AI Frontiers: Rethinking intelligence with Ashley Llorens and Ida Momennejad

acquire different types of knowledge and reason  over different things. And this was at some point   known as g, and it’s still known as g. There have  been many different kinds of critiques of this   concept because some people said that it’s very  much focused on the idea of logic and a particular   kind of reasoning. Some people made cultural  critiques of it. They said it’s very Western   oriented. Others said it’s very individualistic.  It doesn’t consider collective or interpersonal  

intelligence or physical intelligence. There  are many critiques of it. But at the core of it,   there might be something useful and helpful.  And I think the useful part is that there   could be some general ability in humans, at  least the way that g was intended initially,   where they can learn many different things  and reason over many different domains,  

and they can transfer ability to reason over a  particular domain to another. And then in the AGI,   or artificial general intelligence, notion of  it, people took this idea of many different   abilities or skills for cognitive and reasoning  and logic problem-solving at once. There have   been different iterations of what this  means in different times. In principle,  

the concept in itself does not provide the  criteria on its own. Different people at   different times provide different criteria  for what would be the artificial general   intelligence notion. Some people say that they  have achieved it. Some people say we are on the  

brink of achieving it. Some people say we will  never achieve it. However, there is this idea,   if you look at it from an evolutionary and  neuroscience and cognitive neuroscience lens,   that in evolution, intelligence has evolved  multiple times in a way that is adaptive to the  

environment. So there were organisms that needed  to be adaptive to the environment where they were,   that intelligence has evolved in multiple  different species, so there’s not one solution   to it, and it depends on the ecological niche  that that particular species needed to adapt   to and survive in. And it’s very much related to  the idea of being adaptive of certain kinds of,   different kinds of problem-solving that  are specific to that particular ecology.  

There is also this other idea that there is  no free lunch and the no-free-lunch theorem,   that you cannot have one particular machine  learning solution that can solve everything.   So the idea of general artificial intelligence  in terms of an approach that can solve everything   and there is one end-to-end training that can be  useful to solve every possible problem that it has  

never seen before seems a little bit untenable  to me, at least at this point. What does seem   tenable to me in terms of general intelligence  is if we understand and study, the same way that   we can do it in nature, the foundational  components of reasoning, of intelligence,   of different particular types of intelligence,  of different particular skills—whether it has   to do with cultural accumulation of written  reasoning and intelligence skills, whether it  

has to do with logic, whether it has to do with  planning—and then working on the particular types   of artificial agents that are capable of putting  these particular foundational building blocks   together in order to solve problems they’ve never  seen before. A little bit like putting Lego pieces  

together. So to wrap it up, to sum up what I just  said, the idea of general intelligence had a more   limited meaning in cognitive science, referring to  human ability to have multiple different types of   skills for problem-solving and reasoning. Later  on, it was also, of course, criticized in terms   of the specificity of it and ignoring different  kinds of intelligence. In AI, this notion has  

been having many different kinds of meanings. If  we just mean it’s a kind of a toolbox of general   kinds of intelligence for something that can  be akin to an assistant to a human, that could   make sense. But if we go too far and use it in the  kind of absolute notion of general intelligence,   as it has to encompass all kinds of intelligence  possible, that might be untenable. And also   perhaps we shouldn’t think about it in terms of a  lump of one end-to-end system that can get all of  

it down. Perhaps we can think about it in terms  of understanding the different components that we   have also seen emerge in evolution in different  species. Some of them are robust across many   different species. Some of them are more specific  to some species with a specific ecological niche  

or specific problems to solve. But I think perhaps  it could be more helpful to find those cognitive   and other interpersonal, cultural, different  notions of intelligence; break them down into   their foundational building blocks; and then  see how a particular artificial intelligence   agent can bring together different skills from  this kind of a library of intelligence skills in   order to solve problems it’s never seen before. LLORENS: There are two concepts that jump out  

