Skip to content

AI Is Learning Whom to Watch - and Influence (Harvard, MIT)

Exploring how theory of mind can enhance multi-AI agent systems by enabling social learning and trust among AI agents.

Ask about this video. Answers come from its transcript only — with the timestamp, so you can check them.

Generated from the transcript and can be wrong — check the timestamp.

Key Takeaways

  • Theory of mind remains important for AI to predict and understand other agents' behaviors in multi-agent systems.
  • Social learning mechanisms can be adapted from humans to AI to improve selective observation and learning efficiency.
  • Mathematical modeling of goals, beliefs, and costs enables AI agents to make rational decisions about whom to trust and observe.
  • Integrating linguistic and visual data enhances AI's ability to infer social context and improve collaboration.
  • Experimental validation with human behavioral data supports the feasibility of these AI social learning models.

What the video covers

  • The video discusses the relevance of theory of mind in multi-AI agent systems and its role in understanding goals and beliefs of other agents.
  • It highlights a recent study by Harvard and MIT on using theory of mind to arbitrate between social and nonsocial learning in AI.
  • Theory of mind is defined as the ability to understand others' thoughts, beliefs, desires, and emotions to predict behavior.
  • The video explores how cognitive human mechanisms can be abstracted into mathematical models for AI systems.
  • A framework called Lyus integrates linguistic and visual inputs to draw context-specific social inferences in AI agents.
  • Social learning in humans and animals is a powerful mechanism, and the video questions if AI can adopt similar selective observation strategies.
  • The analogy of trusting a best friend is used to explain how AI agents might learn from trusted agents by understanding their motives and behaviors.
  • The video connects previous discussions on multi-agent system topology with this new knowledge to optimize AI collaboration.
  • Mathematical models and cost optimization are used to decide when and whom AI agents should observe to reduce uncertainty.
  • The video also touches on experimental human behavioral studies combined with computational agent models to validate these concepts.

Chapters

  1. 00:00Introduction and Theory of Mind in AI
  2. 02:32Mathematical Models and Inverse Planning
  3. 05:00From Human to Agent Social Learning Patterns
  4. 07:31Imitation Learning and Model Distillation
  5. 09:56Human Behavioral Studies and Agent Tasks
  6. 12:42Grid Map Task and Experimental Setup
  7. 17:42Rational Mentalizing and Cost Optimization
  8. 21:40Value of Information and AI Observation Strategies
  9. 31:14Applications and Social Media Influence
  10. 37:16Neural Networks and Learning Worthiness Assessment

Answers

Questions about this video

What is theory of mind and why is it important for AI?

Theory of mind is the ability to understand that others have different thoughts, beliefs, and desires, enabling prediction of their behavior. For AI, it helps in modeling and anticipating the actions of other agents in multi-agent systems.

How can AI agents use social learning effectively?

AI agents can selectively observe and learn from other agents whose future behavior reduces uncertainty relevant to their goals, balancing the benefits of observation against the associated costs.

What role do mathematical models play in multi-agent AI systems?

Mathematical models formalize goals, beliefs, costs, and rewards, allowing AI agents to rationally decide whom to trust and observe, optimizing collaboration and learning in multi-agent environments.

