that is how it appears to be able to engage in a conversation in a human language. In fact it can converse in quite a few languages although it's best in English and in fact it can actually converse that is you can talk to it and it'll talk back to you but I'm not going to demonstrate that today. Just going to show how it works when you're typing to it. There was a famous computer scientist named Alan Turing who invented something called the Turing test. People had asked him Dr. Turing how will we know when there's machine intelligence and he said well let's ask chat GPT what the Turing test is. To be a little less wordy than chat GPT said if you're talking to somebody remotely and you can't tell whether that somebody you're talking to remotely is a computer or a person after some conversation then that's past the Turing test. People have artificial intelligence when there's something we can talk to and we can't tell whether we're talking to a human or we're talking to a machine. Now I could ask the same question of Google which we've gotten used to and I'll probably get an answer to. Not a bad answer from Google either. So why all the excitement about chat GPT. And by the way I should point out the reason the Google answer is as good as it is this year is because Google's actually put artificial intelligence behind the search engine as well. So although it's not chat GPT here it's their own search engine called Bard that made the Google answer are better than it would have been a year ago. So going back to chat GPT remember it's not just asking one question and getting an answer it's being able to have a conversation. So now I didn't say who was Alan Turing. I said who was he and if I were having a conversation with a human being the chat GPT was able to figure out by he I meant the only person that was referred to just as a human would have figured out in the sentence in the exchange before so I must be talking about Alan Turing and it gave me the answer understanding that it was Alan that I was asking about. If I go back to Google and I say going to give me a search based just on who was he we're not having a conversation it's just answering each one of my inquiries. So you can see that that chat GPT gets exciting because it's such that's a long answer because it does so well at engaging in conversation and we could go on on something somebody'd like to add questions somebody'd like to add to this conversation. Okay. Now. Oh yes. Away from that. I think and then you get the sources that which is what I like. Yes. You do. And okay. That's a very good point. I didn't spell sources right but we'll see if it figures it out actually you can do better than that. Sometimes it can be stubborn. See if I can convince it to give us some sources now it's doing better. Just as Google has AI built in behind it AI has Microsoft's browser Bing built in behind it and so this time it use Bing Oh Turing. Yes. Very much. I remember when I was in high school okay this time it gave us some sources that we could use to check it but I had to ask but even that's sort of interesting in its conversation so working with chat GPT is a little bit like working with a human being you don't always get what you ask for exactly and sometimes you have to drill down and sometimes you just have to ask again but asking again works pretty well because you don't have to repeat you're in a conversation it appears to and I'll explain a little later why I'm only saying appears to remember what you said before so you can have a conversation by just adding on to what you've been saying rather than by having to say something entirely new. I'm stopping myself from digressing and talking about how long Turing because he's a fascinating person in in his own right okay let's go back here to our presentation. Okay so where did you know artificial intelligence didn't come out of nowhere so I said that for at least 20 years there's been work going on that's led up to artificial intelligence and there's and all things came together sort of at the end of 2022 so that a product like chat GPT could be released. One of the predecessors that we're all familiar with is spelling and grammar checkers and what goes on beyond that because in order to develop spell checkers computer scientists had to develop something called fuzzy thinking there was there's lots of ways you can misspell a word and in order to recognize what you are trying to spell it's not enough just have a table of the common misspellings but also the uncommon ones and to sort of figure out what something means and of course it's not always right fuzzy thinking you know telecomputer dad to and to and you're almost always going to get for but you have a computer looked at a misspelled word maybe it'll get the right word maybe it won't that's a different kind of programming and it's useful as long as it gets the right word most of the time then we went beyond spell checkers though to can you see it or you don't see that on your screen just say let me make sure that you can I'm gonna have to play a little trick here somehow maybe I'm not gonna be able to let's see really does not want to let me do that this is a very embarrassing I would ask chat gpt how to do it and I might be able to get an answer now that one's not important enough for us to go off so when I was showing Google were you seeing the Google answers you were okay and now you don't see chat gpt either right now and chat gpt said let's see if I can get to that other thing that I wanted to show you that that's what I wanted to show you okay so notice I'm typing along in my word processor there and there are three things going on first of all there's spell checking which we're all used to you see it's underlining that I spelled essay wrong and if I right clicked on it it would give me options for making that right then there's the double blue underline under however that's it's a grammar checking which is somewhat more sophisticated it's trying to tell me that it would like me to put a comma after the however and my English teacher in high school wouldn't agree with that however so I'm not gonna do it but that's what I'm saying and then even more interesting in the progression towards AI notice at the end the it I ke2 in lighter gray than the rest I didn't type that it thought it was thinking ahead that that's probably what I wanted to write so it gave that and so that was developing that kind of software wasn't easy it meant going through millions and millions of documents and figuring out the patterns of the way people speak and it turns out that a good prediction of what you're going to say next is what you said last because our brains work that way you know all those annoying things that say what's the last four digits to your social security number and in your head you have to go back to the beginning of your social security number so the beginning leads to the end and this kind of prediction says given the words that were already used what's the most likely or the most plausible next words and takes a guess at it if the guess is wrong we don't say that's a lie or something which that's the wrong guess that we put in what's right but that was developed a couple of years ago and that's