Engineering Ecosystems with AI - Sandy Pentland

okay uh let's get started hello everyone I'm Jim huachao professor of Citizen Transportation at MIT welcome to the MIT Mobility Forum designed by the MIT mobile initiative as I wrote in the email to everyone for this fall in addition to the classic academic presentation we also introduced three sub-series the first one are women leaders in transportation the second one are wills and Deals investment in mobility and the third one is the Synergy between venture capital and the startup immobility right so all three of the sub series we are curating the content at the moment and if you have any suggestions welcome uh to to send to me and you can design according on this uh then today I I'm really glad to have professors independent join us uh actually again three years ago uh Professor penderland helped actually open the MIT Mobility Forum as the first speaker on the topic of social consequences of a Mobility system and today I have a pleasure to have Sandy back again to discuss how to engineering ecosystems with AI this is actually a sneak peek of the kind of speech that Stanley will give at the U.S National Academy of engineering here right so our society needs to design and Implement a whole set of system of people on the technology but it seems that we have a difficulty doing well for instance our system for dealing with the pandemic dealing with climate change with congestion interpretation within quality with financial stress has been less than successful and much of the difficulties as Professor Pender argue are because of the unanticipated human behaviors and perhaps as a consequence the Congress new infrastructure Bill expands the definition of infrastructure to include human and social infrastructure so I'm so glad that today Professor Penman here will talk about how the new approach to engineer ecosystem that better integrate human behavior and also second discuss how the new technologies like a large language model llm such as chat gpp can help Professor penderland directs the MIT connection science is an MIT widening initiative and also previously helped great and direct the MIT media lab is one of the most cited computational scientists in the world so before I pass the Forum to Sandy I'd like to announce a few Norms of the Forum and also get people warm up first one is in the chat I invite everybody to type your organization your city and your local time so I got a sense of the audience here right so I would type MIT Cambridge Massachusetts and now it's a 1205. yeah no India Mumbai yeah if I ever want to type in the chat so that will get a sense that whoa yeah as we can see this diversity of the audience thank you thank you yeah secondly given the topic today it relates to uh the large language model so let me do a like a very brave po to get a sense of the background of the audience relate to the llm so I object the pronoun uh so the first choice is uh you're doing research on LL right making methodological contribution to it the second one is you're applying the large language model your own practice and the third one is your carrier user you try to gpp are aware of it and then last one no later of it right I'll give five more second [Music] okay I will close the poll and share the result uh Sunny you can see the result here sure yeah three percent doing research 16 applying 27 carry user then about 30 aware and the 25 no later yeah thanks so fun that I was to State the norm of this forum right I really want to reinforce that you know in most online sessions the default of the audience is silence uh in this forum we will try to challenge this default right so this is actually a request that everybody in the in the Forum giving you you choose to spend one hour here we invite everybody everybody to contribute one idea this idea can be in the form of a comment or a question and please type them directly into the chat in the second the third part the one will curate the questions from the audience and pass them to Professor penderland for his uh response to that right without further Ado let's welcome Professor Sally Penman president thank you thank you uh so a little intimidating to be here with over 200 uh presumably Specialists and faculty but uh as uh was mentioned this is sort of a warm-up for an even more frightening audience when I came at the U.S national Academy so um as you heard the motivation here uh is that we haven't been doing very well with our uh systems that include people so for instance uh vaccination systems uh some of the financial systems we've had crashes and things uh and so what's going on there and a couple years ago I gave a talk about models of humans uh that could be applied in these areas to be able to account for Cascades panics and other sort of Highly non-linear long-tailed phenomena and as uh General mentioned the new uh Federal infrastructure Bill uh includes social structure as its definition and it's because of the failures as you can sort of see so this is sort of a a little bit of a hint about if you want to get money out of the new federal infrastructure bill um maybe you should you know put a little of this in there and and that will be a differentiating uh element for you um so let me give you just sort of a simple recap and then we'll get off to the sort of llm AI sort of thing so imagine that you had a simple sort of extended census so you had traffic monitors you had other sorts of uh things credit cards we've used phones uh traffic sensors purchasing patterns you can use all sorts of things where you aggregate the data like census data so it's not particularly risky and you begin to ask things that are not normally known which is where does a community workshop and play if they go to one place to shop do they go where else do they go the second time because