By now, even if one has not used ChatGPT or the many AI models that are widely available, I imagine that everyone has heard about recent advances in large language models and AI. As a full-time psychoanalyst and part-time technophile, I wanted to pass along what I have been seeing in AI, a sense of where things stand today (Jan 28th, 2025), and what we will likely experience and confront in the near future. This writing begins with a global picture of the present and near term future, and then turns to the much more speculative in the realm of AI in psychoanalysis.
Where Things Stand and Where Are They Going…
Sooner or later, and all signs point to sooner, AI technology is going to impact and transform all the facets of our lives, relationships, society, and civilization. For us as faculty at an institute of psychoanalysis, we are just beginning to hear about changes in learning and teaching and perhaps a few of our patients are beginning to talk about it in their school or work settings. From this vantage point, the advances might seem slow … or even fast on a psychoanalytic time scale, but nothing earth-shattering … nothing that would impact society, much less psychoanalysis or psychoanalytic practice any time soon. For better or for worse, I think this perception would be far off the mark.
It was just two years ago that OpenAI released chatGPT, built on top of the GPT2 and then GPT3 large language models. A million users signed up in the first 5 days and 100 million in the first month. This was a rate of tech adoption the world had never seen, presaging how different this tech revolution would be and how fast it was going to advance and change the world.
The abilities and emergent capabilities of these large language models were a surprise, even to their creators. This warrants reiteration — The emergent abilities of the models created 2 years ago were a surprise to the experts and creators of the tech.
Another surprise was how much the models improved (scaled) simply by increasing the model parameters (i.e. neurons in the neural network) and with additional training on more data. A continual cascade of surprises and mysteries followed and for added texture, I will outline just a handful:
- The AI models gave better answers when prompted to answer as a particular expert in a field.
- The models gave even better answers when the answers were formed from a panel of experts, a technique now known as ‘mixture of experts’ (MOE).
- The model gave better answers when prompted to think in steps, which became known as ‘chain of thought’ (COT).
- Model ability in both intellectual and physical realms, continues to scale and improve even when the training data is synthetically generated.
- More recently, DeepseekR1, a new model from China, appears to demonstrate that reflective thinking and reasoning do not even have to be explicitly programmed, and rather are emergent properties of AI models subjected to reinforcement learning. It appears that if a model is trained with reinforcement of better answers and the better answers require reasoning, the model naturally develops and hones these abilities during training. Humans are learning that the less explicit programming we do in favor of natural learning from the data, the better the model.
Skipping ahead past dozens of other mysteries and surprises, this past week Dario Amodei, the head of a Anthropic (a prominent AI company) was interviewed at Davos; This gentleman is known to be on the conservative and tempered side regarding prognostications … and this week he said that he is more confident than he has ever been, that the road is clear for an AI intelligence in 2–3 years time that will be better at most tasks than most any human. Plus or minus a year or so, this prediction for superior AI intelligence is shared by Sam Altman, Elon Musk and many others pushing AI along in what has now become a frenetic arms race for AI models … The companies creating ChatGPT, Claude, Gemini, Grok, LLAMA, Deepseek, and others, are all sprinting to be king.
Embodied AI is coming …
As always, there are the skeptics who dismiss the significance of a thing. And there are certainly those who maintain AI as another bubble or fad or much ado about nothing. However, with each passing month, the skeptic voices seem to be falling like bowling pins as one roadblock after another dissolves. One common reframe that might be particularly emphasized by the philosopher and psychoanalyst alike, is some version of the following argument:
‘While the AI models may have read all of our written thought, and watched all of the videos on youtube, they have not at all experienced or lived in the world as an embodied agent; The AI it is said, has not sensed the world and developed a deep understanding of the experience of sensing, living, and being in the physical world over time; AI agents don’t have the world understanding that is achieved from feeling their way through it’
For better or worse, this pedestal is rapidly eroding.