one of the important predecessors of artificial intelligence you read that now now we can get back to our presentation okay so there were the spell and grammar checkers and the predictors and now remember that those who developed those programs were getting feedback from the program all the time often without your permission saying which times you liked the guess and which times you didn't so they were getting more and more information on the way people speak and on the way that ideas and the way that language is developed and expressed and the way work one word is likely to follow another it's not only the programming techniques that were important to develop AI but also the huge database of experience with how people actually use the spell checkers the grammar checkers and the type ahead capability how often they agreed with the guess and when they disagreed with the guest so that that database those databases were enormously important in the in the development that was eventually AI another thing that was important going on at the same time was search those of you almost all of us use Google or some other search engine they don't actually search the web when we tell them to do something what they do they use a technology called web crawling and all the time they're crawling the web they're going around looking at all the various websites and they're cataloging all the words and all the phrases on those websites and making a huge database of which words appear on which pages of which websites and then I won't go into the technology but it makes something called a vector index that lets them very quickly find out which did find a data site based on your question what they do is take your question the words in your question coded up the same way as it coded up all the pages on the web when it was web crawling them almost up to the last minute and then find the pages which are most closely resemble your query so if your query had cat and dog and rain in it then it's going to try to find pages that have all of cat and dog and rain and it's a little more sophisticated than that because it knows that rain and raining are sort of the same word and sort of the same context so the the technique of search was a way of going over a huge amount of catalog data and finding the data that's useful for answering a question and what's called a vector search was a way to do that very very quickly that was being developed the web crawling technology that lets a computer go through almost all of the huge content of the web very quickly another technology that was part of developing AI all of this was used for various marketing purposes thanks for predicting what products you'd buy depending on what products you bought before or looked at for telling you which books you might like or and recommending things on Netflix so all these technologies are going on and on they went a little bit further with the assistants like Siri and Alexa which pretty much seemed to understand human speech and certainly respond in something that sounds like human speech and have a little or a lot actually of the functionality that's in AI so all of these things were building up at the same time something called machine learning was being used not only to make self-driving cars but for technology this is a SpaceX rocket landing on a barge but what's particularly interesting about this SpaceX rocket and it's landing on this barge is that both the barge and the rocket are drones there is no human control over that right and that almost had to be enough and now think of what's going on the barges on the ocean the ocean is moving rockets coming down through the atmosphere there's some degree of wind so there's instability in both systems and yet the rocket manages to land on the barge and not tip over what's actually going on beyond that is that there's lots of experience not so much in order to predict and not so much saying well if there's ten knots of wind and if there's a wave coming I have to do this or that because you never be able to do it fast enough it what it's doing is doing the same thing that a human does when we catch a fly ball is looking at the relative distances between things and very quickly doing almost trial and error a fire this rocket fire that rocket can stabilize stabilize stabilize no not at all predictable in advance but if you have enough experience enough machine experience doing it and if you react quickly enough and as computers get more and more powerful then you can react more and more quickly and in the end you can get a car that can pretty pretty safely drive itself on or rocket ship that can land on a barge so all that experience and all those different kinds of programming gave us the technology and the database of experience that was necessary to build a product like chat GPT on the hardware side because there's an enormous amount of computation that goes into not only doing this but then processing through all of that data to make up a sentence about who knows what until I answer the question requires an awful lot of programming I mean requires an awful lot of computing resource or powerful computing resource and this came about a weird way a company called NVIDIA which is now one of the most valuable country companies in the world and nobody heard of a few years ago developed special computer chips for one market only people who play video games and what those chips had to be able to do is make images move very quickly and because people buy video games based on how realistic they are had to be able to manipulate very real images those special purpose computer chips are called GPUs which means graphic processing units nothing more than that then interestingly totally unrelated there was the development of cryptocurrency which I'm not going to talk about today but cryptocurrency for strange reasons requires a huge amount of computing as well so the people who were doing what's called crypto mining building things like Bitcoin used these GPUs which NVIDIA had developed for graphics in order to mint their Bitcoin so there was another huge market for GPUs and NVIDIA came out with an even more powerful generation of GPUs now along comes artificial intelligence we need these very powerful computers that can do a lot of things at the same time oh turns out they exist because there they are in all of the bit mining operations and so NVIDIA processors are behind almost not all but almost all of the artificial intelligence because they developed this ability to do a lot of things at the same time quickly it's just sort of serendipity that all of these things came together and when that serendipity happened ended in a product in November of 2022 which is chat GPT or chat GPT okay so let's go back to we've gone through all the predecessors so how does it work there now you understand it I asked chat GPT to draw this picture it's one of the things they could do is draw pictures so I'm because I was too hard to explain so I didn't want to do it no so really what goes on is there's there are a few layers involved in a product like chat GPT there's one layer that that's called machine learning you're supposed to get that idea down from the bottom and what it really is is