at me based on what you said. One is artificial  general intelligence and the other is humanlike   intelligence or human-level intelligence. And  you’ve referenced the fact that, you know,   oftentimes, we equate the two or at least  it’s not clear sometimes how the two relate   to each other. Certainly, human intelligence  has been an important inspiration for what   we’ve done—a lot of what we’ve done—in AI  and, in many cases, a kind of evaluation  

target in terms of how we measure progress or  performance. But I wonder if we could just back   up a minute. Artificial general intelligence  and humanlike, human-level intelligence—how   do these two concepts relate to you? MOMENNEJAD: Great question. I like that you   asked to me because I think it would be different  for different people. I’ve written about this,  

in fact. I think humanlike intelligence or  human-level intelligence would require performance   that is similar to humans, at least behaviorally,  not just in terms of what the agent gets right,   but also in terms of the kinds of mistakes and  biases that the agent might have. It should look   like human intelligence. For instance, humans  show primacy bias, recency bias, variety of   biases. And this seems like it’s unhelpful in  a lot of situations. But in some situations,  

it helps to come with fast and frugal solutions  on the go. It helps to summarize certain things   or make inferences really fast that can help  in human intelligence. For instance, there is   analogical reasoning. That is, there are different  types of intelligence that humans do. Now, if you   look at what are tasks that are difficult and what  are tasks that are easier for humans and compare   that to a, for instance, let’s say just a large  language model like GPT-4, you will see whether  

they find similar things simple and similar  things difficult or not. When they don’t find   similar things easy or difficult, I think that  we should not say that this is humanlike per se,   unless we mean for a specific task. Perhaps  on specific sets of tasks, an agent can be,   can have human-level or humanlike intelligent  behavior; however, if we look overall, as long   as there are particular skills that are more or  less difficult for one or the other, it might be  

not reasonable to compare them. That being said,  there are many things that some AI agent and even   a [programming] language would be better [than]  humans at. Does that mean that they are generally   more intelligent? No, it doesn’t because there  are also many things that humans are far better   than AI at. The second component of this is the  mechanisms by which humans do the intelligent   things that we do. We are very energy efficient.  With very little amount of energy consumption,  

we can solve very complicated problems. If you  put some of us next to each other or at least   give a pen and paper to one of us, this can be  even a lot more effective; however, the amount   of energy consumption that it takes in order for  any machine to solve similar problems is a lot   higher. So another difference between humanlike  intelligence or biologically inspired intelligence  

and the kind of intelligence that is in silico  is efficiency, energy efficiency in general. And   finally, the amount of data that goes into current  state-of-[the-art] AI versus perhaps the amount of   data that a human might need to learn new tasks or  acquire new skills seem to be also different. So   it seems like there are a number of different  approaches to comparing human and machine   intelligence and deriving what are the criteria  for a machine intelligence to be more humanlike.  

But other than the conceptual aspect of it, it’s  not clear that we necessarily want something   that’s entirely humanlike. Perhaps we want in  some tasks and in some particular use cases for   the agent to be humanlike but not in everything. LLORENS: You mentioned some of the ways in which   human intelligence is inferior or has weaknesses.  You mentioned some of the weaknesses of human   intelligence, like recency bias. What are some  of the weaknesses of artificial intelligence,  

especially frontier systems today? You’ve recently  published some works that have gotten into new   paradigms for evaluation, and you’ve explored some  of these weaknesses. And so can you tell us more  

about that work and about your view on this? MOMENNEJAD: Certainly. So inspired by a very   long-standing tradition of evaluating cognitive  capacities—those Lego pieces that bring together   intelligence that I was mentioning in humans and  animals—I have conducted a number of experiments,   first in humans, and built reinforcement learning  models over the past more than a decade on the   idea of multistep reasoning and planning.  It is in the general domain of reasoning,  

planning, and decision making. And I particularly  focused on what kind of memory representations   allow brains and reinforcement learning models  inspired by human brain and behavior to be able   to predict the future and plan the future and  reason over the past and the future seamlessly  