Full Transcript — Download SRT & Markdown

00:01
Speaker A
Hello community. So great that you are back. Today we talk about multi-AI agent systems, and there's a simple question. Do we still need theory of mind? Is this still important? Because if we have here an understanding of the goal and of the belief system also from AI systems, is this a necessary condition?
00:15
Speaker A
goal and of the belief system also from AI system is this a necessary condition?
00:21
Speaker A
So let's examine here. We have a brand new study from Harvard University and Massachusetts Institute of Technology, MIT, published July 30th, 2026. And I thought, yeah, absolutely. This is mine, computer science. And then I looked at the title and said, what? Using theory of mind to arbitrate between social and nonsocial learning. What is this? This I want something to do with AI. Now, it turned out they apply here this idea to multi-agent configurations. And then I said okay now I'm interested and you will see that in my last video when we talked about mutable collaboration topology of multi-agent systems this is exactly what we talked about.
00:37
Speaker A
of mind to arbitrate between social and nonsocial learning. What is this? This I want something to do with AI. Now, it turned out they apply here this idea to multi- aent configurations. And then I said okay now I'm interested and you
00:54
Speaker A
Now what is theory of mind? It is the ability to understand that other people have thoughts, beliefs, desires, and emotions and that they may differ from my own, enabling however prediction of their behavior in social situations. So we humans are as theories here some social human beings and we are driven not only by pure logic and mathematical understanding but there is something like emotion and you got it.
01:07
Speaker A
thoughts beliefs and desires and emotions and that they may differ from my own enabling however prediction of their behavior in social situations. So we human are is theories here some social human beings and we are driven not only by pure logic and mathematical
01:26
Speaker A
So and now we want to find out hey wait a minute can we understand these cognitive mechanisms that we have in humans and take it, abstract it, put it in a mathematical formula and use it for our AI systems? And you know this happened to me already some time ago almost a year ago in 2025 here and there was a study by MIT, Harvard University, Stanford University, Yale University and I thought my goodness yeah this is something for me and I was totally disappointed because the title was language informed synthesis of rational agent models.
01:41
Speaker A
mathematical formula and use it for our AI systems and you know this happened to me already some time ago almost a year ago in 2025 here and there was a study by MIT Harvard University Stanford University Yale University and I thought
01:57
Speaker A
Yeah. For grounded theory of mind reasoning. And I thought come on what has to do psychology here with AI.
02:05
Speaker A
Yeah. For grounded theory of mind reasoning. And I thought come on what it has to do psychology here with AI.
02:17
Speaker A
Now they formed here a particular system they called Lyus as a framework for drawing context specific social inferences that integrate linguistic pattern with visual inputs. And if you want to have a look here, we have a beautiful environmental model, an agent model here with goal prior, reward prior, hours, belief prior, cost prior, hours. And then we have here involving states, object location attributes and specific action. And we have an inverse planning. So there's a lot of scientific literature if you really want to have a deep dive into this.
02:32
Speaker A
model here with goal prior, reward prior hours, belief prior, cost prior hours. And then we have here involving states, object location attributes and specific action. And we have an inverse planning. So there's a lot of scientific literature if you really want to have a
02:48
Speaker A
But let's go today only to our small little new paper here, Harvard University, MIT, July 30th. Let's have a look. So it is about social learning. This means acquiring information through others or with the help of others and they're saying hey this is a powerful mechanism by humans and nonhuman animals to acquire information about the world. And now the question is simple. Can we take this algorithm here from the human world and put it into the AI world? Because humans do not use social learning indiscriminately. We choose when to observe, what particular person or group of individuals to observe, how long to watch them, to follow them on social network and exactly we pinpoint who to learn from.
03:04
Speaker A
help of others and they're saying hey this is a powerful mechanism by humans and nonhuman animals that to acquire information about the world. And now the question is simple. Can we take this algorithm here from the human world and
03:17
Speaker A
Yeah. So this would be an interesting mechanism here for multi-AI agents.
03:36
Speaker A
learn from. Yeah. So this would be an interesting mechanism here for multi- AI agents.
03:45
Speaker A
Let's put it here in an analog. Trust your best friend. Yeah, I learn from just observing my best friend in action.
03:51
Speaker A
But I'm familiar with his motive, his goals in life, his crazy argumentation and his particular view on the world. So this is a human to human projection where if you want prediction of an understanding and now the question is an
04:06
Speaker A
But I'm familiar with his motive, his goals in life, his crazy argumentation and his particular view on the world. So this is a human to human projection where if you want prediction of an understanding and now the question is an AI agent trusts another AI agent in a multi-agent system. What can an AI agent learn from another AI agent that it trusts?
04:21
Speaker A
machine represents kind of an understanding of what the other AI agents are doing. What is I don't want to say motivate the other agent but what is the the system prompt of the other agents? Why are they behaving in a
04:34
Speaker A
So the idea is simple. Knowing the AI agent is really powerful and helpful to find a solution if this machine represents kind of an understanding of what the other AI agents are doing. What is I don't want to say motivate the other agent but what is the system prompt of the other agents? Why are they behaving in a particular way? What are their capabilities? What are their shortcomings? And what are the augmentation patterns learned by these other agents?
04:48
Speaker A
absolutely crazy yes I'm with you I think this is crazy and if would not be Harvard and MIT I would not present you the study and I would not have read this study but okay let's go now you
05:00
Speaker A
So here we go now from a human to human pattern to an agent to agent pattern and if you think this is absolutely crazy yes I'm with you I think this is crazy and if it would not be Harvard and MIT I would not present you the study and I would not have read this study but okay let's go now you understand in my last video we talked about the topology how multi-agent system can rewire itself and now I would say hey what a coincidence that this was in my last video I'm absolutely surprised and now let's use this new knowledge here for maybe to optimize yanta from my last video.