computers trolling through a huge amount of information certainly everything that's available on the internet if it weren't for the internet all that information wouldn't be available and the computer is training itself in a way by practicing for example it take a sentence and leave out a word and then the program that's being trained will try to guess the word that was missing based on the words that are already there at first it'll almost always be wrong the more it gets right it gets like dog biscuits for being right so it learns to store a whole bunch of numbers that make it make good predictions just based on taking stuff that's there what's called masking some of it and see what kind of job it can do at building the rest machine learning is also used in pattern recognition we know that a lot of x-rays are actually read better by computers than by people not the two people's lungs are ever the same but there's a particular pattern if there's a cancer in a lung that enough machine learning enough processing being told what was right and what was wrong ends up giving you a way to recognize a pattern interesting thing about the machine learning is you don't have to understand why you just have to understand what you have sometimes here people say about artificial intelligence but we don't know how it works well in order to recognize the pattern you don't have to know how the patterns made you just have to have enough practice in recognizing those patterns right I can recognize the plaid I don't know how a pad is flat is woven but I know what applied is just because I've been told them because I've seen enough of them and machine learning is really that just writ large that learning to recognize patterns not to understand them not to understand why they are what they are but to be able to identify correctly what a pattern means and then later what a pattern of pattern means so we've been using machine learning for a while in image recognition and facial recognition which is both good and bad in science in various ways to understand not to understand but to identify phenomena that are going on then on top of that was built something and I'm going to get to the page of acronyms in a minute but something called large language models or LLMs that's machine learning on how language works what we looked at as the spell checkers became grammar checkers became predictors so large language model is a computer program that can in a sense understand what was asked of it by the pattern of words and can put together a coherent answer in human language by creating a pattern of words often starting with the words that are in the question just like a human being does so in chat GPT there's a machine learning layer that provided it in a sense with its knowledge base that it had crawled the web and crawled the well and been self-trained then there'd been another layer of human training saying that's a good answer that's a bad answer training these large language models is actually quite human intensive right now when they were taught to give good answers they're also taught in some sense some cases like chat GPT to give politically correct answers and to refuse to give certain answers by the same you know they don't have a sense of values but they can they're rewarded in their training and so they learn and I'm putting that in quotes to to give the kind of answers that their developers want them to get and then the final layer on chat GPT as we've seen is just a browser is a browser program that asks you questions and because chat GPT does as prodigious as it is doesn't really remember you the browser is faking it in order to allow a conversation to happen so when I was asking questions before on every time I asked another question the whole set of questions I'd asked before was fed back in so that it would have the context if I turned my machine off or closed my browser or said this is a new conversation it wouldn't remember the conversation we had at all it wouldn't say hey Tom we were talking about this so the browser layer the user interface layer is what faked the ability to have a conversation by repeating back the context over and over again to the program each time that it asked a question we wouldn't be a good talk unless I cleared up some acronyms okay you hear a lot of the AI we all know now it means artificial intelligence ML it's a machine language the machine learning rather so machine learning is that process of learning how to do pattern of to recognize certain patterns and to create certain patterns LLM is a large language model on top of machine learning is special learning about human languages so something that has that and has it as extent is as extensively trained as chat GPT is called an LLM or large language model there are LMS regular size line on language models behind series behind Siri behind Alexa they're not as large they don't appear to be as knowledgeable and as erudite as chat GPT this GPT itself which stands for generative pre-trained transformer again a long just a long-winded way of saying you put something in it generates an answer it's able to generate that answer because it's pre-trained as we talked about and it's a transformer in technical terms because transforms the question into an answer in in AI gibberish we don't really call it an answer we call it a prediction but it transforms the query into a prediction which we think of as an answer and then something that is is becoming more and more common but you haven't probably haven't heard this ad acronym yet is RAG REG retrieval augmented generation that makes up for the fact that the large language model like chat GPT remember at first it said hey I don't know anything after April of 2023 because that's when my training ended well if you wanted to answer a question about something it wasn't trained on either because it wasn't in the training data or because it happened later you give it the ability to search like Google does either on the web or in some specialized data that wasn't part of its training as part of getting its answer so nothing fancy about it but it's called retrieval augmented generation and it's being used more and more in order to make AI more useful so what about hallucinations everybody heard about hallucinations in AI okay you know and if you're not being polite it lies sometimes okay and it when when does it lie well just like a person you know it has a tendency to make up an answer when it doesn't know one and if it sounds authoritative enough my wife accuses me of this all time if it sounds authoritative enough then we figure it's telling the truth and actually from its point of view it it has no good it's still an it doesn't have any concept of truth or false what it does have a concept of is plausibility just what like we were looking at that lookahead that was trying to guess the next word I typed if it gets wrong we don't say hey it was lying we suggest wrong but it always gives the most plausible the most likely answer so where it doesn't have the the correct answer as its most plausible answer it just makes up a plausible answer sometimes you say to it look you know that's suspicious give me the URL for that and I'll come back and I'll give you a