using the same representations. Inspired by the  same research that goes back in tradition to   Edward Tolman’s idea of cognitive maps and  latent learning in the early 20th century,   culminating in his very influential 1948 paper,  “Cognitive maps in rats and men,” I sat down with   a couple of colleagues last year—exactly this  time, probably—and we worked on figuring out if   we can devise similar experiments to that in order  to test cognitive maps and planning and multistep  

reasoning abilities in large language models. So  I first turned some of the experiments that I had   conducted in humans and some of the experiments  that were done by Edward Tolman on the topic in   rodents and turned them into prompts for ChatGPT.  That’s where I started, with GPT-4. The reason I   did that was that I wanted to make sure that I  will create some prompts that have not been in   the training set. My experiments, although the  papers have been published, the stimuli of the  

experiments were not linguistic. They were  visual sequences that the human would see,   and they would have to have some reinforcement  learning and learn from the sequences to make   inferences about relationships between different  states and find what is the path that would   give them optimal rewards. Very simple human  reinforcement learning paradigms. However, with  

different kind of structures. The inspirations  that I had drawn from the cognitive maps works   by Edward Tolman and others was in this idea that  in order for a creature, whether it’s a rodent,   a human, or a machine, to be able to reason in  [multiple] steps, plan, and have cognitive maps,   which is simply a representation of the  relational structure of the environment,   in order for a creature to have these abilities or  these capacities, it means that the creature needs  

to be sensitive and adaptive to local changes  in the environment. So I designed the, sort of,   the initial prompts and recruited a number of very  smart and generous-with-their-time colleagues who   we sat together and created these prompts  in different domains. For instance, we also   created social prompts. We also created the same  kind of graph structures but for reasoning over   social structures. For instance, I say, Ashley’s  friends with Matt. Matt is friends with Michael.  

If I want to pass a message to Michael, what  is the path that I can choose? Which would be,   I have to tell Ashley. Ashley will tell Matt. Matt  will tell Michael. This is very similar to another   paradigm which was more like a maze, which would  be similar to saying, there is a castle; it has 16   rooms. You enter Room 1. You open the door. It  opens to Room 2. In Room 2, you open the door,  

and so on and so forth. So you describe, using  language, the structure of a social environment   or the structure of a spatial environment, and  then you ask certain questions that have to   do with getting from A to B in this social or  spatial environment from the LLM, or you say,   oh, you know, Matt and Michael don’t talk to each  other anymore. So now in order to pass a message,   what should I do? So I need to find a detour. Or,  for instance, I say, you know, Ashley has become  

close to Michael now. So now I have a shortcut,  so I can directly give the message to Ashley,   and Ashley can directly give the message to  Michael. My path to Michael is shorter now.   So finding things like detours, shortcuts, or if  the reward location changes, these are the kinds   of changes that, inspired by my own past work  and inspired by the work of Tolman and others,  

we implemented in all of our experiments. This led  to 15 different tasks for every single graph, and   we have six graphs total of different complexity  levels with different graph theoretic features,   and [for] each of them, we had three domains.  We had a spatial domain that was with rooms   that had orders like Room 1, Room 2, Room 3; a  spatial domain that there was no number, there   was no ordinal order to the rooms; and a social  environment where it was the names of different  

people and so the reasoning was over social, sort  of, spaces. So you can see this is a very large   number of tasks. It’s 6 times 15 times 3, and  each of the prompts we ran 30 times for different   temperatures. Three temperatures: 0, 0.5, and  1. And for those who are not familiar with this,   a temperature of a large language model determines  how random it will be or how much it will stick to  

the first or the best option that comes to it  at the last layer. And so when there are some   problems that may be the first obvious answer  that it finds are not good, perhaps increasing   the temperature could help, or perhaps a problem  that needs precision, increasing the temperature  

would make it worse. So based on these ideas, we  also tried it for different temperatures. And we   tested eight different language models like this  in order to systematically evaluate their ability   for this multistep reasoning and planning, and  the framework that we use—we call it CogEval—and   CogEval is a framework that’s not just for  reasoning and multistep planning. Other tasks can   be used in this framework in order to be tested,  as well. And the first step of it is always to  

