Some Thoughts on Generative Machine Learning/Artificial Intelligenceε

 Since providing one's thoughts on this subject seem to be all the rage right now (August 2026), I might as well have a crack at it.

I started doing some research in machine learning (ML) applied to physical chemistry-based problems in 2021 as part of a postdoctoral stint in Prof. Pratyush Tiwary's lab at the University of Maryland, College Park because I could see how heavily my professional field (computational chemistry) was starting to lean in that direction. I didn't like ML then because it felt a bit like cheating -- instead of coming up with a physics-based theory to describe phenomena in a system (e.g. the "slow" eigenfunctions of the Markov transition matrix or transfer operator) folks started resorting to training very complicated (to me, at the time) neural networks to approximate these sometimes-complex functions. 

So, all the physics became buried in an inscrutable neural network, and there is still no consensus way to dig that information out, although it is currently (again, August 2026) at hot topic of research. But the point is, once there is enough data to train these networks (which there is for e.g. molecular dynamics [MD] simulations of systems whose slow dynamics occurs on the nanosecond to microsecond timescales [you can go up to milliseconds or maybe even seconds now if you have access to a state-of-the-art {in 2026} computing machine such as Anton 3 or use a highly coarse-grained {CG} model), in a lot of cases, if you just want to get something done (e.g. calculate an approximation of the committor function, a free-energy difference, or some other thermodynamic property), you can get a good-enough approximation using this type of ML technique. Even worse (or better, depending on the reader's perspective), the quantities calculated using these ML techniques (which are again data-driven and can by-pass theory completely) can be in even better agreement with the ground truth (experimental data) than a rigorous theory that has too many simplifications in it or misses the mark a bit with one or two assumptions. For me, this means, despite being trained in my PhD by a theoretical chemist, I basically do strictly computation now (although I didn't do that much theory as a PhD student anyway). 

But the point is this -- there has been a trend in physical chemistry (at least in my neck of the woods) where even quite rigorous theorists have shifted away towards developing theory and have instead pivoted towards getting nice results for more complicated systems (maybe -- there are still a lot of studies out there examining the dihedral transitions in alanine dipeptide). In many senses such a pivot is great since these results tend to give some insights into useful applied areas such as drug discovery, materials engineering, and general model creation. 

On the other hand, it seems (to me, at least) that the general quality of the scientific literature has gone down in the past five years and many people are willing to make small changes to an established ML model (very often a version of AlphaFold), show that this change marginally improves a cherry-picked metric of goodness, and submit a pre-print to a pre-print server, and this manuscript may never see the light of day as a peer-reviewed publication. 

Where am I going with this? The point is, many people, for the past year or so (again, I am typing this in 2026) are worried that generative models will make their jobs irrelevant because generative models are already technically proficient at the (electronic or digital) aspects of their jobs as the humans are. This leads to many economic concerns (e.g. what will happen when global unemployment rates reach X%? How will workers re-train when jobs have become more and more technically skilled over time? What will be left for humans to do now that generative ML models are as or more proficient than we are?). For me, I know that e.g. the current thinking version of Google's Gemini model is better than I am at writing code (it's the only significant thing I use it for, and I am still writing code -- I mostly use it debug and write some functions that I would find tedious and make many mistakes writing). For me, so far, it has made doing science a lot less fun and stressful because I know that I could be "more productive" (i.e. generate more data and result) if I used generative models more widely. 

But to me, this is the same issue I didn't find pleasant in 2021 -- it feels people have become focused on productivity (in the capitalistic, "how many widgets have you made today?" sense) and don't care as much about the process of doing something, in this case, scientific research. Now, one can argue that people still care about the process, but that the process now involves utilizing generative models effectively. And if that is how sometime wants to re-define the process of doing a task (such as scientific research), I think that is a valid viewpoint. But, to me, the analogy is buying a table from IKEA, taking it home, and assembling it, then inviting people over and telling them you made it from scratch. That's a lie -- I assembled it; I didn't make it from scratch.

