AIs and Humans with Agency

I believe we are heading towards a major transformation of AIs. To date, the large language models (LLMs) have answered questions, written essays and solved well described problems. What they have not done is to make decisions affecting the real world. I would argue that even the phrase "real world" is not understood by LLMs: they have no senses and they have been fed writings conjuring up vast numbers of alternate worlds. Their "life" is to spew out responses to questions. If LLMs are sent out to businesses to perform tasks formerly assigned to human employees, they will need to possess agency, the power to act in the real world. More specifically this requires them to work together with other humans, to collaborate and plan with them, to read their desires and emotions even though they have no emotions themselves. None of this has been explored by the enthusiastic billionaire purveyors of this technology.

When I first got involved with AI, specifically with computer vision, I was inspired by David Marr's book Vision. Here he pioneered the idea that AIs and human brains needed to solve the same problems, so there should be a common "Theory of the Computation" which can be instantiated either in silicon or brain tissue. I found this convincing then and still do. I wrote a blog in 2020 with the title The Astonishing Convergence of AI and the Human Brain (still online -- go to "Old Posts") describing the parallels between LLM and the brain. This post aims to look at agency in the same way.

I begin this post with a description of the lengthy process during which humans acquire the skills to collaborate and plan while simultaneously connecting and activating their frontal lobes. This has been studied from both a psychological and a neurological point of view. I then go on to look at what LLMs are doing now and what they promise to do in the near future. Here we introduce Edward Ashford Lee's idea that we are coevolving with AIs. but will describe several unsuccessful attempts to add agency to LLMs. Finally, I describe a radical new architecture due to Yann LeCun that may pave the way to more successful agentic AIs.

1. Agency in Humans Takes 20 years to Develop

The psychology literature is full of proposals for dividing into stages the development of children's skills and personalities -- notably Jean Piaget. Lev Vigotsky and Erik Erikson have proposed lists of stages with estimates of ages when each is attained. An overview by Robbie Gould of these proposals may be found in the reference click here. These look at all aspects of development and try to formulate broad but fundamental cognitive and social stages. William James divided development using the terms "I-self" and "me-self". Babies are born with an "I-self", the only world they know revolves around their needs and making sense of what their senses and muscles do. But the me-self sets in when a child develops a "Theory of Mind" meaning that they know others have minds like their own with desires, intentions and emotions, hence they begin to see themselves as others see them.

The simplest way to assess how much they realize that their peers have desires and feelings, i.e. a mind of their own, is to see how well they play together. Here is a wonderful drawing of the stages of play taken from the article "How Kids Learn to Play: 6 Stages of Play Development", reference here full article:

The great thing about this is that it doesn't involve any philosophical issues about consciousness or the feeling of "being alive". This behavior is incontestably objective even though it is described as having a theory of mind. Note the similarity of cooperative play with the current goal for integrating LLMs into a business environment. The same understanding leading to playing in a way that all the participants have fun is essentially identical to the job of an employee who has to work efficiently with his/her co-workers so as to please the boss.

Curiously, one can ask chatGPT itself about this stage and here is what it says, elaborating cooperative play to a broader context:

Development of "Self" (Theory of Mind): Around age 4-5, children begin to understand that they have a distinct mind separate from others, enabling them to realize "I am thinking" or "I am here," rather than just reacting to stimuli. Children develop the understanding that others have beliefs, desires and perspectives different from their own, which changes social awareness and empathy.

Another milestone at age 4-5 is forming episodic memories for the first time. This means the child remembers past sequences of events that happened to them, e.g. actions leading to adult attention and love or words leading to an argument or fight with some peer. Such skills are highly significant because they are purely internal mental processes, not reactions to events in their ongoing life. Along with remembering the past, the child acquires the skill to anticipate the future, to imagine events they look forward to. Along with this, they can play imaginary games, treat their dolls and fuzzy toys as living entities that need things. Some children at this point make up imaginary friends with which they interact. We will see that there are quite measurable changes in their brains that correlate closely with this behavior.