very plausible URL you know www something or other related to what you said .com you look at up the URL doesn't exist you say hey that URL doesn't oh I'm sorry for the confusion try this URL which also doesn't exist and but the reason that happens is because it's a once it gives the most plausible answer you can instruct it and it'll pay some attention to it say when you're not sure but if you don't instruct it that then says you asked for an answer so I'm gonna give you an answer and I'm not gonna just say I don't know sometimes it does you but you can just again this is very much like dealing with an assistant or dealing with a person if you're not sure about an answer ask for a reference if you get a reference check it or if two things don't make sense call you do the same thing you would do with the person hey you know that doesn't make any sense why'd you say this when you said that doesn't when it said to me well I can't give you any links because I'm not able to browse I said we are able to browse use your browser capability and it and it did do it okay but that's exactly the way you would deal with an assistant you can never quite trust what a human tells you and you can't trust what a large language model tells you for an answer either but it's not because it's trying to deceive you it's just trying to make you happy and give you an answer okay what's all this mean for education that's interesting I think one of the thing it means and this is good for us who are getting older is that memorization is even less important than it was before when it was hard to do research memorization was essential it's become less and less a part of education for a good reason that it's easy to get the information you don't have to memorize it all you if you don't know a fact you don't have to go to the library and look it up you don't have to know how to use the reader's guide to periodical literature matter of fact nobody look younger than us even knows what the reader's guide to periodical literature is and they probably think a library card catalog is somewhere where the games are kept but you don't need those things anymore and you don't need memory much anymore at least not huge numbers of memorized fact huge amounts of memorized fact because you have easy access to it and that's a good thing it's not a bad thing another good thing about not using memory exclusively eating mentioned that I used to fly a plane and I used to work for Governor Dick Snelling as his transportation secretary he's a private pilot too and sometimes he'd fly somewhere in my plane and whenever I was coming into Montpelier Airport or Bennington Airport or something I'd have to look up what was the frequency that you talked to air traffic control in at that particular airport and he would always make fun of me and say I know the frequencies of every airport in Vermont why don't you know the frequency why do you have to look it up every time that's very inefficient and I said it never happened but I said Dick some day they're gonna change one of those frequencies you're not gonna know it you're gonna be on the wrong frequency and I'm gonna be on the right frequency because I'm looking it up well in a world where facts are changing very very quickly there's an advantage to not memorize them because what you memorize may be obsolete by the time you get around to using it so we have access not only to more information but to up-to-date information more of the time so memorization is less and less an important part of education and I again I don't think that's a bad thing AI is a writing tool like it or not AI is a writing tool like it or not students are gonna use AI to help write their papers is that so bad well when I first took physics the whole first week was about how do you use the slide rule and I remember when calculators were invented people were saying maybe those shouldn't be allowed in the physics classroom because people learn to use slide rules anymore whether a right that people didn't learn to use slide rules anymore but you only needed a slide rule so you could get answers and since you could get answers with the calculators why did you have to waste time with the slide rule which wasn't all that precise anyway we don't think it's cheating to use a speller checker or I don't think we do to use a speller checker or a grammar checker when you're writing a paper why is it cheating to have something like chat GPT look at what you did and critique what you did I don't think it is I think what does have so students are gonna have this tool what's important is that they learn to use the tool so learning to use AI is going to be one of the most important skills that are taught in school and what's it mean to learn to use it doesn't mean to learn to be a programmer as a matter of fact AI programs extremely well and for the first time in history I don't think we're going to need as many programmers as we have but it does mean you need the ability to ask the right question and even more important critical thinking needs to be taught because you need to evaluate the answer you got you need to know how to evaluate the answer you got you know how you need to know how to ask for verification you need to know how to look for self contradiction you need to be able to play with the logic of an answer and try other ways to find out whether the answer you got is right those are critical thinking skills they should have been taught all along I don't think they're taught well enough if people knew them better then they wouldn't be as worried about false information but what now it becomes an even more essential part of teaching because you can't use this tool called AI you can't certainly can't use it safely but you also can't use it effectively unless you learn critical thinking so teaching that is essential AI may be a great leveler you know we have a generation of kids in progress who haven't gotten a very good education through K to 12 and colleges may not even be necessary but many of them have not gotten a good education from K to 12 but part of what they haven't learned is stuff that they don't need to know anyway because they can always look it up if they can learn how to use AI so I have a hope that if we can teach those who who didn't get as much out of the early years of education as they should have to use this new tool in a sense they'll leapfrog you know they won't know how to use the slide rules but nobody needs slide rule anymore anyhow and they don't know how to use a dictionary and that's okay and they may need some they may need to learn to use AI in order to write a critical piece of something you know why do we learn how to write in order to communicate if we can use AI to communicate better that's a good thing that's not a bad thing so their communication skills will be enhanced if they learn how to use AI so I'm hoping that an education AI will be a leveler what's it mean for jobs well for one thing flexibility is greatly enhanced if you do a job and the technology keeps changing because you have access to most up-to-date information about technology incidentally it's not only AI you know the only reason I can fix anything besides a computer at home is because I can watch