people's behavior is extremely patterned and having this sort of aggregate Mobility data which is real time uh can be real time uh poses minimal privacy risks Than People widely understand that it's not something that's a very threatening thing but look at the sort of thing that you can do so this is just a simple example recent publication um we looked at vaccination rates across the whole U.S and you can say well okay we know that people move from one place to the other and of course when they move they infect each other so there's a a transfer that happens and we can ask well so if you're in uh one infected area and you go to a unvaccinated area you're likely to do some damage how does that affect your vaccination strategy you can see a little graph off at the right that you could do uniform uh uh vaccination Campaign which will so one in percent increase in the vaccination rate actually does a pretty good job of a little over two percent at reducing uh infection rate random sampling where you just try to scattered around does the same uh politically it's often uh uh there's a push to do the least vaccinated or the minority uh communities and that in fact is a little better but if you take the people uh that are most Central in the mobility Network and remember this is not just going to work and it's not just uh um cars or public transportation it's all the mobility and you focus on them you can get an eight percent increase uh decrease in infection rate from that one percent in investment in increased vaccination rate and actually if you apply this sort of optimization techniques that are common in transportation you can reach almost 10 percent uh reduction that is hugely better than what we actually did and what it's doing is bringing in very simple information about patterns and Mobility uh within people it changes what you do pretty dramatically so the place that my group has been focused on and what I talked about a few years ago is this notion of social exploration so uh people explore in different communities than the ones that they lived in and that results in diffusion of ideas right so this is a lot of studies about this across the whole country in most continents uh all continents on earth most countries and it's been shown that this social exploration this uh intersect interaction between groups with different cultures different opportunities different skills is a strong causal factor and I mean typically we see numbers in the 30 to 50 percent of the variance range in the spread of commerce between communities uh historical uh study just came out showing that immigration from different communities accounted for a huge fraction of the Innovation rates in the United States during the 1800s and for growth and Commercial capabilities between countries I said a big study we did in China and as a consequence um this notion of sort of exposure between people very much like the disease exposure but this is exposure to transmit ideas and opportunities and skills predicts GDP growth in cities and neighborhoods around the world and we've done studies on four continents and found again quite regularly between 30 and 50 percent of the variance in GDP growth is accounted for by this Factor even after you control for centrality of the neighborhood education etc etc etc so it's huge but it's typically neglected so if you look at average person uh this is a where they visit diagram the big things are where they the big circles are where they visit more often the big arrows are the frequency with which they uh go from one place to the other and you can see that there's a core uh and then there's this periphery of places that people visit very rarely um different people even of exactly the same income exactly the same cultural background have different types of exploration some people are very worried shy uh and they don't explore much some people are ambulent and fearless and they explore a lot and what ends up happening is the ones that explore a lot even controlling for other factors you might imagine uh make more money over a period of time there's also uh their kids grow up to to be more socially mobile many many good things and as I mentioned we've done studies in four continents and this is pretty much universally true so you might want to say okay well can we design Transportation Systems to promote mixing between uh different communities communities with different skills different opportunities and the first thing to notice is is that we've shot ourselves in the foot in the last couple of years with the pandemic so on the left hand side at the top where it says April 2019 the color there the blues are places where there's mixing between different communities in the city of Boston and then during lockdown that's April 2020 obviously much less mixing but it hasn't come back that's October 21. it's still not back so we still have much less integration between our communities and of course that's hurting all sorts of things particularly minority communities and so Nick Carlos working for General uh uh as his advisor asked the question how can we design transportation systems that better promote this sort of mixing and this is an example of using public libraries as co-working spots and which public libraries would you establish the co-working spots in if you wanted to promote the most mixing so yeah he found that these four locations five locations are the best out of the current uh Public Library things and if you want you should look at Nick's thesis it's wonderful um so um the question then arises can we do a lot better than just having people rubbing shoulders and an observation is is that we really need to more so than we think so this is a recent PhD thesis of mine Isabella luaza uh looking at the U.S government's ownet job transition