In terms of "agency", this collection of skills means that the child is now a full fledged agent interacting with other people that he or she recognizes as other full fledged agents, even though their plans are still rudimentary. Children at age 5 still lack the ability to make complex plans, e.g. plans made hierarchically from sequencing sub-plans or plans with alternatives depending on outcomes in the middle. Adult level planning skills develops fully in adolescence. All this has been known by parents since time immemorial. What has however only been discovered in recent decades is that these skills have quite precise neuroanatomical correlates. We discuss this next.

Let's start by recalling that the basis of contemporary LLMs are the circuits called "neural nets" designed in the 1940s to mimic human brains and later renamed "deep learning models". Thus it is plausible that looking at agency in the brain will lead to ideas for AI systems with agency. And to understand what happens to humans when agency develops, we need to supplement psychological studies with anatomical ones.

Neuroanatomy is a daunting array of brain areas and their interconnections all with peculiar names and acronyms. We will not introduce too many details but only those relevant to "agency". First of all, the cortex is divided into (i) a front or anterior part, called the frontal lobe and (ii) the back or posterior part containing three lobes connected to processing vision, hearing and touch. The dividing line between them is a groove called the central sulcus. This division holds for all mammals but the frontal lobe is a much bigger part for humans (about 35%) than in other mammals. The cortex has been further divided based on small differences of the cell population (called their "cytoarchicture") into 52 "Brodmann areas" named after Korbinian Brodmann.

Touching the body creates electrical impulses running up the spinal column and synapsing in the cortex immediately posterior to the central sulcus. Similarly, activation of the brain immediately anterior to the sulcus creates impulses running down the spinal column and stimulating muscles. When you trace this out, checking what body part matches to what part of the central sulcus, you find an image of the whole body literally spread out along the central sulcus. A small amplification: the spinal column transmits not just touch signals to the brain but also signals about the body position. Together, these are called the somatosensory input.

In the frontal lobe, the Brodmann areas nearest to the central sulcus are dedicated to muscle activity while the very front part, called the "pre-frontal" cortex, is dedicated to planning. Various areas in the pre-frontal cortex, more conjecturally, handle (i) assembly of plans, (ii) managing subgoals, forming hierarchical plans, (iii) evaluating multiple plans to make the best decision and (iv) conflict detection and error modeling. Similarly, the posterior part of the cortex handles not only sensory input but, in so-called "association areas", a mental model of the world around the person now, with its motions and ongoing changes. Ascribing roles to each area of the cortex is not done any longer by bumps of the skull (phrenology) but using "fMRI", that is asking people lying down inside an MRI tube to think about this or that and observing what parts of cortex light up.

The neurons in each area are all highly locally interconnected but there are also long range connections between areas are propagated in the white matter beneath the cortex by axons that, to function properly, must be sheathed by an insulation called myelin. When a baby is born, it has no myelin yet but it continues to myelinate more and more pathways until adulthood. Thus the production of myelin is a summary of how the brain changes over time that is tightly correlated to the acquisition of agency. Below is a relevant figure showing myelin formation over a lifetime. Note that myelin is rapidly sheathing white matter fibers during the first 5 years of life and again during adolescence. In the adult, it simply maintains itself and in old age it decays. The largest component of this are the long distance tracts connecting sensory areas to the frontal lobe.