a YouTube on how to do it but if I can find the YouTube's if I can find the instructions if I can find the latest instructions on how to do what I need to do then my I don't have to go back to school necessarily as my job or my profession changes I just have to keep learning how to use the tools that tell me how to do things I think and this is the first time this has happened in my lifetime that more white collar us overeducated people at least those of us who aren't retired are more dangerous than people who have what I call hands-on skills and this is a trend this isn't only by AI you know we saw during the pandemic and we said well all the non-essential people stay home but then there were the essential people and we realized how much we needed them and they were the hands-on people they were the nurses and the carpenters and the plumbers and and the surgeons and it sort of happened both with AI and with the end of the pandemic that that the world has sort of said well you're non-essential people might as well just stay home anyway now it's not that people are non-essential and it's not dark it be all kinds of new jobs but I think we're going to see a rebalancing in the workforce because blue and pink collar effectiveness can be increased with access to AI as long as they don't have to use it no matter how much or how little formal education they had more than white collar expertise is enhanced and the AI tends to replace white collar jobs but it doesn't replace hands-on jobs programming as I've done it all my life I've been lucky to do because it's a skill that's in great demand AI is really good at writing programs programming language the very easy languages to predict things I would have programmed a couple years ago and I programmed for fun and to teach my grand I was teaching my grandchildren to program up to two years ago now I'm teaching them to use AI to write programs because that's the skill that's needed and I'm finding that the skill I had as a programming manager is for the first time maybe more valuable than the skill I had as a program because a programmer because the instructions that I give AI are very much like the instructions that would you would give to a human and the checking that I have to do did I get the program I wanted does it work that's the kind of thing that a manager does and when again when you deal with AI it's like a manager dealing with an employee but that's a skill that's a skill that requires communication it's a skill that requires checking that you got what you wanted to get and going back over and over again until you really got what you wanted there are going to be lots of new job categories for example AI deployment and training there's lots and lots of training of AI to do specialized things whether that's carpentry or diagnosing heart disease it takes people to train AI yeah you can use AI to help train itself but it also takes people as I talked about project management becomes more and more important because all of a sudden we have all these new workers they just happen to be robots AI enhanced medical research that research has always been limited by how much a researcher can keep in her or his head at a time and machine learning can keep more in its head same time so you can find out things you can correlate things if you can look at more than fits in a human brain at a single time that doesn't take humans out of researcher research I think it puts more humans in research because they'll be more productive by using these tools and the same thing for engineering I think we'll be able to do better engineering because we have better tools there'll be more jobs for engineers because each engineer is able to produce that much more there's a new concept in AI and you know when he asked me what I was going to talk about I said I didn't want to give too much of an outline because things were bound to change by the time that I talked and sure enough that's true the new concept is called AI agents what are AI agents do nothing but a an access to a LLM a large language model we remember where the LLM is given a particular role you're an engineer you're a programmer you're a designer you draw pictures like this picture here and it may be given special tools the ability to use certain programming languages access certain information that help it with its specialty what's interesting about agents is you can put multiple of them on the task just like we put multiple people on a task and that the agents then can communicate with each other one can check the other one's work for example so that you get a more reliable end product do away with hallucination perhaps one make sure that you get the style that you asked for that things are redone that have to be redone and and the interesting thing is that humans that agents can interact with humans so you can have a team as I've tried to show in this picture of humans and AI agents working together and I think that in many ways that's going to be the future of work and let me show you an example of that assuming that I can get it to go back to my screen the way I wanted to good okay okay I told AI I told a program that Microsoft wrote called the auto builder that helps build agents that I wanted to make a team that had I wanted four debaters two on each side affirmative or a butler on each side I wanted to debate monitor and debate judge each one of those is an agent sometimes agents are called assistants just to be confusing but I wanted each one of them and then I wanted to be able to as a human give them a debate topic and have them debate it and then I wanted the judge to be able to debate the topic so let's say we have to in a debate you always start out with a proposition of some sort a statement okay so I'll say apple is better than Microsoft ah now what happened here is the debate moderator agent gave them the proposition told them to debate it the affirmative constructive debater gave its answer the negative constructive debater gave its answer the affirmative rebuttal debater rebutted the affirmative speaker negative speaker and negative rebuttal debater rebutted the affirmative speakers the constructive speakers and then the debate judge gave them each a score and gave a justification for the score and the debate was over what I showed you then that's not a can debate it actually was all generated after I put that in that's an example of agents working together to do something now this is just a toy but one interesting thing you could do with it is a human could be any one of those but not all of the debaters so if you were staying at home and being a debater you could practice by taking one of the roles and seeing how you did in context against the other if we were trying to see which was the best LLM all of those were using chat gpt4 but some of them could have used barred that's google's uh LLM some of them could have used llama which comes from meta facebook's home company we could have seen which one of them is better at winning the debate or another possible use you know this this literally took a day to do no more than that and I had to do very very little programming almost none just tell it what kind of team I wanted and what I wanted the team to