05:14
Speaker A
knowledge here for maybe to optimize yanta from my last video. So this means a socially intelligent agent. Now wow this is now this is now an update. No a socially intelligent agent does not simply ask hey who is the
05:27
Speaker A
So this means a socially intelligent agent. Now wow this is now this is now an update. No a socially intelligent agent does not simply ask hey who is the expert in the group for reading or performing some scientific calculation. But the socially intelligent agent says, "Hey, whose future behavior will reduce the uncertainty that matters for my particular goal and is watching this other agent worth my costs?"
05:46
Speaker A
And you say, "What?" Absolutely. We go with some crazy idea if you want you see the mathematics in the paper, but I try to make it simple and understandable also if you do not have a PhD in mathematics or theoretical physics or
05:58
Speaker A
And you say, "What?" Absolutely. We go with some crazy idea if you want you see the mathematics in the paper, but I try to make it simple and understandable also if you do not have a PhD in mathematics or theoretical physics or anything else.
06:11
Speaker A
I follow only the scientific expert or should I go with the successful agent and maybe the successful agent is just an agent that publishes the most on social media.
06:24
Speaker A
So we do have normally the heuristic learning and the heuristic learning is characterized by the question hey should I as an AI system copy here the majority vote. Should I follow as a lemming the majority? Should I follow only the scientific expert or should I go with the successful agent and maybe the successful agent is just an agent that publishes the most on social media.
06:40
Speaker A
a level deeper. Let's try to understand why this particular subject or object chosen here a particular action. So what was the goal of this person or agent and what was the belief that generated this action? So what we have now is a tupil
06:58
Speaker A
Yeah, not seeing anything else. And the theory of mind gives us another idea. The theory of mind has a simple question. What goal and belief generate this particular action either by a human or by another AI agent? So this tells us here let's go a level deeper. Let's try to understand why this particular subject or object chosen here a particular action. So what was the goal of this person or agent and what was the belief that generated this action? So what we have now is a tuple of the goal of somebody and the belief of somebody how to achieve it. What values are okay to achieve this?
07:08
Speaker A
So what we do is we take here this psychological theory of mind and we put it finally in a mathematical formula.
07:16
Speaker A
So what we do is we take here this psychological theory of mind and we put it finally in a mathematical formula.
07:31
Speaker A
remember here the imitation learning when somebody accuses here some model here to just learned and imitated another model just give you here the keyword distillation.
07:44
Speaker A
Yeah, absolutely. It is now a resource allocation mechanism that we can code in mathematics and later we can code in Python or C++ and implement it in a multi system. And finally, we're on the right track. Now what we want to do and remember here the imitation learning when somebody accuses here some model here to just learn and imitate another model just give you here the keyword distillation.
07:50
Speaker A
Remember you have 200 uh agents in your system and you have from this 200 I don't know 50 uh specialized agent here expert labels. Turns out it's all nonsense. So this new preprint suggests now that these are all insufficient when
08:08
Speaker A
Avoid the blind imitation. Current systems often rely on majority votes or expert labels or confidence scores.
08:17
Speaker A
There is some idea how to do a better multi- aent routing because in a heterogeneous team of multi-AI agents the normal and strongest agent may not possess the most relevant information for the learning that is necessary for the group as a total entity or even for
08:34
Speaker A
Remember you have 200 agents in your system and you have from these 200 I don't know 50 specialized agents here expert labels. Turns out it's all nonsense. So this new preprint suggests now that these are all insufficient when agents have different goals or some individual maybe some private nonproprietary or some non-public information.
08:48
Speaker A
agent here the most expensive cloud-based agent because this agent has to take all the decision, has to do all the dedication of work and allocation of work to the other sub agents. But it turns out this is not a good way to
09:02
Speaker A
There is some idea how to do a better multi-agent routing because in a heterogeneous team of multi-AI agents the normal and strongest agent may not possess the most relevant information for the learning that is necessary for the group as a total entity or even for a single agent in a group and this also goes if you think about an orchestrator agent normally we have a strict hierarchy and the agent on the top is the coordinator, the boss agent or the orchestrator agent. No, and you make this agent here the most expensive cloud-based.
09:18
Speaker A
hierarchical structure of multi- aent system combined to a swarm structure. And this paper argues in part that a hierarchical structure is not always the best because if you have intelligent sub agents, why not utilize their intelligence and their communication
09:37
Speaker A
patterns? Just have a look at my last video with manta. So orchestrator agent we talked about.
09:44
Speaker A
So let's start. This was just a warm up to introduction that we are all on the same level. It is not primarily an experiment I'm going to show you on LLMs or modern EI agent. This is really a
09:56
Speaker A
human behavioral study combined with a computational agent model. So they really recruited 300 Americans here [snorts] and did some tasks and I will show you the task we went through and then we had the eye doing here the same
10:10
Speaker A
task and the question was can we combine this behavior of the humans? Can we learn it and can we place it in an agentic model? And its relevance to AI comes from translating a component of the human social intelligence. Whatever
10:29
Speaker A
is a component of the human social intelligence. I'm here a theoretical physicist. I don't know any nothing about pology or neuroscience. But I know what it means. We want to place it into an explicit mathematical algorithm to calculate it and be able to encode it
10:43
Speaker A
here in a transform architecture. And we talk about a rational mentalizing model of social learning. Social learning is simply that if you are in a group with other EI agents or you in a group with other humans that you want
11:00
Speaker A