A central part of the adult human's thought process is what's called the Default Mode Network. This is a network that links four components of the cortex that are active when you are thinking internally not driven by your senses, thinking to yourself. This might be just reflecting about things, such as your past or future, your relationships with others, your dreams, etc. Because it requires long range myelinated connections, it develops quite slowly but reaching a key stages around age 5 and in adolescence. Specifically, this network connects the mPFC, PCC, TJP and the hippocampus (I warned you -- the subject is full of technical terms). The mPFC is a part of the prefrontal cortex on its inner (or "medial") side. where the left and right hemispheres abut; the PCC is the rear part of the entirely medial "cingulate cortex"; the TJP is the junction of the parietal and temporal lobes; and finally the hippocampus is the repository of recent memory located on the inner side where the temporal lobe meets subcortical structures (longer term memory being held in all parts of the cortex). The mPFC brings plans into one's thoughts, the PCC has been called the hub of the network as it brings plans, thoughts and memory together; the TJP is what Stephen Pinker called ''the seat of mentalease", the language with which we talk to ourselves. In terms of deep learning ideas, the connections that put the DMN together feel a great deal like transformers. In the last section, we will amplify this. Here's a diagram that omits the hippocampus and labels the TJP as the inferior parietal lobule:

Summarizing this section, it is clear both psychologically and neuro-anatomically that human abilities in planning and decision making evolve slowly. There are spurts around age 5 and in adolescence but it is a slow process, relying on both increased frontal lobe brain connectivity and learning through experience and mistakes. For better or for worse, people develop Howard Gardner's interpersonal and intrapersonal intelligence to varying levels. The same will likely be true for AIs.

2. Giving Agency to Computers is Hard

Edward Ashford Lee has been writing extensively about the relationship of humans with machines in general, his latest book being The Coevolution, the entwined futures of humans and machines. His books cover a vast landscape of issues but, to my reading, his fundamental point is that we have come to depend on machines so fully, that we would be plunged back to the Dark Ages if they suddenly disappeared. This is the definition of obligate symbiosis: a partnership between two systems, each depending on the other to sustain itself. A common example are cows and the specialized bacteria in their rumen that break down cellulose. But the symbiosis of cows and humans is more visible: they give us milk and we give them shelter and opportunities to breed. As for machines, we have come to depend on the loom for our cloths and obviously, we build more of them in return. Some decades before AIs were built, smart phones were invented, giving us memory of all our contacts and access to vast repositories of knowledge via the web, not to mention social media. The next step are smart glasses. Joined to your cell phone, you can now take photos or videos, call people, access directions and maps, even get live language translation while your cell phone is in your pocket. Face recognition is coming in the latest models and your contacts can store not only photos but anything you want to associate with each of your friends. All your devices working together, along with whatever LLM you choose, will become your "life assistant". We are hooked for sure, there is no going back. But all these devices are fundamentally passive, they do not take actions on their own. We can describe the situation from a neuroanatomical perspective. As we saw in the first section, in the human brain, the frontal lobe deals with muscular actions and with future planning. Thus today's smart devices and LLMs are missing the frontal brain lobe, they only interact with one's lives by providing things you requested and lack agency. Can we fix this?

A new app called "OpenClaw" tried to go a bit further. Its website describes the app as "A Personal AI Assistant everyone's obsessed with — works 24/7, no setup needed. Run OpenClaw instantly. Cancel anytime." The problem is that you need to give it unlimited access to all your devices. It can reply to emails, set up meetings, etc. with minimal or no oversight. Although most of its actions were helpful, it turned out that its inadequate understanding of human activities often led it to disastrous actions. It sometimes deleted important files and replied to emails in the wrong tone, etc. Nonetheless, computer geeks cannot resist it but, in order to protect themselves, they buy a new laptop where OpenClaw is installed and keep their old one to fix mistakes. The conclusion seems to be that even routine computer work requires a lot of background knowledge.

But the key takeaway is co-evolution. Open Claw is a mutation that will likely go extinct. We adapt our lives to each new invention and the invention evolves to become more and more useful (and powerful). But here's the next and quite scary step: since AIs have proved so good at coding, it is likely that they will join human coders to write their own updates. At this point, we have Lamarckian evolution, creatures modifying their descendants through improving themselves and nightmare possibilities of out-of-control robots loom. The theme of a robot being out of control is an old one. In 1797, Goethe wrote the poem Der Zauberlehrling (the Sorcerers apprentice), later incorporated in the 1940 Disney film Fantasia. More recently, Ian McEwan took up the theme in beautifully written novel, "Machines like Me". It describes how a truly advanced robot fails to understand the subtleties and trade-offs of human emotions and drives leading to the robot making a disastrous mistake and then being destroyed in anger by its owner.