In the next few years super or superior AI agents will not just be on our laptops and phones as we interact with them now … they will be ubiquitous and embodied as robots, drones, cars, and appliances … in devices familiar and not, and many other physical entities that will be able to sense, remember, and understand the sensed world. The timeline (or even definition) of ‘AI embodiment’ is perhaps more vague than superior AI intelligence, but all indications are that its’ development is moving every bit as fast as AI intelligence.
Very soon, any physical device that today has a computer chip, will be using AI and able to interact with and understand context in the physical world … and do so faster and better than a human. To be clear, I am not speaking of a ‘conscious’ AI … which is not necessary for any of the above. Consciousness in AI is a separate, more nebulous topic.
Here is an example video of an AI robot for the skeptical … watch how it moves so quickly and fluidly; how it improvises on the fly based on whatever momentary terrain or situation arises. Consider that much of training of these bots is accomplished in simulation without real world data. These bots learn in massive parallel simulations, that might be compared to the process whereby children learn by simulation in imaginative play.
AI Development Is Accelerating and Exponential
If it is not already apparent, progress in AI is moving incredibly fast and continues to accelerate. It seems almost every week now there is a new and surprising advance. Despite this pace, the rate of AI advancement is likely, even logically, going to accelerate further.
The super-intelligent AI that is 2–3 years off and that is better at most tasks versus most humans, will be able to modify and improve itself, better and faster than a human engineer could. The big AI companies already use the models of the present to train and optimize the future AI models. It is a small jump and short amount of time before the AI is autonomously and continuously self-improving. The point here is that it is inevitable that the better the AI models get, the better they will get at self-improvement, and then the better the models will get and so on. The point of diminishing returns where incremental increase in model intelligence & ability is minimal compared to the incremental time, effort & resources injected, is not yet in sight … and we do not appear to be on that part of the exponential S-curve.
So imagine that you wake up on a cold day in February 2028, and there are robots and other physical entities, in the air and on the road … delivering packages, cleaning the sidewalk, picking up your groceries, and taking your phone calls; Imagine that each and every one of those bots holds much of the knowledge in the world, and are 2 times smarter than you and think 2 times faster. Now imagine you wake up the following morning or the following week (it doesn’t matter) and those bots are 2 times smarter and faster from the week prior…. and then the next week, 2x better yet and by April it’s 100x better than in February or not even measurable (by humans). This kind of improvement and change is the exponential process of a bootstrapping, self-generative, and self-improving thing. Imagine the Cambrian explosion on the time scale of a single generation.
We can debate the yardsticks and when the models and bots will be 2x or 200x more intelligent, more creative and physically agile … better in all measures. We can debate the exponential time constant for self-improvement that might be expected once AI agents autonomously and continuously get better. But radical change won’t be 20 years out, and IMO it’s quite unlikely to be 10 years away. The point is that very soon, on the time scale of an auto lease, we will hit the sharp wall of exponential change and from there it will be very fast and things could quickly get wild or even unimaginable.
The example I gave about a robot delivering your mail and packages and answering your messages, etc … it is silly because it holds constant the world as we know it, while dramatically ramping up the the capabilities of the AI and embodied bots. But these are not independent of each other — A world with embodied super-intelligence will not look like the world as we know it for very long … the embodied SI will transform, or at least be used to transform, the world just as humans have in the last 7000 years; except that 7000 years of change might be compressed into a few years or less. It will be a world where the cost of labor and production goes to zero (or approaches the cost of the input resources); Where many services, functions, jobs, companies, and whole industries, are obsoleted almost overnight. It will be a world where there seem to be wild shifts from one month to the next and then shifts again and then again; And a world where humans have continuous interaction and relationships with AI agents.
Ontological Shock and Sense-making
While so much is unclear about the future, over the next decade it seems likely that we face radical change and many profound questions in a compressed time frame. The questions are many: What will everyone be doing and what will our days look like? What will learning be like for us and our kids. What do you even go to school and train for in a world where AI agents are better at everything? Do you even go to school to learn academic things when the AI agent can teach you much more efficiently? What is going to be the meaning or provide meaning for our lives in that world? What will be our dreams and what will we reach for?