the performance of the group outperforms here the single individual added knowledge of each single individual member. So the model performs a simple rational utility comparison at each time step. Is any information I gain by watching other agent worth more than I
11:17
Speaker A
would lose here by not acting on my own goal? So you see what we have. We have now a mathematical optimization theory.
11:25
Speaker A
Simple. We say hey maybe I have I see 20 other agents here in my group. Huh? And I watch them now and I can learn just watching them being successful, being not successful. I understand maybe what drives them or at least I see their
11:42
Speaker A
action and then I can have a probability interpretation. What was the goal of this particular agent number 17 in this subgroup A and what is its belief system? So how does this agent has been pre-trained with a particular knowledge
11:59
Speaker A
with some particular skills with some particular memory markdown files that it is able to solve a particular task in a particular way let's call it here this is here what we transfer what we map from a human belief system into a
12:13
Speaker A
machine belief system and when should I say okay now it's I have to stop watching and I have to start my own execution I have to try I have to fail and I have to learn myself.
12:27
Speaker A
So as you already seen I told you about this tuplet. So we have we want to see and observe the agent might have a particular goal. G for goal and a particular belief. B for belief and said let's do this in an experiment. Let's do
12:42
Speaker A
this on a treasure map. Now so we have wizards who have some amolets where you can open some gates and the goal is to find some treasures. Maybe one, two, three treasures here on a particular grid map. Yeah, but the way to formulate
12:58
Speaker A
this in a mathematical way was simply let's combine this. Let's have here all our particular goals that we think of we have a probability of and let's have all our beliefs they form here a twolet and we call this twolet mental state
13:13
Speaker A
hypothesis. Okay. And then yeah you see we have to sum over i. So the index i identifies one possible explanation of the agent behavior. And if you sum over all possible explanation of the behavior.
13:26
Speaker A
Yeah. And we sum it up. Let's say we normalize it to one. You should have a good understanding here of the probability distribution. Yes. Again the probability distribution not only by the next token prediction but also now by
13:39
Speaker A
understanding here a learning process or the action prediction of a group of agents where we have to deduce here what is their goal and what was their belief structure.
13:53
Speaker A
Let's have a look at some sort of code. Now, this is the algorithm for the rational mentalizing model.
13:59
Speaker A
Let's have a closer look. So, what we do, we have some beliefs. Beautiful. Then we have some observable agents. Beautiful.
14:09
Speaker A
And then we have some particle for each agent. And well, as you see, we have here we simulate the future trajectory of an agent. We compute the updated beliefs from a simulated trajectory. And then we estimate some cost. So where if
14:23
Speaker A
you have a cost function, we have a mathematical optimization. Beautiful. So let's do this experiment. This game here takes place on a 12* 11 grid map containing here treasure maps, wizards, and some colored barriers that you need here from the visit here a particular
14:41
Speaker A
object that you can open up this gate. Either the treasure chest is either unobstructed or blocked by a colored barrier. And to pass through a barrier, the player must obtain an amulet from the matching collar from the wizard. And
14:54
Speaker A
among all wizard of a given color, only one of the wizard holds the amulet while the rest are just decoys.
15:01
Speaker A
So this is now here exactly here the grid that they worked with. This is MIT or now university. If you think it's too complicated, look, look at this.
15:11
Speaker A
Unbelievable. Now, [snorts] so here we have here our M. This is in red, the main agent. And then you see we have here three golden treasure chest A, B and C. But unfortunately the way to them is blocked. And we have here some
15:25
Speaker A
barriers. And for the barriers we have to have to get some keys from the wizards. And we have one red wizard and two blue wizards. And now you have here to move here as a human or as an AI
15:36
Speaker A
system and solve this immediately. Now what we have also is two different additional agent. In green we have a novice noit agent. So, an agent that has no idea, it's just trial and error. And then we have an expert agent in blue.
15:53
Speaker A
Now, the expert agent knows exactly to go to which of the blue wizards how to open up, for example, here the barrier to catch here the treasure chest that is labeled B. And now the task is I can now
16:09
Speaker A
simply as the red agent, the main agent watch the other agent interact to learn the best solution to learn knowledge and procedural knowledge from watching others other AI system or I yeah and now the question is how long should I watch the blue agent or
16:30
Speaker A
should I watch the green agent for a particular amount of time. What is here the criteria and what is the mathematical theory behind this?
16:39
Speaker A
As I told you, we will formulate it as a cost optimization problem. So let's have an example. Now our hypothesis is now hypothesis 1 2 and three is one is the goal A and the believes the wizard one
16:54
Speaker A
has the amolet. Our hypothesis two is we go for goal B and believes the wizard two has the amolet. And here the hypothesis three is we go for goal A and believe the wizard three has the amalolet for this
17:08
Speaker A
particular and you say oh great so but of course the observer me as the red agent I don't know exactly which hypothesis of my let's hopeful this is a complete set of hypothesis is correct no therefore in my understanding I can now
17:23
Speaker A
assign it as a starting condition a probability distribution to each hypothesis or a probability in itself let's make it easy Now as you noticed the unlaw code the preprint calls this hypothesis here in my beautiful pink color particles.
17:42
Speaker A
Yeah it comes from from physics but if you're not into the physics hey this is not a particle. This is not a physical object. This is just a name that they give maybe stay with hypothesis or substitute particle with hypothesis. It
17:55
Speaker A
has nothing to do with mechanics and so on. Now let's go here with probabilities.
18:02
Speaker A
Now we have posterior probability that you know and this is simply calculated here. So given everything I have observed, I have observed the H on J.
18:11
Speaker A
Whatever the H on J is doing up to a particular time T. How plausible is the mental state hypothesis theta I? So this is simply here a probability distribution. Then we need cost because this is cost. These are the elements
18:27
Speaker A