and skills data and what she found is something that people really hadn't noticed much before which is is that the the medium skill the low skill and the high school skill communities don't intermix in other words if you get a job in your mid-20s that's a mid-skill job so you know you're helping run a restaurant or something you will never transition to a job that's a high skill if you're in a high skill job when you get out of school you will stay in that income category in that skill category and never have medium skill things and if you're low skill you're just out of luck you don't get into either of them except quite rarely so this is a complete failure of what you might call continuing education it obviously has our current educational system is not doing the best it could but once they get out of school they're frozen and that's really uh terrible so what can we do about that and the hopeful signs are that things like large language models things like chat GPT help medium skill workers more than high skill so that's really interesting in many different tasks we see that like programming writing things like that if you give the same tool to medium skill workers they become more productive they get a greater increment of productivity than if you give it to high skill workers and that suggests that maybe what we want to do is we want to be able to use this capability to be able to break that notion of that observed segregation that lock-in and of lack of skills um and the first question is to ask why is this possibly true and uh the various sorts of uh literature on this uh which has been one of our major uh uh Endeavors is to contribute to that literature and you see some of the uh citations there is how do you describe Human Behavior Uh how do you describe how humans make decisions and the typical thing of course is that you select an action a posterior probability of a of an action being good based on your estimated likelihood but in fact that's not what people do what people do is quite literally multiply that by the popularity the action among their community and they use that as a prior this makes a Bayesian estimation obviously um and the fact that most models don't include that popularity of action in the immediate network is why they miss things like uh panics and Cascades and other sorts of uh of trending things and one of the reasons that these AI tools might help medium scale uh think people more than high skilled is it can suggest actions that you hadn't thought of so if you're not integrated into the work Community or into the the the commercial Community you don't quote-unquote know the ropes if you're high skilled you probably do the fact that you don't know the ropes means you won't think of the right things to do and you can't of course therefore evaluate them as a good or not good line of action so how can you get uh AIS to do this well uh you can build AI that augment this social exploration and remember this is a major causal factor in income growth both for individuals and for communities uh and what they actually do is they estimate this from online conversations so I haven't endowed chair at MIT I got it uh when Marvin Minsky who named the field AI uh retired and what he always used to tell me is the most important thing in AI is that people weren't doing is common sense how do you get a sense of what everybody else thinks and that's what these llms do it's really interesting they are a statistical assessment of what everybody is saying uh when you use them of course you have to ask Kamina to what community so uh current llms are weird because they're trained on things like Reddit and Reddit is full of craziness if you train your llm on say physics journals you'll get something that's really pretty good about physics won't be 100 right but it will be the sort of common wisdom about physics and just to give you the 90-second tutorial because people are not familiar with it but there's an enormous amount of press you can imagine uh that there's a space of words that people say in any one particular language and I'm going to steal slides from Stephen Wolfram here who uh has done some of the best work on on reasoning systems and you can imagine that sentences are transitions between these roots so you see these conversations moving around in circles among the words and of course millions and billions of people are having these conversations so there's all these trade skills that are created over time among the word tokens and what the llms do is they use a neural net architecture to compress that and pull out the major uh paths or features of conversation among all of the people and these are huge so they can have a huge amount of uh conversations that they uh account for um and incidentally you hear all this stuff about large models recently a type of model that came out of uh MIT is called liquid neural networks and it is four or five orders of magnitude more efficient in terms of nodes and neurons than most of the ones that you hear about so size is not necessarily connected to uh performance here so what what these llms do is it gives you a probability distribution of the next word so if I gave it a prompt which is the best thing about AI is its ability to then it makes an estimation of the most likely next word which in this case is learn but it could be with a little less likelihood predict or make or understand and most of these llms add noise so you don't always get learned sometimes you get predict occasionally you get make and so on and so forth and in fact all of the possible completions of that prompt that beginning sentence make had a long tail distribution so that depending on the noise you can get words like run combine catch and talk and what it does then of course is it just recycles says the best thing about AI is ability to and then it