And, to me, that is the fundamental change when sufficiently advanced tools are introduced to any task -- the process changes despite the initial and final states being roughly the same (let's say the reactant and product are sampled from the same distribution). Conceptually, I think of this as a function composition: let's say the initial state lives in a domain A and the task maps something in A (say, some pieces of wood and some screws) into a range B (say, a table of unspecified wobbliness). A minimally expressive function to perform this mapping is given by f, i.e. y = f(x), with x in A and y in B (I cannot find an equation editor here). Now, with the tool, the mapping is given by a function composition; now the mapping from A to B is performed by y = g(f(x)), where g maps f(x) to some space intermediate between A and B (let's say it maps the pre-table materials into a box filled with table legs, screws, and tabletop on a shelf at the Dybbølsbro IKEA). Now, the things in this space (let's call it A') more closely resembles the things in B (tables) than those in A (raw table materials). Because of this, the work required to make something in B from the required something in A' is less than that required going from A directly. And this is what tools have always done -- the classical simple machines literally generative leverage, and we have grown so used to this fact that we don't really complain about using pulleys to lift pianos to fifth floor apartment units. 

But the issue with the current generative models is that they are reducing the work required to do mental tasks. And, while the human brain is not a muscle, it must still be trained, first to learn material and skills and then retain them. Now, if we change the function mapping example above such that A is now a set of ideas (say, the the loose concept for a novel) and B is now the space where great literature lives, we can use generative models as a tool (or a not-so-simple machine, if you will) to go from idea space to novel space. And depending on how heavily we use it or how expressive a model we use, we can do a lot of work (develop all the characters, the plot, and write the entire novel ourselves, using only the generative model to e.g. proofread the final result) or we can do almost nothing (i.e. we prompt a model to "write the biggest, most beautiful novel about how great the United States of America in the style of Shakespeare and in less than 200 pages") and just make sure the output is semi-coherent (or just publish as-is -- such a prompt will generative a piece of literature that is simply tremendous and will be on everyone's best-seller list). 

For practical things, this reduction in work can be quite useful, especially in the medical field, where a lot of diagnosis involves heavy pattern matching and the recording keeping should be very, very high quality (meaning the training data for these models is quite good). However, here is a two-sided pinch. One, it means that doctors are perhaps both less valuable (even though they have spent many long years training for this role) and maybe less trustworthy ("ChatGPT says 'tis but a flesh wound!"). And the doctors lose the ability to practice and think critically because there is only a certain capacity required to make each diagnosis, e.g. less say it takes 3 work units to go from the space of symptoms to the sample in the diagnosis space known as "strep throat." Now, something like ChatGPT is currently good enough to perform that 3 units of work itself, meaning that the doctor need only do 0 units of work, can have Claude do the paperwork, then head home early. 

However, if less work needs to be done, then either 1) not as much training is required (so being a medical doctor will no longer be as high-profile a job [at least in the USA, where job status can be quite important]) or 2) the doctor will slowly become less sharp at these types of diagnoses because the brain is not being stimulated enough to retain this skill. Of course, this sort of atrophy as already been noted in humans physical bodies: I read a while back that those who practice distance running (e.g. marathoners) have the bone density comparable to those of our hunter-gatherer ancestors. This means that, those who do not perform an equivalent amount of load-bearing work have a much lower bone density (and likely physical strength, etc.) than humans had "back in the day." Of course, this leads to negative effects, e.g. bone diseases such as osteoporosis.