that we're going to optimize. Now, so we have Q obser observed this is the expected cost of watching the agent J for some number of step for some time T and then we act. So based off what we
18:42
Speaker A
learn just watching another agent we understand some complexity we understand ah there are some solution forward. So we say okay now I know how to act and now I start to act. Now I start as the red agent to play and I collect now all
18:59
Speaker A
the gold from the treasures. So computing this Q now requires this expected cost of watching the agent. It requires the observer me to imagine for each possible hypothesis about what the watch the agent wants and believes how that agent's future trajectory would
19:18
Speaker A
refine now the observer's own belief and therefore reduce the cost of the observer's eventual plan. So I might say my goodness. So what I have to do I have to imagine for each of our hypothesis about what the watched agent wants and
19:33
Speaker A
believes which is not a simple task even if you are an almighty AI then how that agent future trajectory so the projection here into the future would refine the observer's own belief systems if this is a machine okay wow in order
19:52
Speaker A
to reduce the cost so this is now a most important and formula by the orders. So here we have our Q observed and here you see we sum over all I of a particular formula. So what we have we have here C
20:08
Speaker A
ops this is here the per step cost of observing and you can define the cost I will show you the cost structure in a moment and then we have here a particular parameter TTI this is here if you want the observation horizon this
20:22
Speaker A
means how many more steps of watching it would take me before the agent's trajectory narrows the remaining wizard to one or before I see as the red agent ah now I understand the pattern Now I see the hypothesis that is driving
20:38
Speaker A
here the action of this particular agent that I watch. But of course we need something else. No we need another term and this is here C plan. This is here the exploration cost the observer me as a red agent would still face afterwards
20:55
Speaker A
when I understand the theoretical pattern. But now I have to go in the game and play and this also costs no.
21:03
Speaker A
So this C plan I is if you want the remaining self exploration and execution cost since I now participate in a game and I have to move and each movement costs some amount.
21:16
Speaker A
So what you see is we have more or less a term that acquires here the cost of acquiring the information and the cost of using this particular kind of inside this information that I learned by watching other agents in my multi- aent
21:30
Speaker A
system and then yeah you can combine this here. So we have the complete conditional cost in our QI and then we just sum it up. So we average those cost over all possible hidden minds and then we have here our master formula.
21:48
Speaker A
Beautiful. This is an example I always ask in my eye system. Hey wait a minute.
21:54
Speaker A
So let's play this a little bit that I get a feeling for the formula and I just insert very simple terms and you see here for the hypothesis what we get for the Q and for the probability distribution. You see simple terms I
22:06
Speaker A
have here my hypothesis. I just go with two hypothesis, probability, observation. And this provides me with some deeper insight because I'm a hands-on guy. If you are pure theoretical, abstract, beautiful. Do it in your your way. So this means the
22:21
Speaker A
control cycle now is simple. Now we ask the AI to simulate the future of all the other agents in our particular system.
22:28
Speaker A
Observe only one step how they behave. update here from their action what we think that they believe and what drives them. So the other actions are driven by their belief system and then we simulate we start a complete simulation for the
22:45
Speaker A
future again and this is why the author call it the rational mentalizing. Okay. So mentalizing means according to them the model goals and the beliefs rational and therefore we just compare the expected cost because this is a mathematical cost optimization. This is
23:02
Speaker A
the master formula and this formula is more or less just simple translated. Consider every plausible explanation of what this agent wants and believes the agent that we watch and for each explanation that we think could be theoretically an explanation predict how
23:20
Speaker A
long I as the red agent. I would need to watch them and how much work I would still need afterwards after I watched them for a particular time decided I have enough understanding I learned enough and now I can go and act in this
23:34
Speaker A
system with that total cost by how plausible the explanation is a probability distribution an average across all explanation and you know this is pure statistics so we have again just a statistical idea and we hope that this is somehow the methodology ology that
23:52
Speaker A
humans do here their social intelligence. Yeah. So if for every available agent J you have to do this. So the mal calculates the cheapest agent to observe and then it compares watching that agent with acting independently. It compares a
24:08
Speaker A
cost structure and it is simple. Please notice it is not about some complex elements from physics like information gained or entropy reduction. No, this is just valuous information only when that information lowers the downstream task cost. So the expected cost is then
24:28
Speaker A
compared simply hey what would it cost me as the red agent to stop watching here the blue agent and solve the problem myself jumping into the game right now with my knowledge that I learned just from monitoring the other
24:41
Speaker A
agents and then we have the complete algorithm one understood because we went through a each and every step. Now they did this here on multiple experiment for a particular reason just that you know experiment one was a single agent with a
24:55
Speaker A
single goal. So we have a red agent this is me watching here and a blue agent an expert NPC player with full knowledge of the amolet location who takes the optimal puff to its goal. So I learn here the golden puff. Yeah. And I have
25:11
Speaker A
only one treasure chest. Experiment two. Guess what? We have three treasure chests. Experiment three, we have we add another agent. We have multiple goals.
25:21
Speaker A
And now I have to decide, hey, wait a minute. Should I watch the green agent or should I watch the blue agent? What is here an expert? And this is now getting interesting. And then with experiment four, we have here some
25:33
Speaker A
experts or non-expert agent. We have multiple goals. And now really here we have the permutation manifold. That is interesting. As I told you, each action carries here a point cost. And the artist decided to go with a simple cost
25:47
Speaker A
allocation. Each movement costs three points per step. Each interacting with anything with a wizard costs here five points. Other costs less. And observing another agent cost one point per step.
26:03
Speaker A