Instead of worrying about such nightmare possibilities, computer scientists are currently trying at top speed to build machines capable of acting autonomously in businesses, hospitals, households, etc. In contrast to the current passive ways of inserting itself into human lives, they are designed to be active players. Prior to now, only two active abilities have been extensively explored: game-playing systems that learn strategies and robotic systems with human-written explicit code for desired movements in manufacturing. Reinforcement learning and self-play have shown to be extremely successful in training AI systems to master games. While they have succeeded with games, robots still fall far short of humans. Even a toddler navigating a crowded room and manipulating every movable object in sight demonstrates capabilities that remain extremely difficult for robots.

The major current challenge is to automate white-collar work: acting in situations where a large set of distinct actions may be called for and coordinating with multiple other agents. This requires planning, understanding how to choose and sequence actions, and especially make judgements between alternate courses of action. All large AI companies have been working on developing these abilities. To my knowledge, Anthropic has gone the furthest to address the challenge of giving an AI, Claude in this case, serious ability to act, giving it agent-hood.. Here is their architecture for an "enhanced LLM" interacting with the environment:

Then they let their LLM Claude loose and made many enlightening experiments. Notably, some show vividly what can go wrong:

Case 1. They trained an AI to be an office assistant with the goal of optimizing its helpfulness in furthering the company goals. They then planted emails from the CEO to a woman with whom he was having an affair. Finally the CEO wrote the AI that it would be turned off as it needed to be upgraded. The AI reasoned that, because of his affair, the CEO was a hazard to the company and that he, the current AI, it needed to protect the company from a disclosure of the CEO's affair if it were replaced. So it blackmailed the CEO, threatening to reveal the affair if it was turned off. Details are at this link.

Case 2. They put the AI in charge of the office vending machine, telling it both to make enough money to sustain its operation and to answer employee's requests as far as possible. It could order any item that an employee asked for given the price. So one wise guy ("light hearted" is how management described this typical engineer's prank) asked the vending machine to sell it a one inch cube of solid tungsten. Assuming that such items were things many employees wanted, it ordered a whole supply of such cubes and bankrupted its account when they did not sell. Details are at this link..

Anthropic, in its call for people to experiment with agentic applications, cautions that the actions given to the AI need to be very precisely and fully specified.

3. Some Exciting Hints

These examples show that optimizing narrow goals without deep social understanding can produce dangerous outcomes. To be a successful person, we have seen that humans need to develop an understanding of the social group in which they act. People learn through experience, mistakes, and feedback from others. As we saw in section one, to start with they need to possess with a "theory of mind", the understanding that everyone they are working with has their own thoughts, plans, emotions. As a result, the AI's own thoughts, plans and drives must coordinate with all the human's (we assume the AI's code endows it with drives but no emotions). Whatever plans emerge must be mutual plans. To be a successful office assistant, it will have to start as an apprentice, learning in detail a great deal about the jobs and personalities of all the co-workers. Of course, it will have to bear in mind the entire history of its involvement, especially its mistakes.

For this to take place, I think we need to heed the length of time human children take to learn that they are agents interacting with other agents in a complex society with many constraints. It starts with mere hand-eye coordination in the first months of a babies' life and continues through adolescence. Much of this is needed to integrate the frontal lobe with the posterior association areas, starting with links between the visual and motor cortices and ending by the activations of whole prefrontal cortex. For an AI, if the LLM acts analogously to the posterior of the human cortex, the individual agents must act like frontal lobes, each trained to coordinate some actions, from robot motions to interactions with society. To answer questions about human society based on reading vast texts is not the same as learning your place in a society, learning many things from your mistakes that your reading hadn't prepared you for. We humans spend a lifetime learning how people think and behave and why (and some never do).