adds create create worlds crates worlds that and it keeps going until it hits an end token and you have a sentence and that's all it's doing is it's giving you samples it's like a Monte Carlo sample of a probability structure which is built from all the word transitions that it observes in its huge amount of training and if you repeat that training with noise then you can get a probability distribution of what people think about a given B and of course that's exactly what you need to make decisions so so that's the 90-second version of it let me show you some consequences so uh this is something we did some years ago personal investing it turns out that uh people copy each other they learn from each other but they have limited capacity and if you use a computer tool to augment their social exploration and say hey look at this too other people like you do this what you find is they make more money and as you add more and more strategies to them as suggestions they make more money until they hit a point where there's a cognitive limit and they just can't keep it all in their head so that's really pretty interesting you could actually make people invest better cool here's one paper uh that came out of Wharton where they were asking uh llms and people for novel ideas for low-cost products in the first column in the middle uh to the left here is humans and the vertical action is uh purchase a purchase intent as it uh as evaluated by a large number of third-party people and you see the people come up with you know things that are pretty good but gpt4 does a little better on average and gpd4 with some prompted by some good examples um does even better so it's coming up with things that are on average a little more marketable some that are quite a bit more marketable and if you have the humans rated on uh novelty you find something that's a little bit different you find that the humans are a little better about novelty uh but not hugely and certainly not significantly uh and so people take this of course as the idea that oh llms are creative it's artificial general intelligence but no that's not true at all what you're doing is producing tools that uh bring together all the conversations that people have and that can be used to aid social exploration there's nothing new in there except a little bit of statistical Randomness would play with words uh and uh you needn't be afraid that it's going to take over the world on the other hand you're going to have to use stuff like this to make sure that you consider all the examples that you should do and many of my friends in the Sciences use these tools to say well are there other experiments I should consider or other tools so let's go to that so we all know about science citations and they make a sort of language so it's a little different than the llms are trained on you could train it on citations for instance what is the space of citations look like and it has many of the statistical properties that you see in llms and in language and this is an interesting uh little movie we created so you'll see at the top it says what year so it's doing all the way from uh let's see from the late 30s up to the present and there's that blue star which is a paper that was done back in the 30s and uh what you're seeing each dot is a paper in the physics literature and what this looks like is an amoeba searching the space so people are triggered by each other to write things that are nearby they try to find the edges so that amoeba moves in various ways and when the when the amoeba of science gets in the region of that blue star which is sitting there all lonely with no citations for 30 40 years right suddenly it gets a huge number of citations and what's interesting is you can actually do a pretty good job of predicting citation count before you write the paper so this is an example of precision of predicting a citation count uh the orange but at the top is uh uh the predictions based on social characteristics of the paper where the paper fits in this embedding space of the community can you imagine that if you know NSF and other people gave funding based on things that would contribute more citations uh now that you can actually do a halfway decent job of predicting and what's even more shocking is that this works with law cases so what what judgments are going to be most influential and patents which patents are going to be cited too and this is a paper that we're just completing and is submitted for review here's another one so RNA is also a language right it's the language of biology and when you sample microbiomes or you sample the sea water you get this huge number of little fragments of RNA but those fragments are are like the words in natural language and they co-occur a lot more than is would be uh statistically random I mean hugely more and they form a structure which you can analyze to find things that go together just like those words that I showed you with um uh with natural language uh llms and when you find these sort of clumps of things that are working together then you can say well what is the effect of those well here's one effect so this is a result of looking at some 2 000 cows and their microbiome over a long period of time it's a little spin-off that we have and uh it's reducing methane emissions and the uh sort of state-of-the-art uh uh naive treatment is you give all the cows this different sort of feedstock but by analyzing the microbiome you can almost double the efficacy of that and reducing uh uh methane so just to sort of point out what this means this would mean if you did all the cows in the world this way that would mean uh one or two maybe up to three percent of global emissions just by taking care of the microbiome in these cows pretty interesting um so the takeout messages here are we can improve social