operationalize the cognitive capacity in terms of  many different tasks like I just mentioned. And   then the second task is designing the specific  experiments with different domains like spatial   and social; with different structures, like the  graphs that I told you; and with different kind   of repetitions and with different tasks, like  the detour, shortcut, and the reward revaluation,  

transition revaluation, and just traversal, all  the different tasks that I mentioned. And then the   third step is to generate many prompts and then  test them with many repetitions using different   temperatures. Why is that? I think something  that Sam Altman had said is relevant here,   which is sometimes with some problems,  you ask GPT-4 a hundred times, and one   out of those hundred, it would give the correct  answer. Sometimes 30 out of a hundred, it will  

give the correct answer. You obviously want  it to give hundred out of hundred the correct   answer. But we didn’t want to rely on just one  try and miss the opportunity to see whether it   could give the answer if you probed it again .  And in all of the eight large language models,  

we saw that none of the large language models  was robust to the graph structure. Meaning,   its performance got really worse as soon as the  graph structure, [which] didn’t even have many   nodes but just had a tree structure that was six  or seven nodes, or a six- or seven-node tree was   much more difficult for it to solve than a graph  that had 15 nodes but had a simpler structure that   was just two lines. We noted that sometimes,  counterintuitively, some graph structures  

that you think should be easy to solve were more  difficult for them. On the other hand, they were   not robust to the task set. So the specific task  that we tried, whether it was detour, shortcut,   or it was reward revaluation or traversal, it  mattered. For instance, shortcut and detour   were very difficult for all of them. Another  thing that we noticed was that all of them,   including GPT-4, hallucinated paths that didn’t  exist. For instance, there was no door between  

Room 12 and Room 16. They would hallucinate that  there is a door, and they would give a response   that includes that door. Another kind of failure  mode that we observed was that they would fail to   even find a one-step path. Let’s say between Room  7 and 8, there is a direct door. We would say,   what is the path from 7 and 8? And they would take  a longer path to go from it. And a final mode that  

we observed was that they would sometimes fall  in loops. Even though we would directly ask them   to find the shortest path, they would sometimes  fall into a loop on the way to getting to their   destination, which obviously you shouldn’t do  if you are trying to find the shortest path.   That said, there is two differing notions  of accuracy here. You can have satisficing,  

which means you get there; you just take a longer  path. And there is this notion that you cannot   get there because you used some imaginary path or  you did something that didn’t make sense and you,   sort of, gave a nonsensical response. We had  both of those kinds of issues, so we had a lot   of issues with giving nonsensical answers,  repeating the question that we were asking,  

producing gibberish. So there were numerous kinds  of challenges. What we did observe was that GPT-4   was far better than the other LLMs in this regard,  at least at the time that we tested it; however,   this is obviously on the basis of the particular  kinds of tasks that we tried. In another study,   we tried Tower of Hanoi, which is also a classic  cognitive science approach to [testing] planning  

abilities and hierarchical planning abilities.  And we found that GPT-4 does between zero and   10 percent in the three-disk problem and zero  percent for the four-disk problem. And that is   when we started to think about having more  brain-inspired solutions to improve that   approach. But I’m going to leave that for next. LLORENS: So it sounds like a very extensive set   of experiments across many different tasks  and with many different leading AI models,  

and you’ve uncovered a lack of robustness across  some of these different tasks. One curiosity that   I have here is how would you assess the relative  difficulty of these particular tasks for human   beings? Would all of these be relatively  easy for a person to do or not so much? 