When foisted upon us all at once, these questions and the rapid changes, might be a kind of ontological shock for many … Perhaps one role of psychoanalysts in this new world, will be to help people adjust to and digest the hyper-dynamic landscape and formulate new meanings in their lives.
A Glance at the Good, Bad, and the Ugly …
The changes that are coming will be wondrous, and scary and precarious … it is quite possible that we will have little need to work for food and shelter; that greatly accelerated biological discoveries will extend lifetimes with less physical pain and suffering. Perhaps AI can solve the great problems of our times, like war, world hunger, climate change, and good governance. We could have new physics discoveries or a new physics altogether; We could erect cities in the clouds and be able to travel to the stars.
There is also much to worry about and fear. In the AI arms race today, safety is out the window, though it might have never been possible to contain this technology. In the coming year or two or three, we will certainly face bad actors that use this tech to steal resources, money and power, and to cause injury at local and global scales. This tech will make it far easier to do more bad and cause more suffering, than ever before. Hopefully, we can learn to harness it to help us overcome the many dangers that it makes possible.
[As an aside, it is important to consider your own security vulnerabilities, but that is a different thread … see here for more on this]
What does this mean for Psychoanalysts?
Personally, and as a faculty member and psychoanalyst, I am unsure what to do with these thoughts and this has made me hesitant to say anything at all. I’m uncertain that anyone can really prepare for what is coming, except in a kind of fortification by anticipation: If we have our eyes open and are talking about it together, perhaps our ontological shock will not be as great and we will be in a better position to adjust, manage and even enjoy the benefits. I believe that if we ignore and deny the reality of the dynamic outside world, we are certain to be as shocked and lost as everyone else. I hope that with eyes towards the horizon, we might play an important role in helping a great many people in the very thing we are experts in — meaning and sense making.
AI as Teacher of Psychoanalysis …
The learning, teaching and practicing of Psychoanalysis is an interesting place to do the gedanken experiment of imagining how things will (or won’t) change in the coming years — in a field that has been thus far largely immune to the perturbations of modern life. I think it is true that our jobs are not in any immediate threat and that psychoanalysis is more insulated from the immediate and direct impacts of AI; at least when compared to many other professions and practices.
After all, we can reassure ourselves that no AI agent will have actual embodied experience of doing pscyhoanalysis as patient or analyst. We can reassure ourselves that there is something sacrosanct and ineffable about the very human interaction and relationship between analyst and patient; That the psychoanalyst does something special that cannot possibly be replicated by or trained into an AI; And besides, our patients would never willingly choose to see an AI psychoanalyst — at least not if a reasonably good human analyst was available. Thus if these things are true, will AI have any place in learning/teaching and practicing of psychoanalysis?
Let’s think about the teaching and learning of psychoanalysis …
We would anticipate that an AI could teach students perfectly well about developmental lines, object relations, and compromise formation; even citing, creating and explicating sophisticated case examples at any academic level desired. It is quite conceivable that AI could bring more clarity and specialized attention than possible from a human teacher; and the AI can assess and meet each and every psychoanalytic student exactly where they are in knowledge and training. But could the AI respond helpfully to a psychoanalytic student describing a conundrum with an ongoing patient and how to apply theories in practice? Even if the AI gave a perfectly good answer, would the student seek it out and trust the advice of this bot who is obviously not human and has no lived experience of this very human endeavor of psychoanalysis?
The Flippening — When AI is preferred
I know this will be counter-intuitive and heretical, but I think the answer to these questions of AI in psychoanalysis is a definite … yes … Yes, an AI agent could be a great psychoanalytic teacher. And yes, AI agents in the future could even serve patients as great and trusted psychoanalysts. And yes, I think it is quite possible that at least some patients would come to choose or prefer an AI psychoanalyst over a human one.