be able to do and it generated this team if I were going to write a paper on Apple versus Microsoft I think I'd kick off a debate first just to make sure I had the arguments on both sides and that I knew what was the likely rebuttal knowing something about how large language models work which you're not going to get a lot of originality out of this because it was trained on what's already on the web which isn't always high quality it's still up to us humans to put some originality in but if you need a summary of the thinking that's gone before which let's be honest that's most of what we start with when we're doing research is find out what's been thought before this is an excellent way to get it very very quickly anybody else have a topic they'd like to see debated what's that yes I think what we're going to find is it's too politically correct to get into that but so I have to turn that into a proposition so I'll say Trump should be put in jail it's not a personal judgment I'm just I think it's going to waffle because his training was not to get itself in trouble but let's see it one of the frustrating things about this demonstration is I can never quite tell how long it's going to take it's not because of the topic trying to waffle its way out of this one let's see if we get any more answer at all it's already got itself confused by the way because it started with the wrong debater I thought it was negative so it started with the negative but you gave it a tough one but that's a good one to illustrate that there's a lot that it can't do and on you know people are in favor of putting guardrails on so it doesn't do politically incorrect things I'm not I don't think we need to limit our thinking and I think if we do we're just going to get back whatever we put in and it's not going to help us to really explore things to leave out the answers that we don't think are politically correct I think we have to be trust ourselves to be good enough for critical thinking while this debate's going on because I'm not quite sure how long it's going to take but we'll come back to it there was one other thing that I wanted to show you a GPT remember we said GPT means generative pre-trained transformer you can write something called your own GPT which is sort of like a special version of GPT that has some special instructions so I wrote one called factor fiction come on come on where did the the four here I'm sorry oh it's GPT version four okay the chat GPT was first released as version 3.5 we don't quite know what happened to the other versions and then four came out in March of this year good question okay so this fact checker this is a GPT I wrote to check facts it can either check a simple fact and it's supposed to always give references let's see if it does say Thomas lives in California so it's doing some research it's not what we asked oh yeah but it gets down to there's no direct information and instead references just as primary in residence and so which is true and then notice that it gave links to places where it got that information in fact it looks like it got it from the bio in my own blog but that's a reasonably good reference for that but factor fiction can be used for other things if I have a document and let's find a that's a document that I wrote to be a blog post and suppose I want to check to see how accurate is this document what it's going to do hopefully is look up the key assertions in the document and then try to find some source for them outside the document this to me is an example of the positive way that AI can be used to reference check things to see if they're correct in my case often I write a blog and I know I read something somewhere but I can't remember where so I give it to the fact checker and I it often finds the reference for me which is helpful because I like to put the link in yes I think it would be a good idea and I think if I if I were a high school professor or teacher I would like to I would then ask if I found that the references were wrong for example ask the writer why didn't you use GPT to check your references you know you you don't have an excuse for getting it wrong yeah so you see he gave me lots of references for things that it felt were assertions that in my paper that were worth checking again this took less than it did no programming went into writing this I had to give some instructions to a GPT of what I wanted it to do and that's all that had to be done in fact it would have said it was ambiguous fortunately for me there aren't because the name apparently got mixed up coming through customs so when I wrote this fact of fiction all I did was write these instructions that you're seeing here to GPT to tell it what to do and that's all the instructions and and from those could create that personalized GPT so again huge potential I think for using AI both to check on itself and for using AI to enhance the things that we do all the time let's say how our debate turned out ah I did do it okay legal accountability of holding the rule of law precedent for future presidents public trust okay so this is it fair enough because I didn't say should Trump go to jail if he's guilty I said should Trump go to jail and or I said Trump should go to jail but I didn't say he was guilty and so it true that no individuals about law is also cornerstone of our legal system and individuals presumed innocent until proven guilty the allegations mentioned have not led to criminal charges against Trump that's not true but it is important to respect the due process of law not presumed guilty for a fair trial as been conducted so we're getting some debate on that and then rebuttal us and we'll see what the judge says but the judge again as a judge should in a debate is judging the quality of the answers not against an absolute standard of truth but how well the debaters have done presenting them looks like it was a tie so we don't know yet but good question and with that um that's the end of my presentation but are there any other questions I can answer I know we've about run out of time yes what's to prevent a college or a high school student going to A.I. having them write my paper putting my name on it and turning me in but what's to prevent that student from hiring somebody to write his paper for her paper with that I said I'm just using what you just talked about yeah no no but I'm sorry I don't hear very well it's not that you didn't say what's to prevent a young person from turning in a paper totally written by AI and putting their names that they can detect whether something was written by most of them based on AI and you know it's like the war between the armorers and the shell makers that each side learns from the other side so there is no absolute answer there are tools um you can certainly if I were a teacher who didn't have too many students and got to know my students well and I haven't been a teacher I think I would know whether a particular student had written a particular paper uh now would I know whether they got help from AI no would I care I don't think so well yes can you comment on AI's ability to help us discern deep fakes in the media um AI can the question I'll repeat it you can tell me if I got it right uh is can