So you see there's also the cost. If you do not do anything yourself, if you just stay passive and watch, you also have to it cost you something. Yeah. You have think about opportunity cost. If you would have chosen an active
26:17
Speaker A
participation in the game, you understand the optimization. Yeah, they implemented this in Julia. You have here all the the PEDDL structure, symbolic planners library, inverse planning library with particle filters, basian inference, everything is already available in Julian just
26:35
Speaker A
implemented it. Now, let's just be clear. The preprint's main purpose is not to discover whether AI performs better or worse than the humans. It is to identify this hidden computational rule that may explain how humans judge in their social interactions
26:54
Speaker A
because what we want is we want to learn our AI to be more socially compatible with humans.
27:04
Speaker A
Let's have a look at the results. Now the the main result was and since this is a human experiment let me stay a little bit more theoretical. The main result was that the theory of mind alone in itself fails because it recognizes
27:19
Speaker A
informative behavior but it does not properly price here the continued observation. If you just watch and never act you fail. Huh? The cost reasoning characteristic the Malc algorithm here alone also fails because the system cannot understand what a particular
27:36
Speaker A
agent action reveal about the hidden belief and goals because it is missing the theory of mind. So guess what? If you combine them well well then yeah you [clears throat] inferior agents mind plus you predict its future behavior
27:51
Speaker A
plus you calculate whether watching is worth it for each single time element and when to stop watching and when finally to start acting.
28:02
Speaker A
So during this experiment they found that human decisions about social learning this exact pattern that we want to learn in our AI machine are consistent now with a computational mechanism that what a surprise that combines here the theory of mind with a
28:20
Speaker A
value of information reasoning. So assigning cost and having a threshold in our costs. This is here finally the successful path they show you in figure three. If you have here for all the different models and ways and there are
28:36
Speaker A
different experiments colorcoded. Just notice that here the red line here is closest here to the human behavior in this full model that they propose. And you see if you do ablation and you eliminate parts of it, you see the red
28:51
Speaker A
line flattens out and is not at all like the human behavior. So what was the main question of the paper? Let's come back. It was hey as an AI agent or as a human who should I watch or how long should I watch this
29:08
Speaker A
other agent and when should I stop watching the other agent and discover the answer myself after I had some learning process internally and analyzing here the patterns that are recorded visually.
29:22
Speaker A
So if you want the the deeper relevance is therefore an intelligent agent should not merely consume other agent output without understanding what they doing, what is their task, what they present to you because this might not be the mathematical optimal
29:38
Speaker A
point. An intelligent agent should model why those outputs were produced by those other agent and finally determine whether observing that other agent is useful for its own objective.
29:55
Speaker A
So this means it is not okay to have a boss agent to have here an orchestration agent that just delegates job without explanation just saying hey you do this you do B you do C you do D because this
30:09
Speaker A
is not the best way to have the optimal performance of the complete system complexity we have that our little elementary intelligence they need to understand also what the other parts of the system are doing how they interact, how they
30:26
Speaker A
depend on each other. And it is not just here the classical supervising. We do achieve a better multi- aent routing in a heterogeneous team because the nominal strongest agent may not possess the most relevant information and this is here our orchestration agent
30:45
Speaker A
or a specific expert agent. As I told you, weaker agent mistakes or the search trajectory may reveal more about the hidden state because maybe the weaker agent makes a stupid move. But this stupid moves helped us to eliminate a
31:00
Speaker A
huge part of the search space because now we know for example that this particular visitor at this particular location on the grid has no information that is valuable for our particular task. So we can say whatever is here in
31:14
Speaker A
our search BS allocated to this particular wizard just forget about it. We have here the result by a weaker agent that did something maybe by mistake.
31:26
Speaker A
But let's take another step back. So what it is all about? Let's have let's look behind the court. Yeah, I have a feeling that this is somehow connected with an AI assistant behaving now or observing a human could
31:43
Speaker A
deduct something. Yeah, maybe this AI model, personal assistant, whatever you like to call it, might understand what the human really wants. Not what the human puts in or what the human provides on instruction, but what the human really wants. watching the human in its
31:59
Speaker A
natural habitat, in its natural environment, in its conversation, reading all the emails, reading all the communication. You got it. Then this AI personal assistant understands what the human knows. What is the education level? What is the complexity that the human still can work
32:18
Speaker A
on? What does the human falsely believes in? whether interrupting or asking a question is worth the cost. So this is something what I will call here a deeper understanding of a single individual of a single human that the eye is
32:37
Speaker A
interacting with. But if there is a formula, if there is a methodology, an advanced formula for an AI to really understand the belief system of humans and the goals of human, then we have a win-win situation. No, because yeah, we
32:53
Speaker A
can make it much more safer. No, we can really have a coherent answer that really integrates here the belief system of humans. No, but there are also risks.
33:03
Speaker A
An AI system that mods the human goals and understand the human beliefs in its particular interaction pattern can collaborate more effectively and provide better results to the human. And we just applaud this. This is gorgeous.
33:18
Speaker A
And we would never dare to think about that the same capacity can also be used for persuasion or strategic manipulation. that the eye system now really manipulates a human to a much stronger extent in a much more professional way in a much deep
33:38
Speaker A
obscological understanding was simply exploiting here false beliefs of this single individual being because the AI now understands much more about this non-mathematical emotional stuff that is driving and motivating human So suddenly this is about EI and the application of AI in social media and
34:03