exploration between communities using knowledge of social interaction patterns and we can enhance those using AI by sticking to this notion that uh enhancement of context of social exploration may be a safe and effective way to address many serious problems and then a little bit of advertising for my more recent books and thanks thank you great thank you so much Sani this I I really enjoy this the way you your frame AI as a two to aid social exploration right I really like this this angle maybe what I'm quite relate to that is that many people ask the question about what AI player uh convergence role versus Divergent Row in terms of a different group of people Etc so here you you made the comment that LM may help the media skill more than the the high skill there right so I want to broaden it to say what what's your thought about ai's General impact on society and from the convergence versus the Divergence in that then the second one related to that is the the fact that LM helped media skill worker more than High School worker it seems is after the fact that we we empirically test it and that's the result but in their way we can engineering a system like LM with that purpose in mind right so we want design some two so that it converts achieve a conversion row beforehand even is it is it possible or just we just add to the mercy of the the the whatever we come out of this right well so what I see is is that there's an enormous race not to make bigger models no one's really trying to do that but to build specialized uh uh and actually quite smaller models so in programming and what is the data they're going to use to train on what's going to be experts and so the things that come out of that model are not going to help experts all that much but they are going to help and this is the evidence they help mid-skilled uh programmers and I showed about you know sort of marketers right so this is out of Wharton so if we try to I mean not people will use this technology in bad ways yep okay people use all sorts of technology in bad ways uh it's the people not the technology so what we want to do is try and figure out how we can use this technology to help people what I'm suggesting is given what we know about how people make decisions and how they learn that helping them with presenting them with the common sense what people normally do do is a great strategy it leaves the person in charge so the ethical a lot of the ethical questions are are minimized and it seems to be a major way where we can result in upskilling of people and perhaps transition them people from mid-skill to high skill jobs which is what we need as the AI deploys we need people who are able to deal with this stuff so I would suggest that what we have to do is focus there is how do we build things that help mid-skill people to go to High skill by telling them what quote unquote everybody knows in companies this is called the company culture people know like if you work for Ernst young you know what are the way to approach things but if you're a new guy interested in young you don't know that this could be very useful for that hmm thank you yeah so that relates to another point you mentioned that continuing education we've been calling for this for multiple decades right say we should have this lifelong learning uh the the school is just the starting point the learning Etc but based on your empirical findings we we largely failed to to achieve any results right yeah so here do you see this this lrm or AI can be opportunity really change that failure here right that's actually related to part of the the single project we've been collaborating on how AI can improve human capital on this right right give some thought on how this AI can play a better role in this continual education part of it well so this is embedded continuing education so you know if you're a new person in a business or a medium-skilled person trying to become more skilled this is something you can use to tell you what the common sense is and what other people do which will you will learn as a human uh and uh over time and so you be you get the the skills of how to do the job by looking at the things that are suggested that other people do and considering them for your uh situation and then of course you can reflect on these too you can say well I think this is the best choice and you can ask the llm what other people will say and it'll give you a sort of uh cultural average in terms of a response which you can choose to ignore of course right but but it helps you guide your thinking and your learning and that idea of having embedded learning is uh pretty clearly got some advantages over having to take time off work take separate classes blah blah blah yeah and then uh the the you mentioned at the the first half of the presentation that uh your library done a lot of work in in terms of this social interaction in the in the physical sense right people even when it's traveling in another neighborhood and maybe shopping another neighborhood also the fact that covet really heard that quite a bit in this right so that's a one way to uh to boost the social interaction is try through this uh physical infrastructure augmentation like the way we do and then the second half you presented like a AI or larger Lego model as another alternative way to boost the social exploration here right so how do you see the two potentially uh interact with each other or helping each other or substituting each other between the physical interaction way of versus the AI based way of boosting social interactions well I think the human interaction is the primary thing because there's all the stuff that isn't in language it's body language there's attitude Foods there's emotions and while some of those are a little have echoes in language they're not there completely and