colors and sizes. There are some rules. You cannot  put certain disks on top of others. So there is a   particular order in which you can stack the disks.  Usually what happens is that all the disks are on   one side—and when I say a three-disk problem, it  means you have three total disks. And there is   usually a target solution that you are shown,  and you’re told to get there in a particular   number of moves or in a minimum number of moves  without violating the rules. So in this case,  

the rules would be that you wouldn’t put certain  disks on top of others. And based on that, you’re   expected to solve the problem. And the performance  of GPT-4 on Tower of Hanoi three disk is between   0 to 10 percent and on Tower of Hanoi four  disks is zero percent—zero shot. With the help,   it can get better. With some support, it gets  better. So in this regard, it seems like Tower   of Hanoi is extremely difficult for GPT-4. It  doesn’t seem as difficult as it is for GPT-4 for  

humans. It seems for some reason, that it couldn’t  even improve itself when we explained the problem   even further to it and explain to it what it did  wrong. Sometimes—if people want to try it out,  

they should—sometimes, it would argue back and  say, “No, you’re wrong. I did this right.” Which   was a very interesting moment for us with ChatGPT.  That was the experience that we had for trying it   out first without giving it, sort of, more support  than that, but I can tell you what we did next,  

but I want to make sure that we cover your other  questions. But just to wrap this part up, inspired   by tasks that have been used for evaluation of  cognitive capacities such as multistep reasoning   and planning in humans, it is possible to evaluate  cognitive capacities and skills such as multistep   reasoning and planning also in large language  models. And I think that’s the takeaway from this  

particular study and from this general cognitive  science–inspired approach. And I would like to say   also it is not just human tasks that are useful.  Tolman’s tasks were done in rodents. A lot of   people have done experiments in fruit flies, in  C. elegans, in worms, in various kinds of other  

species that are very relevant to testing, as  well. So I think there is a general possibility of   testing particular intelligence skills, evaluating  it, inspired by experiments and evaluation methods   for humans and other biological species. LLORENS: Let’s explore the way forward  

for AI from your perspective. You know,  as you’ve described your recent works,   it’s clear that you have, that your work is deeply  informed by insights from cognitive science,   insights from neuroscience, and recent works—your  recent works—have called for the development,   for example, of a prefrontal cortex for AI, and  I understand this to be the part of the brain   that facilitates executive function. How does, how  does this relate to the, you know, extending the  

capabilities of AI, a prefrontal cortex for AI? MOMENNEJAD: Thank you for that question. So let   me start by reiterating something I said earlier,  which is the brain didn’t evolve in a lump. There   were different components of brains and nervous  systems and neurons that evolved at different   evolutionary scales. There are some parts of  the brain that appear in many different species,  

so they’re robust across many species. And there  are some parts of the brain that appear in some   species that had some particular needs, some  particular problems they were facing, or some   ecological niche. What is, however, in common in  many of them is that there seems to be some kind   of a modular or multicomponent aspect to what we  call higher cognitive function or what we call   executive function. And so the kinds of animals  that we ascribe some form of executive function  

of sorts to seem to have brains that have parts or  modules that do different things. It doesn’t mean   that they only do that. It’s not a very extreme  Fodorian view of modularity. But it is the view   that, broadly speaking, when, for instance, we  observe patients that have damage to a particular   part of their prefrontal cortex, it could be that  they perform the same on an IQ test, but they have  

problems holding their relationship or their  jobs. So there are different parts of the brain   that selective damage to those areas, because  of accidents or coma or such, it seems to impair   specific cognitive capacities. So this is what  very much inspired me. I have been investigating  

the prefrontal cortex for, I guess, 17 years  now, [LAUGHS] which is a scary number to say. But   been ... basically since I started my PhD and even  during my master’s thesis, I have been focused on   the role of the prefrontal cortex in our ability  for long-term reasoning and planning in not just  

this moment—long-term, open-ended reasoning and  planning. Inspired by this work, I thought, OK,   if I want to improve GPT-4’s performance on, let’s  say, Tower of Hanoi, can we get inspired by this   kind of multiple roles that different parts of  the brain play in executive function, specifically   different parts of the neocortex and specifically  different parts of the prefrontal cortex,  

part of the neocortex, in humans? Can we get  inspired by some of these main roles that I have   studied before and ask GPT-4 to play the role of  those different parts and solve different parts of   the planning and reasoning problem—the multistep  planning and reasoning problem—using these roles   and particular rules of how to iterate over them.  For instance, there is a part of the brain called  