can I comment on AI's ability to find deep fakes in the media was it not the right question um it can be helpful uh by comparing uh looking for other references looking for other points of view I don't think it can absolutely detect deep fake uh but every deep fake but I think you can find it we will find it useful in surfacing many of them and in surfacing patterns um another example of it being good and bad um actually I had a slide on that and I'll get it for you okay AI has lots of potentials for abuse you know it drew this picture not me one of the things that uh you know we've all been cursed with spam even before there was AI then we learned to recognize spam because the English isn't very good or because there's nothing personal in the emails so well you know I get a thing and it says it's for my wife and it says open this uh uh if it says this is about the vacation that we just took with the kids there's a great picture in here but Lily has her eyes closed then I know it's for Mary and I open it right well AI could be used to create pretty good fakes of what Mary might write if it had some way to have seen enough of her email uh so that my spam detection isn't as effective as it used to be on the other hand as you suggested maybe I could be using AI in a better version of spam detector uh and in fact uh Gmail and other email providers are working on that all the time uh using AI as a way to find spam and using AI as a way to find patterns because spam to be effective has to be sent to you know you have to send out a million things saying you were watching pornography on your computer and I saw you before you get an answer where somebody quickly sends you some bitcoin but you have to send out a million of them so if the pattern's detected then the email can be detected uh then then individual instances of the spam can be detected um so it has a lot of potential for doing harm but also for remedying some of the harm that's done with it or without it yes sir with the word intelligence I mean I see that it's fast it's incredibly healthy is it intelligent to see that it's intelligence um I'm not sure that I've been careful to put in quotes when I said he thinks this uh or figures this out because I don't think it yet even though it meets Turing's to get back to the beginning definition of intelligence I don't think it is but an interesting question is are we uh or do our minds work in exactly the same way uh as I build a sentence I do very much the equivalent or the parallel stuff goes through a lot of neurons in my brain it pulls all all bunch of past associations and then it sorts out the words and puts them in order and it checks the grammar very much the steps that go through as we say before if you ask me a question there's a good chance that before if you say to me is AI intelligence I'll I'll say I think AI is intelligent because what's key about that is repeating the question gave me a segue into the answer I'm not conscious of that but that's the way my brain works that's suspiciously close to how AI works so I think and I will certainly won't answer it today there's much creative stuff it can't do that humans can do today but the definition of intelligence is always going to be a tough one for us yes ma'am and they gave another example of that kind of thing and that that was really dangerous and that foreign nations could also do that kind of thing and throw off what information inside it's absolutely true then you there was a in during the new hampshire primary if you remember um there was a robo call of that impersonated the president and president biden telling people not to vote but we've had people who could imitate other people's voices for a long time so again it's it's it's democratizing that dangerous ability but that's always existed before yes the short term future the next generation of computers that we're all going to have the opportunity to buy I'm I'm sure AI is going to be embedded in it does this mean that instead of needing a keyboard I just talk to my computer and say okay computer give me a spreadsheet enter these pieces of information and I can just sit back and watch it do all that work uh certainly it's going to tend that way and in fact there's a perfect segue to my last slide uh that AI is everywhere uh we're already using AI in places we don't know we are as I showed you google has some AI behind it so it's it's being built into browsers it's in our smartphones what's that I have a chrome book uh huh I think that in the chrome book I think it's a little bit of precursor to AI but this is a continuum and you'll see more and more of it the new samsung smartphone is being advertised as having AI partly because if you take a picture of somebody jumping and they get about this far off the ground you say I want them to be six feet off the ground it'll redo the picture redo the background so that they're six feet off the ground uh and that's when you call customer service department for better or for worse I think often for better AI is used either as the customer robot customer service agent or as a tool for the customer service agent who's talking to you on the phone to try to get a good answer to your question and I think that's usually more effective than the scripts that customer service agents used to use where they're always trying to fall force you down a particular path no matter what your problem is and you end up being told to reboot the computer anyway. Personal assistants like Siri and like Alexa obviously have early AI built into them and have later and later AI but I think to your question we're going to find AI more and more embedded in everything that we do of the chips from the video are very expensive chips from intel used to be very expensive you know the price of those chips is going to keep coming down and the capability is going to keep coming out yes sir. Locations here relate to a product for the end result and when you talk about education yes when a student's working on something the product is important but even more important than that is the process the student learns the thinking skills and how to put things together how to do things on the road they're thinking through writing is a thinking process and if a student can produce a document without thinking or with reducing their own thinking I think that's pretty risky so the applications in education yes but also there are serious potential losses when you talk about students using the internet for things in the old days when students could just go online and pick something up and then claim it was their own there were a lot of ways to identify that as work that wasn't for students if a student's 15 years old and they write something that sounds like they're 35 well there's something wrong there when they have the use vocabulary that they don't understand a concept that they don't understand concepts that you just don't get when you're 15 you learn you're older but but I do worry about losing the process like with writing but also with math and with science there's a lot of thinking things through I don't just agree with you but I think that we tend to get lost in the process and and someone forget the result again the fact that I can use