uh and you know we need we have a real need for human interaction so uh when we segregate ours our communities you can see all the things that happen one of the main things that happens is distrust and distrust results in polarization and we all can see a lot of that at the moment it's it's probably one of the major problems that we face as a society is just trust between uh different strata different communities Etc um so the physical is is absolutely critical on the other hand you also have this more sort of cognitive skills and experience that you've had and if you can accelerate that um that strikes me as a good thing uh and uh so that you you know like if you go to the party here's the things you ought to remember that they do at that sort of party right it's like you know don't wear that sort of dress don't say these sorts of things whatever it is you use the social norms and so if we can build things that do that I think that that will help the situation I see looking at the uh the commas here yeah you could do a lot of bad things too let's not do that okay that's the correct answer to all of those comments don't do that uh instead let's focus on the good things and then we'll figure out ways to uh discourage prohibit Etc the bad things uh I do a lot on that uh it's just that the U.N uh um meet board meeting about how to do things like that one of the primary things is you have to keep track of what this stuff is doing we don't do that in our society we you know we we don't know if the algorithms are plus or minus we don't know who they hurt and who they help because we don't keep track we have to keep track back and that has to be visible to other people so that we can find the bad things and we can figure out how to fix them if you don't know what's going on you're never going to fix it and that's another sort of major source of uh problem in our society lack of transparency and accountability right thank you so last set of question from me uh then what passed to to move on to curate to the audience question I encourage again people please type your your questions and comments to that right uh so it's a Sandy last night you recommended the paper the optimal human AI system for me to do as a as a pre-reading that in that paper so in that paper you you mentioned about this exploration versus exploitation right particularly uh the the difference between individual decision making versus the group decision making right it seems that AI can help the individual in a very different way of how AI can help with the group right for example you mentioned that as a group instead of choose the maximal likes good action you could distribute action into the frequency of action proportional to the likelihood so that you have enough exploitation but also you have enough room for for exploration there right so so this angle uh maybe the first question is how do you see the difference between AR helping individual versus having a group right that's why the second one is in the paper you mentioned that this optimality Union from the human AI system design has a solid this optimality property in the demand that have expressed the learning occur right for example you mentioned Finance right so what about the demands that we do not have such explicit learning outcome right then how do we think of about this better desire of this human AI system so the fundamental thing here is is that most of our culture uh focuses on individuals making decisions essentially by themselves rational individuals but that rarely describes humans uh you can go to school for 20 years and learn to do that it's hard that's why it takes 20 years in fact most of our decision is culturally bound and uh and that's a good thing because that means that we can learn from the experience of others in terms of say sampling Theory or optimal estimation or any of those sort of classic ways of looking at determining best strategies learning from others is is a huge win it multiplies your abilities by uh you know orders of magnitude and that's the sort of social thing in any particular Community the learnings of the community we often give it names like culture uh this is the way we do things here and it can be wrong but it it is sort of the Learned answers of this group of people and um you know approaching that is uh as opposed to changing the individual is often something that's easier because you're adding other opportunities to the to the community and people have that rationale element and they will tend to make those choices in terms of objective outcomes where you can see things uh quickly and and in a very hard way um yes learning stuff happens pretty well there because you can see the result in other things and this is one of the major problems we have as a society is we do things that look good in the short term but in the long term they're terrible um so the short term long climate change is a good example of this right the finance is another good example of this there are many things like this and typically the only way to from a mathematical point of view to do that is to aggregate over larger and larger numbers more and more experience this is a sort of a question of ergoticity um so you know if it's just me making decisions I might do stuff that looks really good but then it kills me later if I have samples from hundreds of thousands of people I can say hey wait a second for some of these people it looks pretty bad and the trend is the wrong way so um that's like learning from a broader Community which we uh tend not to do because we don't have transparency and accountability we don't know outcomes of actions uh in some way that is reliable and and uh and truthy to use that word I like that word thanks I belong for the audience questions thank you Sandy for that fascinating talk the chat as always