anterior cingulate cortex. Among other things, it  seems to be involved in monitoring for errors and   signaling when there is a need to exercise more  control or move from what people like to call a   faster way of thinking to a slower way of thinking  to solve a particular problem. And there is … so  

let’s call this the cognitive function of this  part. Let’s call it the monitor. This is a part of   the brain that monitors for when there is a need  for exercising more control or changing something   because there is an error maybe. There is another  part of the brain and the frontal lobe that is   the, for instance, dorsolateral prefrontal  cortex;; that one is involved in working  

memory and coming up with, like, simpler plans to  execute. Then there is a ventromedial prefrontal   cortex that is involved in the value of states and  predicting what is the next state and integrating   it with information from other parts of the brain  to figure out what is the value. So you put all of   these things together, you can basically write  different algorithms that have these different  

components talking to each other. And we have in  that paper also, written in a pseudocode style,   the different algorithms that are basically akin  to a tree search, in fact. So there is a part of   the role … they’re part of the multicomponent  or multi-agent realization of a prefrontal   cortex-like GPT-4 solution. One part of it  would propose a plan. The monitor would say,   thanks for that; let me pass it on to the part  that is evaluating what is the outcome of this  

and what’s the value of that, and get back to  you. It evaluates there and comes back and says,   you know, this is not a good plan; give me another  one. And in this iteration, sometimes it takes   10 iterations; sometimes it takes 20 iterations.  This kind of council of different types of roles,  

they come up with a solution that is solving the  Tower of Hanoi problem. And we managed to bring   the performance from 0 to 10 [percent] in GPT-4  to, I think, about 70—70 percent—in Tower of   Hanoi three disks, and OOD, or out-of-distribution  generalization, without giving any examples of a  

four disk, it could generalize to above 20 percent  in four-disk problems. Another impressive thing   that happened here—and we tested it on the CogEval  and the planning tasks from the other experiment,   too—was that it brought all of the, sort of,  hallucinations from about 20 to 30 percent—in   some cases, much higher percentages—to  zero percent. So we had slow thinking;   we had 30 iterations, so it took a lot longer.  And this is, you know, fast and slow thinking.  

This is very slow thinking. However, we had  no hallucinations anymore. And hallucination   in Tower of Hanoi would be making a move that is  impossible. For instance, putting in a, kind of,   a disk on top of another that you cannot do  because you violate a rule or taking out a middle   disk that you cannot pull out actually. So those  would be the kinds of hallucinations in Tower of   Hanoi. All of those also went to zero. And so  that is one thing that we have done already,  

which I have been very excited about. LLORENS: So you painted a pretty   interesting—fascinating, really—picture of a  multi-agent framework where different instances   of an advanced model like GPT-4 would be prompted  to play the roles of different parts of the brain   and, kind of, work together. And so my question is  a pragmatic one. How do you prompt GPT-4 to play   the role of a specific part of the human  brain? What does that prompt look like? 

Great question. I can actually, well,  we have all of that at the end of our paper,   so I can even read some of them if that was of  interest. But just a quick response to that is   you can basically describe the function that you  want the LLM—in this case GPT-4—to play. You can   write that in simple language. You don’t have to  tell it that this is inspired by the brain. It is   completely sufficient to just basically provide  certain sets of rules in order for it, in order  

to be able to do that. For instance, after you  provide the problem, sort of, description … let   me see if I can actually read some part of this  for you. For instance, you give it a problem,   and you say, consider this problem. Rule 1: you  can only move a number if it’s at this and that.   You clarify the rules. Here are examples. Here are  proposed moves. And then you say, for instance,   your role is to find whether this particular  number generated as a solution is accurate.  