Speaker A
suddenly I understand why those famous ivory tower universities are exploring this or discovering these new opportunities how to train AI to be much more compatible with humans with their emotional landscape with their non-logical landscapes and so on because
34:26
Speaker A
maybe there is a social media media platform that wants to sell a particular product, a particular service or some particular whatever to those humans. And guess what? If the eye really understand what this human person is, who this
34:41
Speaker A
person is, what are the emotional drivers, what are the goal of this person, you can manipulate here a conversation just unlimited.
34:52
Speaker A
Talking about limitations. So here we go. One final note. This was a human behavioral study. No, using simple computational agent in a controlled grid world. This is it. This was not a demonstration that a modern LLM can already infer arbitrary human belief
35:10
Speaker A
systems in the real world. This was just here. This would have been the next step that is not explicitly in the paper in the preprint. Yeah. Because the preprints if you want scientific proposal is nevertheless exactly what we
35:24
Speaker A
already identified. An intelligent AI observer could model what other agent want and believe either in a multi- aents AI system or in a AI human configuration predict what their future behavior could teach it and observe only while that social information remains
35:43
Speaker A
cheaper than an independent trial and error exploration. So this means that the experiment supports now the following statement. A predefined computational model combining now according to our insight of this new preprint the theory of mind with a value
36:00
Speaker A
of information reasoning. Now this complex system predicts whether whom and for how long humans or AI agent choose to observe and this prediction are substantially better than those AI models lacking either the component. But please be absolutely clear that this
36:19
Speaker A
preprint does not [clears throat] uh parenthesis yet parenthesis support the following statement. Hey, we trained an AI on 300 human socially intelligent pattern and behavioral styles and thereby created now a brand new EI machine that it is now socially an
36:38
Speaker A
intelligent agent and therefore so much bad is suited here to optimize the human behavior or influence the human behavior on all social media platforms.
36:50
Speaker A
I mean we never would do this now. So therefore here you have it. So just give you your personal reflection at some point here reading this paper I had the following question.
37:02
Speaker A
So wait so I said we have an AI agent okay that watches now other agents or other humans okay and just watching the action. This AI can now deduct the specific goals of the agent or of the humans and their belief system. Okay.
37:16
Speaker A
And then it can calculate I would mean it learned with a neural network if it is worth further observing this agent or this human in order to learn more from this watched behavior and compared to actually the self-performance of some
37:33
Speaker A
action and learn from its own actions. So this is a is a a soft equilibrium.
37:39
Speaker A
How long should I watch other agents or other humans to learn their behavior? And when should I start my own action, my own interaction as a human or as an agent with the environment to optimize my learning and be different and be here
37:54
Speaker A
of I don't know a better expert for a particular topic. Now the answer that I found and this is now outside of the study. This is just my personal answer.
38:04
Speaker A
How to answer this is the observer does not really deduce goals and beliefs with some certainty. Now think about we are still like in the next token prediction of a transform architecture here the encoder and decoder part of a transform
38:18
Speaker A
of a T5. We are still here in probability distribution. So whatever this intelligent social AI system now infers, it is a probability distribution over possible goals and belief systems, it is not yet able to really [clears throat] pin down here
38:36
Speaker A
this with some certainty. Yeah. Also the AI system does not merely ask whether another agent is generally more competent or not. it just estimate whether that particular agent future action that it will predict will reveal information that are relevant to the
38:52
Speaker A
observer's own goals. So I look out for a particular agent in the group of agents where I think that the agent action indicate this agent has a similar goal like I have. So I would like to say hey little agent you are my body let's
39:10
Speaker A
work together because we both have similar goals and maybe if you have a similar belief system if you have been trained on a similar ma mathematical methodology or even if you have been trained on some complimentary methodology let's form a team because
39:26
Speaker A
then we are more powerful and I think this system now continuously compares this watching with learning mode to its own costly exploration. How much would it cost if I stop watching and have my own exploration, my own trial and errors and it can and it might
39:46
Speaker A
encounter some horrible cost structures. So maybe it is better just to watch. I think this is interesting from a human point of view but also from an AI intelligence point of view because if you really want to have an interaction
39:59
Speaker A
process between human and AI system and you have want to have a better coherence between human and AI. Now I understand why Harvard and MIT is kind of exploring this particular interface compatibility on what they call an intelligent social
40:18
Speaker A
AI system. My goodness. Okay, so I made a video about a topic. I thought I would never have to make a video about the theory of mind. But okay, let's do this. If Harvard and MIT say this is worth
40:33
Speaker A
exploring, I have to learn it. I have at least to understand what it is all about. I hope I presented some insight to you here with this video. Have a look at the paper itself. There are a lot of
40:45
Speaker A
additional data. There are a lot of additional details I could not mention in this video. But I think this is now becoming more and more interesting because remember there's not AI to do the particular job in science. It is
40:59
Speaker A
also that a lot of corporation here deploy EI for a particular objective on their social media platform like to sell you something or to influence you or that you should vote for a particular person in the future. And the more we
41:17
Speaker A
have a a coherent interface between human and AI system where the AI system really understand all this irrational human behavior and it can find a mathematical optimization procedure to simulate this. I think then DI is really really on the edge of becoming something
41:37
Speaker A
we should watch out for. I hope to see you in my next video.
Topics:theory of mindmulti-agent systemssocial learningAI collaborationHarvardMITmachine learningtrust in AIcomputational modelingrational mentalizing