has exploded with comments and questions and I'll be sure to send you a copy so that you can go through it before your talk uh later on so I'm going to combine two threads of questions into one uh so you mentioned how most llms have been trained on Reddit and that's probably not the best source yeah it's a joke also but yeah yeah true better joke so I mean the mentioned that you know uh company cultures is one you give the example of how company cultures can be transferred you know in terms of knowledge transfer Workforce training and development so company cultures have a legacy of racism gender bias built into them so will AI perpetuate these cultures is this really a social benefit so I'm just trying to combine you know two these two questions into one together okay yeah so let me do that okay so first of all um the current generation of llms are trained omnivorously on everything including Reddit and that's probably a major source so we have a project called the provenance project which you can Google which is going through and serving open uh data sets about things that you care about like is this truly used by Third parties uh you know was this done under human subjects approval blah blah blah blah blah so that you can get much cleaner training than just having to like go over Reddit and I think that's the trend that will happen everybody's racing to establish good training corpai that um that can be used in a legal way and can be used in a way that is authoritative that won't cure all the problems but it'll cure a fair amount and then the the company culture and racism so forth um that's auditing transparency and accountability the real problem I mean look um the real problem is we don't know until somebody you know collects the evidence which is very difficult makes the argument which is very difficult there ought to be standard ways of monitoring these things and computers could do that it's not hard it's not expensive but then there would be a public repository where you could say okay so here's the company policies is it racist well statistical tests should take you a good solid you know 20 seconds to answer that action as opposed to five years of your life I remember a talk I was at at Oxford where a leading person in this field got up and talked about bias and everything and the Justice minister of Kenya said what you say may be true but if you've seen our current system you know judges Etc the humans in our system are incredibly biased and we don't call them to account either so I'm a big advocate of we ought to have auditing of everything we ought to be able to get feedback about actions versus sensitive categories and about outcomes both short and long and then we'd have a hope of being able to find policies actions that are good in both the short and the long term until we have that sort of auditing that sort of data uh I think it's it's all heuristic at best I know that's not terribly popular but that's why the sustainable development goals at the UN have data metrics is Because unless you have data metrics that people agree on it's all a bunch of hot air yeah yeah I mean yeah that makes sense you know auditing of an AI system to the through the data Matrix that you we proposed yeah so you initially in your talk you mentioned that llms may be helpful to medium skill workers uh so here there are two aspects so one is helping them with discrete tasks versus insights so how can AI enhance uh the medium skills workers with insights or can they not enhance them with insights and just help them uh you know with the way you mentioned it and but not you know take off take their jobs in the future so what sort of is the balance over there you know enhancing their jobs now but possibly take over their jobs in the future or can AI actually give them the insights they require to move up from a medium to a high skill uh job so so look um I don't have all the answers okay um what I do know is or what I believe is that um we can design things that will help lower skilled worker is to become higher skilled and that's a major step towards solving the problems that you just did will it solve it all probably not um but if we have better less segregation in our society so there's better spread of opportunities and we have a continuous upskilling going on it seems like we're likely to be better off than if we don't do those things if we continue in sort of a segregated siled Society I see there's a lot of things about privacy and yes you can do stupid things uh don't okay so one of the major threads of what we do is around privacy and how you do auditing how you do other things like that where you don't give up personal data at all uh you know I was the one that led the discussion at Davos that turns into the gdpr Privacy is sort of at the core of what we do um so you know uh take a look at it right take a look at sort of things that are happening there it's now possible to audit things in ways that are were impossible only a few years ago in terms of privacy and in terms of security there are some inevitable trade-offs uh if you're going to hold people accountable then you have to know who to be accountable but one can make a Judicial uh uh path for that uh and then of course the Judiciary has problems and so forth but but without data without knowing what's happening uh it's going to be very hard to do anything sensible so it's great that you got you brought a point of value privacy as well so there was one comment with mentioned that they'd be curious as to what percentage of conversations that happen across the globe on any given day are digital and uh you know there's this whole aspects of you know your car listening onto you all the time Alexa and you know Google home are listening to all these conversations and it being transferred so just your thoughts