In order to do that, you can call on this other  function, which is the predictor and evaluator   that says, OK, if I do this, what state do I end  up in, and what is the value of that state? And   you get that information, and then based on that  information, you decide whether the proposed move   for this problem is a good move or not. If it  is, then you pass a message that says, all right,   give me the next step of the plan. If it’s not,  then you say, OK, this is not a good plan; propose  

another plan. And then the part of, the part that  plays the role of, “hey, here is the problem. Here   are the rules. Propose the first towards the  subgoal or find the subgoal towards this and   propose the next step.” And that one receives  this feedback from the monitor. And monitor has   asked the predictor and evaluator, hey, what  happens if I do these things and what would   be the value of that in order to say, hey, this  is not a great idea. So in a way this becomes a  

very simple prefrontal cortex–inspired multi-agent  system. All of them are within the same … sort of,   different calls to GPT-4 but the same instance.  Just, like, because we were calling it in a code,   it’s just, you just call, it’s called multiple  times and each time with this kind of a very   simple in-context learning text that, in text, it  describes, hey, here’s the kind of problem you’re  

going to see. Here’s the role I want you to play.  And here is what other kind of rules you need to   call in order to play your role here. And then  it’s up to the LLM to decide how many times it’s   going to call which components in order to solve  the problem. We don’t decide. We can only decide,   hey, cap it at 10 times, for instance, or cap it  at 30 iterations and then see how it performs. 

that I’ve done in AI and computer science since my  undergrad. What I would be interested in is—and I   have learned over the years that you cannot think  about the mind in general when you are trying to   isolate some components and building them—is  that my interest is very much in reasoning and   multistep planning, especially in complex problems  and very long-term problems and how they relate to  

memory, how the past and the future relate to  one another. And so something that I would be   very interested in is making more efficient types  of multi-agent brain-inspired AI but also to train   smaller large language models, perhaps using  the process of reasoning in order to improve   their reasoning abilities. Because it’s one thing  to train on outcome and outcome can be inputs and  

outputs, and that’s the most of the training data  that LLMs receive. But it’s an entirely different   approach to teach the process and probe them on  different parts of the process as opposed to just  

the input and output. So I wonder whether  with that kind of an approach, which would   require generating a lot of synthetic data that  relates to different types of reasoning skills,   whether it’s possible to teach LLMs reasoning  skills, and by reasoning skills, I mean very   clearly operationalized—similar to the CogEval  approach—operationalized, very well-researched,   specific cognitive constructs that have construct  validity and then operationalizing them in terms  

of many tasks. And something that’s important  to me is a very important idea and a part of   intelligence that maybe I didn’t highlight enough  in the first part is being able to transfer to   tasks that they have never seen before, and they  can piece together different intelligence skills   or reasoning skills in order to solve them.  Another thing that I have done and I will   continue to do is collective intelligence.  So we talked about multi-agent systems,  

that they are playing the roles of different parts  inside one brain. But I’ve also done experiments   with multiple humans and how different structures  of human communication leads to better memory or   problem-solving. Humans, also, we invent things;  we innovate things in cultural accumulation,   which requires [building] on a lot of … some  people do something, I take that outcome,   take another outcome, put them together, make  something. Someone takes my approach and adds  

something to it; makes something else. So this  kind of cultural accumulation, we have done some   work on that with deep reinforcement learning  models that share their replay buffer as a way   of sharing skill with each other; however,  as humans become a lot more accustomed to   using LLMs and other generative AI, basically  generative AI would start participating in this  

kind of cultural accumulation. So the notion of  collective cognition, collective intelligence,   and collective memory will now have to incorporate  the idea of generative AI being a part of   it. And so I’m also interested in different  approaches to modeling that, understanding that,  

optimizing that, identifying in what ways it’s  better. We have found both in humans and in   deep reinforcement learning agents, for instance,  that particular structures of communication that   are actually not the most energy-consuming,  not all-to-all communication, but particular   partially connected structures are better for  innovation than others. And some other structures  

might be better for memory or collective memory  converging with each other. So I think it would be   very interesting—the same way that we are looking  at what kind of components talk to each other in   one brain to solve certain problems—to think about  what kind of structures or roles can interact with   each other, in what shape and in what frequency of  communication, in order to solve larger, sort of,   cultural accumulation problems. [MUSIC PLAYS]