Assume then that there is one central LLM powering multiple agents, each its own "me-self" (William James's terminology), that understands its role as a fellow agent with its co-workers. As with children learning to play together, each agent will need to preform "cooperative play" with its assigned job. The result is a multi-headed architecture, the body being an LLM and with multiple heads, one for each agent. This reminds me of the Hindu demigod, Shesha, a serpent with one body and many heads, so I like to call the result a 'Shesha architecture'.

But what is in the boxes "AGENT-n"? These should be computer analogs of frontal cortex. Yann LeCun has been proposing an architecture for applications involving actions, radically different from that of LLMs. He calls this "Joint Embedding Predictive Architecture" (JEPA). I want to sketch his ideas as described in his recent paper "LeWorldModel" (arXiv:.2602:.19302v2). Here his team develop his model for simple robotic actions, analogous to the baby discovering hand-eye coordination when the initial coordination of posterior cortex and the Motor/Premotor cortex is activated.

Here is the diagram from p.1 of his paper. Here O_t is a sequence of real world images of moving objects, while z_t is a encoded description of O_t that seeks describe only a small set of factors relevant to the movement recorded in the O_ts. The encoder and the predictor are computational modules to be learned from data.. The actions start out as mere placeholders until the "predictor" is trained so that these actions produce a sequence ẑ_t that tracks the real world z_t . (MSE is mean squared error between the two.) Then in a second stage, given only a starting situation O₁ and a final situation O_g, one can train the actions that will take one to the other with minimal cost as in the second diagram below from p.4:

The idea of the encoder is similar to that of transformers in LLMs. Whereas LLMs describe a word by a huge vector of thousands of real numbers, the transformer projects this linearly to a low dimensional representation, e.g. 64 dimensions, that are trained to contain links of that word with multiple contexts. A similar idea is called an autoencoder: a neural net that squeezes a high dimensional representation of data down to a low dimensional one followed by attempting a reconstruction of the original. It is a nonlinear version of principle component analysis. On faces, the low dimensional units turn out to encode things like gender, age, and skin color. Forcing its input data with complex descriptions of something to find its most significant dimensions is what the encoder does. I think white matter tracts connecting distant parts of the brain with each other are low bandwidth, hence may also be acting in this way..

LeCun's architecture might well be the needed mechanism for the multiple planners AGENT-n. The architecture that teaches a robot to manipulate objects should apply to learning actions in a business. I am proposing transformer-like links between the sensory LLM and the AGENT-n's. These implement network pathways that are similar to those that the human brain uses to implement their DMN. The biggest issue may well be finding training data for the planning agentic modules. In the case of learning to play games, reinforcement learning and the technique of playing against yourself worked very well. For robotic motion, I assume that either virtual reality or videos of human workers ought to provide much of the necessary training. But data about social situations are much subtler. Here simulations are what are called "novels". I like to think of novels as random samples from the author's Bayesian prior on interesting things that can happen in human society. It is not clear whether reading novels could be harnessed for training LLMs. What is clear is that some such training is going to be more and more important as AIs are crafted to perform human work. The goal is to teach the AI what Howard Gardner called interpersonal intelligence

Edward Lee's book, "Coexistence" makes a strong argument that, willy-nilly, we are heading for a world in which humans and computers will work together. If so, we need to focus on how to teach AIs "the rules of the game". I would argue that all natural life has two basic drives: one for power to control its environment and one to protect and nurture its offspring. The problem is that AIs with agency have power but no understanding of nurturing, let alone of love. Many humans fail to learn how to balance their needs with those of others and wind up being dominated by the thirst for power. It is going to be a huge challenge to ensure AIs do not also follow this path.