on you know you mentioned privacy the gdpr in in EU so are these really being used what are the ways that one can you know have a like hard turn off button on these devices well so currently they're not being used right first of all some of these things are relatively new from a technological point of view so that's not to surprise in summer the economic incentives of the big companies are not aligned with our incentives so one of the things that we uh try to work on is distributed systems where you control your data absolutely right and you can work with your community to get insights that you want and you can choose the community so if you look at transformers.mit.edu you can see a lot of the sort of stuff that we're we're thinking about in that sort of area because you know people want to learn from their friends and so forth that they don't want to like post it on Reddit that's crazy um there is this uh you know this really sort of delicate dance you have to do with data and privacy and accountability and the good news is is that the the mathematical tools are being and have been uh largely developed I think to do to do a lot of this there's still work to do uh the bad news is is that the current systems don't use these um and we have to cook up uh ways to encourage quote unquote those things which is going to be a combination of Law and economic incentives of some sort and that's a whole nother you could have a long conversation about how you change the economic incentives but people are are thinking about it they're doing it we're involved in that conversation um uh it's not being ignored it's just a difficult conversation let me just actually put something out there just so people understand okay um uh General uh Chief Secretary General Secretary G of China right so it's like the largest um representative Marxism Lottery's voice for Marxism recently said data is a new primary means of of production along with capital and labor and if you think about that what he's saying is is that classic Marxism is done it's now not a battle between capital and labor it's a battle between data capital and labor and that sort of gives you a sense of the magnitude of this problem and if you look at what Society did with capital and and labor it took a century or more for instance to form labor unions to pressure companies to establish principles to get laws enacted and uh the same thing with capital with agricultural Banks and credit unions and it's not a fixed thing it's not like you can do it once and it's done it evolves over time so currently we're in uh a new evolutionary phase of of Labor and a new evolutionary phase of capital the problem with data is we don't have any institutions we don't have any norms for it basically it's new and so we're back in the robber baron era of capital we're back in the uh uh the early Industrial Age where kids were working 14 hours a day that's where we are with data okay just face it and what we have to do is we have to like develop the right sort of Institutions to be able to deal with this now uh critical element of society something that is really you take seriously it's not just an extra thing or an irritant it's a major part of society yeah right ready to that is just yesterday uh the Senate Majority Leader Chuck Schumer is trying to organize a group of tech uh companies say let's let's think about something on AI regulation and what's the real federal government should play Etc right uh first of all what's your view on what federal government wrote you know what was the General era they showed into me but more specifically relate to the fact that AI can be used a tool to boost the social exploration on that aspect right is there anything that's necessary from the policy perspective to boost that or any concerns that we need to address well we work with the EU on their AI regulation and I work with Brookings which is sort of helped shape the Biden Administration and work with other countries too um and there are lots of details in there um the primary thing that I Come Away with is we don't really understand all the risks and dangers we don't really understand what are the good things uh it needs some more exploration but that has to be safe exploration and that's why I'm such a big fan of auditing so that you know perhaps the audits are not public but it ought to be the case that if you worry about something that you can get a court order and an examination of what it's doing against any sort of particular problem and that should be a matter of minutes not years of a dollar or two not millions of dollars in our legal system a lot of the big problems are identifying noticing that there's something wrong and then doing discovery which is a legal term which is hugely expensive and slow and then litigation we could take that and turn that into something that you know your average person could do for a dollar ninety five uh in you know half an hour that would be progress okay and fortunately that's part of the vision that people have uh they don't do quite as far as I do but I'll notice that the first actually operational AI regulation out of Singapore where what they're doing is auditing the AI if you want to release a product and you claim it does X you have to prove it before you get licensed to do it sounds like not a bad idea and the only thing that's wrong with their regulation is I want them to audit it every month okay because World changes uh and uh things go off the rails right that that's that's wonderful uh we'll get to the end of the hour now and this is such a brilliant discussion thank you so much Sandy everybody be join me thank Professor pentile for the presentation on the conversation it's really really wonderful aspect pleasure hope hope it helps this conversation broadly indeed India

2023-09-24 15:52

Other news