So, what this means is that we have historically seen changes in human body composition due to tool usage. So, the same can be expected with the heavy usage of generative models -- we will simply become worse at making things (for now, digitally) than we were before we resorted to using such tools. But, you might say, we can make better things now, because we have tools doing at least some of the work for us. And I will say, yes this is true, but only from the "productive" or starting-ending point perspective, but not necessarily from the process perspective. What I mean is this: if a doctor needed to make a diagnosis requiring 20 work units, but they are only capable, despite their training, of making a diagnosis of 15 work units, using a generative model as a "pulley" to perform the extra 5 units of work required to make a successful diagnosis is the proper use of the tool, and it is probably a good thing. But, not all doctors will do this -- some will not bother to strive for the 20 work unit diagnosis and will only look to diagnoses of 5 or 10 work units. These doctors will probably be out of a job as the hospital figures that they don't have to pay an AI company as much for a generative model that, on average, make such diagnoses with the same accuracy (and doesn't need a vacation allocation or sick days [well, servers sometimes crash...]). 

So, it seems like those who still strive to push their limits (e.g. the marathoners) will stay relevant in the their current occupations while some who don't strive (e.g. the sedentary) might get washed away. But, I think there are some remaining issues. First, if these models continue to improve (and become self-improving), then they will be able to perform all the mappings that humans are currently capable of. If it also becomes more efficient to chain (non-human) generative models to perform all tasks, then we seem out of luck, since it doesn't make much thermodynamic sense for humans to stay in the loop. Even if this is  not the case, I think a second issue is that, as is the case with bone density, the bulk of humanity could begin to suffer due to equivalent ailments of the mind; after all, the marathoners are only average bone density compared to our ancestors. In this case, those who would be considered extremely intelligent or clever or mentally competent in the future would be considered intellectually average today.

I think the question is, is that the outcome we want? As we have already decided that technology is good (I would agree technology has overall reduced human suffering), it will probably go that way. But, I would also argue that over-reliance on technology is not great for human wellbeing (e.g. car-centric societies; social media; advanced weaponry). But also, in many cases, the most advanced technology is not the best solution (it is never the most robust [linear models are good]), e.g. the fastest (and most efficient) way to get around most cities is via bicycle; analog communication (physically talking to someone) is always more efficient when the relevant parties are physically in close proximity (no need to sample a Poisson waiting time distribution); and writing a physical letter to someone in place of a text message or email will almost always elicit a stronger response from the recipient precisely because it signals that the writer utilized a higher number of personal work units to communicate a message in that way. And this is also a fundamental tenant of biology (ask Claude about the handicap principle): generally the more work we personally put into a project 1) the more self-satisfied we are and 2) the more impressed are peers are. So, those doctors who only put in 5 work units are probably on the way out, anyway.

Because of that handicap principle (the 5-work-unit doctors are self-selecting even without the use of more advanced technology [they may have more luck switching fields, but that is beyond the scope of this writing exercise]) and the finite resources inside our skulls, how much generative AI becomes adopted in the future will come down to 1) the strength of the handicap principle (i.e. how much do we value human work on an individual level [e.g. there are still human cashiers in addition to self-checkouts, and I think there will be a number of people who prefer human-driven taxis over Waymos]) and 2) how much policy choice at the government level favors humans over silicon-based workers (i.e. how much do we value human work on a collective or macroscopic level?). I think, when the dust settles (for now), we will end up in the cashier/self-checkout situation when there is still a high premium placed on human labor. Then, the question becomes how do economies tolerate an e.g. 20% unemployment rate? I don't know, but first we will probably see another reduction in the standard work week, finally down below 40 hours in the USA.

And, that brings me to the last point, which is that I don't think the emergence of generative AI models a first-order phase transition as people think it is. Automation has been a part of work and life for a while now, and it has pretty much been monotonically increasing as a function of time. And, digitally, it has been easy to generate things of dubious quality for a while now (Photoshop; lifting code from Stack Exchange; basically anything on social media). And history has also shown these to be self-correcting, e.g. forged results result in retracted papers, ruining careers and many studies have noted harmful effects of social media on people, resulting in legislation and lawsuits in multiple countries and talks of regulation in others. To me it seems most likely this will end up the case will generative AI -- it will be used, but not by everyone and not all the time. And it will likely become important to practice some things "the old fashioned way" to retain decent psychological and mental aptitude. And, once a hard-fought equilibrium is reached, us humans will have to learn how to deal with the androids from Planet 9.