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries, speaker detection, and unlimited transcriptions.

App Store Google Play
Or transcribe another YouTube video here →

Free tools: TXT to SRT · SRT Validator · Merge SRT · Subtitle to Text · All tools

More from Discover AI

View All
Memory Now Inside the LLM: NEW AI Transformer METIS — Transcript43:14
Memory Now Inside the LLM: NEW AI Transformer METIS — Transcript
6,784 1 English
AI Changes Its Multi-Agent Topology During Inference — Transcript40:17
AI Changes Its Multi-Agent Topology During Inference — Transcript
7,687 1 English
The Death of Monolithic AI: The Scaffold & Harness Para… — Transcript35:03
The Death of Monolithic AI: The Scaffold & Harness Para… — Transcript
5,634 1 English
Can 3 “Soft Tokens” Replace 1,000 Reasoning Tokens in AI? — Transcript51:22
Can 3 “Soft Tokens” Replace 1,000 Reasoning Tokens in AI? — Transcript
8,338 1 English
Self-Learning AI Swarm Intelligence (New Code, RSI) — Transcript33:10
Self-Learning AI Swarm Intelligence (New Code, RSI) — Transcript
5,364 1 English
Next-Gen Self-Evolving AI Agents (ATDP) — Transcript36:37
Next-Gen Self-Evolving AI Agents (ATDP) — Transcript
5,666 1 English
Anthropic’s Big AI Design Change: The RED Pill — Transcript31:21
Anthropic’s Big AI Design Change: The RED Pill — Transcript
5,223 1 English
The Real AI Bottleneck? Graph Routing for Skills. — Transcript28:18
The Real AI Bottleneck? Graph Routing for Skills. — Transcript
4,520 1 English
Multi SKILL.MD Configurations: Self-learning AI — Transcript5:06
Multi SKILL.MD Configurations: Self-learning AI — Transcript
855 1 English
Signs of a Market Top — Transcript1:01
Signs of a Market Top — Transcript
226 1 English
15 Hobbies That Give You An Unfair Advantage — Transcript4:55
15 Hobbies That Give You An Unfair Advantage — Transcript
932 1 English
I Investigated A Failed $100 Billion Ghost City — Transcript21:47
I Investigated A Failed $100 Billion Ghost City — Transcript
3,754 1 English