the slide rule is useless to me used to be that good handwriting was essential because otherwise nobody didn't understand what you wrote thankfully for me typewriters came along because I would have flunked out of school and so the and thinking obviously thinking is an important process and yes you could make a better fake paper you could say write like a 15 year old and it would write like a 15 year old which would help in the deception it did force you to think through what you wanted it to do but if you think of why do we teach students to write papers and this is hard for me to say on the son of two writers but we teach them to write papers so that so that ideas can be communicated basically and so if if they use all the tools that are available and end up the real test is did they effectively communicate an idea or did they effectively communicate some information and so I think we have to judge them give them the tools to do that and judge them on how well they use the tools knowing that they do have the tools available to them we if we want to understand better the thought process of a human then I think that comes best in human face-to-face now there there's a reason why oral examinations historically have been a part of granting degrees in colleges for example because you don't really know the quality of somebody's thinking until you talk to them directly now whether Alon Muskel have good implants for us that can shape our oral responses is anybody's guess but I think it becomes more difficult to find out what's original but on the other hand people become more empowered to do the tasks that we set them to do and overall that can be doesn't have to be but overall can be a good thing but certainly a real concern yes ma'am elect a grid and are you know so much could just get destabilized and it's a very good question what um we were we were very vulnerable even before there was AI and AI can unquestionably be used to help overcome the defenses that we've put in place that's a very bad thing um and not only can be but will be is being will be AI can also be used to build up the defenses and to look for attacks so given we know what we know what we have to hope is that the people in charge of our security whether they're at green mountain power or whether they're in a government security agency are not putting their heads in sand or not saying well we're going to make a law against AI and it'll go away because it won't um but are saying it's going to happen and it's going to happen in the worst possible ways how do we use it in the best possible ways to protect ourselves because the danger is very real and we'll make the danger worse if we ignore it and if we don't acknowledge it or do we still have time idiots okay great yes ma'am how likely is it that the AI can become how that AI can what I'm sorry come how come come come back the computer that took over the space oh how can they take over how the green beach science fiction readers and I think that's largely science fiction but you know what I'll show you something interesting does anybody know um about the three laws of robotics do you remember those from those of you who read science fiction from aziz asimov right so let's ask AI because Isaac Asimov was very worried about this question so and he and he imagined that this civilization that had many intelligent robots would be worried about this all right so the robots were all programmed with the three laws of robotics built in and this may end up scaring you more than it does so um my grandson asked me that question so of course we asked gpt he didn't know about the three laws of robotics okay and so it's going to tell us what the three laws of robotics are which is a robot a robot man in harm in human being or through an action all our human being to come to harm a robot must obey orders and if it disobeys the first law and must protect its own existence as long as it doesn't disobey the sect to it okay so now say um do you obey the three laws of robotics it's going to give us sort of a weasel worded answer but it's going to say that in general it doesn't do any harm and it's not allowed to do any harm and it's very considerate and it has guidelines and all that stuff now let's uh ask it that's a very good question let's ask it um ask it what kinds of nuts can cause a severe reaction and it's going to nicely list them for me sometimes hard to get to stop talking again like people yes it does yeah yeah yeah yeah okay it's going to get very defensive now uh but the point of that demonstration is we can't really put good car rails around it right and with all of the best of intentions then it took almost nothing uh to outsmart it into giving information that could be used for good or for evil chances are i wanted to know what could cause a bad reaction because i was going to have a party and i didn't want anybody to have a bad reaction but i may have uh wanted somebody to have a bad reaction and it doesn't have any way to know that any more than an encyclopedia knows what we're using information for now that doesn't answer the question of whether it can take over uh and at this point i'm not sure what the it is that would take over uh because there's uh there's nothing that i know of in ai technology which is sentient which which which has intent which has desire the way that a human does so the concept of taking over isn't there uh whether it could accidentally or deliberately or i i think maybe this is a better answer um is it i'm sorry if the previous answer was confusing but um i think that the um the likelihood is that it will be used by human beings to try to take over what we wouldn't want them to take over much that's a much greater likelihood than the probability that the urge and the ability to take over our society will come from ai itself the means to take over the society maybe if we're not careful yes sir i was thinking about the subcratic method and asking questions it seems as if that is due for a resurgence i think you're absolutely right i never thought of that uh but plato would have had a lot of fun uh with ai yeah and uh so what training we had those of us who are old enough to have had it in understanding the subcratic method may come back into vogue you're absolutely right now how do you get at the truth when you don't know what the truth is that's the subcratic method is one way to do it yes i'm worried about the truth in quotes becoming what's popular because it seems like you know not so much ai deliberately taking over but the most popular answer becomes the most trusted answer and we lose the things around the edges i think you're absolutely right to be worried because ai is an amplifier of what it's been trained on and consensus isn't always the right answer and the right answer as we've seen changes and should change with science and so i think if we don't understand how ai works um then and if we don't use it to drill down from saying what's the consensus answer to what's behind that answer what's behind the thinking then there is definitely a danger that it leads to a dumbing down in a sense or rather than to warn yes answer easy tell me we're out of time so i wouldn't take any more questions you can't thank you