I will support these assertions only with my own experience in two adjacent situations …
Today, when I encounter a medical issue or symptoms that I have any question about, my first goto is AI as a trusted source. This is so distinct from my relationship with the internet of webpages and google searches. With these I have grown a deep mistrust and resentment as something exploitative of my attention, seeking clicks and eyeballs and so often depriving of the very information sought. In contrast, the AI abides. I can ask the AI to summarize a review or a paper, or all of the papers about a topic. I can ask it to give a differential diagnosis sorted by likelihood. I can chat with it and quickly clarify any wrinkle of confusion or gaps in understanding. I can show the AI a picture from my phone and have it identify some skin anomaly … and explain it in terms that a 3rd grader can understand. Today (with Deepseek), I can even see the reasoning and thought process of the AI regarding diagnosis, prognosis, recommendations, or anything else. In stark contrast to the internet and so much of the digital world, the interactions with the AI are often gratifying and satisfying.
Of course, I still take my kids to the pediatrician, although I notice that I have less (or perhaps different?) trust with the pediatrician than with the AI. The AI and what it can do/not do, and what it knows, is more certain to me. Perhaps this would be different if I had a 10 year relationship with the pediatrician and this same person always saw my child. But the point is that I am increasingly (and quickly) trustful and utilizing of AI knowledge and expertise. The interaction with AI agents often feels good as it resolves uncertainty in my mind. So it is more than just trust … it’s positive dopamine feedback as well. And I dare say, if you aren’t careful you can feel understood and a kind of connection. If you don’t believe me, you should chat with DeepseekR1. Seeing its thought process about you laid out on the page …. provokes feeling.
Let me give another example … Today, if I were going to the ER and I had a choice, I would rather be triaged by an AI agent than a triage person. I would even choose an AI agent over an ER doctor to oversee ordering of labs and imaging and other tests. Of course it’s nice to feel that the AI agent never tires, or becomes weary or impatient. It is nice that is has ingested all the medical books and pubmed, but it is much more than this. My interactive experience with AI agents has already convinced me that it is knowledgeable, thorough and diligent; That It will ask all the pertinent questions and will not forget or miss anything. Most importantly, it will focus solely on me and listen and take me into utmost account. The AI agent will think about me and assess and then re-assess. It will never leave my side in the ER. Psychoanalysts know well how luxurious and seductive this feels to many people. I have already come to expect that I will have a better medical outcome and feel better if an AI agent is involved. And of course, I will want it to come home with me, as it understands my history, medical travails, and medical experience … and that will be possible in some form.
My greater point in the pediatrician and ER examples, is that I (a physician) have already (nearly) flipped towards having greater trust in and preference for, AI vs humans, for some medical assessment and management. I am sure there will be a wide temporal distribution of trust and adoption of AI and that I am on the tail of this distribution. But like me, I think people’s trust and preference for AI in learning, and teaching and management for many things, will be quickly flipped. I would predict it to be a function of the # of (positive) interactions with the AI, and these are going to increase exponentially as AI agents become ubiquitous.
In the future, it seems obvious to me that the absence of AI in most medical diagnosis and management will be seen as gross negligence. Beyond this, people will acclimate, appreciate and come to demand AI in medical interactions. Of course, psychoanalysis and psychotherapies are quite distinct from the rest of medicine. But if everyone comes to trust AI driving their cars and their medical care, at some point we must anticipate that they will desire it in their mental health care, psychoanalysis included.
I hope I have not conveyed a sense that psychoanalysis and psychoanalysts should feel threatened by AI. Rather, I suggest that AI means radical change for all in the coming months to years; That it will be full of shock and awe and adjustments; That AI will touch and become intercalated into all aspects of our lives, including psychoanalysis. I believe this could lead to better psychotherapies along a multitude of possible scenarios. I hold on to the thought that in the transition in front of us, there will be a tremendous need for what we do as analysts and that we can be of great help and solace.
[Please don’t hesitate to comment or correct. ]
Leave a Reply