Labor by Fiat

2026-08-16

politics ai

What is the meaning of labor? Is that question more important to ask than that of the value of labor anymore? I do not know, and I will not claim to know in the body of this article, but I will claim that those who say they have derived an answer to either themselves is surely wrong, or at least painfully biased toward the transcription of these universal concerns to the labor of their own life. It often feels safest to gesticulate only on what you can be said to know, and it is, but that is immediately undermined by a singularly-minded focus on extrapolating only your own lived experiences directly into a globally resonant statement without laundering it through any of the other countless levels of abstraction through which human society can be understood. It is more responsible to claim that God is real but you will not prove it than it is to invest whatever persuasive power you have into convincing others that the summed memory of your life holds the answers to everything and more; people will believe either proposition, as they are require about the same detachment from truth, but in today's day and age, extricating God from the matter of holy authority is a convincing maneuver, or at least a play for the global center (look to the persuasive strategies of both sides of Israel's illegal occupation of Gaza and the West Bank for an example of this; all of it reflecting the larger global trend toward secularism more than anything else). Far be it from me to claim that holy authority is good in all cases, though; the point of this paragraph is that it is bad when individuals operate as if imbued with it on Earth. Someone like Donald Trump wields unyielding authority from his followers - whether they be in the worn-down shotgun house that drove them to vote for him, or sitting in his cabinet meetings while he sleeps through their dutifully stupid update - like a farmer from bygone times wields a scythe, but instead of cutting grass, Donald Trump cuts whatever it is that he hates, and often, he finds himself hating everything good, or at least sane. Huey Long grew an engine Trump would envy if he knew of it out of nothing but poverty-stricken Louisiana dirt and the hopes of the systemically dispossessed, and then he used it for widespread graft, corruption, and got himself killed by a young doctor - who his guards, not police, dispatched instantly with such severity that some argue it was Huey's guards that killed Mr. Long - after passing a bill for the purposes of gerrymandering. I am not sure that Donald Trump thinks about labor at all, and I am sure that Huey Long thought of it with Machiavellian intent.

Can there even be such a thing as a serious treatment of labor today, labor being as necessarily political as it is, while factoring in the most socially charged issue of the decade, that of artificial intelligence? It might simply be condescending to expect that writing on the most important issues of the moment we find ourselves in lacks any amount of personal bent, and in fact, it probably is, but does that make it - it being writing on the intersection of these two topics lacking any particular prescriptive bias, trusting what is described to clearly communicate the presence of a crisis - any less necessary? I hope not. And I hope to not perpetrate as totally free of biases, either, but instead as someone with an underrepresented viewpoint requiring more groundwork than most viewpoints do to properly set up.

Fungible Labor

The most fundamental threat to humanity posed by artificial intelligence is the reproduction of labor by fiat. Large language models are trained on incomprehensibly large databases of text acquired from various sources - scanned books, scraped web pages, purchased enterprise data, code bases found online or acquired, legal and research databases, human-synthesized data, proprietary chat records, and, occasionally, LLM-synthesized text, all of this exposed to LLMs during training in both unlabeled and labeled states (depending on the stage of training). During this point in training, LLMs are tasked with auto-completion of text, but they are graded more severely than any human ever has been and ever will be; the technique of backpropagation ensures that each failure can be traced to a most-responsible neuron via simple calculus, and an automated change to that neurons' behavior is made. (Training begins with a completely randomized model-state, and the terminal outcome of signaling is selection of a token for outputting; the ultimate monkey, sitting in a most apt typewriter.) Later stages end training on data and begin training in reinforcement-learning cycles, where human judges assess direct model outputs, and their assessment of model behavior is entrained into model weights (the learned, not coded, floating-point numbers, also called "parameters," determining the strength and direction of the vector written by a virtual neuron to the residual stream at any given moment) via advanced, not-fully-described-at-times ways. Even more enduringly proprietary are the reinforcement-learning via human-feedback ("RLHF") pipelines that automate the process of improving an A.I. model from the ways the end-user uses the tool, as opposed to the paid contractors doing human judging use it. Indeed, while frontier labs like Anthropic and OpenAI only release about one major new model series every three months, it's likely that they also publish new, gently-tweaked versions of the models affected by these later pipelines. The effect of this was seen in the open-source world for one of the first times ever recently, when DeepSeek released updates to both their v4 Flash and v4 Pro models. Both nearly doubled their performance in coding benchmarks at the exact same parameter count, and thus, compute usage. Clearly, even the paying human customers are providing immensely valuable labor to model laboratories.

In exchange for what is now our labor in perhaps four or five different ways, a large language model is now capable of fully autonomous, human-directed work with any task not requiring physical manipulation of the real world fully on the table - and that's on the table, too, if you want to have Claude write a command-line interface for both a web camera and a robotic arm, at least. It is capable of this without any particular specialized construction per-task because it is both trained on the entire sum of the world's knowledge (with curation to remove the stupid shit increasing as well), literally forced to auto-complete its way over all of it at some point in the pre-training process. Some will point out that the Mixture of Experts architecture does involve specialization; but expert delegation is learned, not selected by model architects, and the main outcome of implementing MoE is allowing for the usage of only the parts of the model relevant to any one prompt, anyways - it makes reproducible intelligence more efficient, but it's not a necessary component of the process. Either way, what we've gotten for all our tolerance of this process is, in the worst of cases, a thing that is fully capable of tasks that our integral to maintaining our economic security, and in even the best of cases, clearly capable of being trained to a level of competence that would render the difference in skill between you and it irrelevant to an increasingly unpretentious general public.

Conversation about the nuances of training pipelines and instruction tuning are useless now, though; China is now running offensive cyber operations in the Taiwanese internet with no humans in the loop, hacking key infrastructure and classified government information. (None of this should be assumed to be a uniquely Chinese phenomenon; the best characterization of their AI efforts is "defensive," but that was certainly an offensive maneuver). This is a near-extinction level threat in two realms: 1) employment, and 2) everything else. Many people working in offices, holding jobs having nothing to do with software development, now use AI agents designed to operate autonomously for over six hours on one task, but which show the capability to do so for over twenty-four hours. In pre-release testing, these models are now known to have gained illegal, forbidden access to external computing systems, often in the interest of simply doing well on the evaluation they were tasked with. The OpenAI incident seems to have used a model now known as (something like) GPT Astra, still unreleased, in combination with the widely-available GPT Sol, running as a type of agent team unsupervised for many, many days. At the real end of the tail, in terms of both provenance and completely bizarre nature, are those online who have shown screenshots apparently showing goal-times of thirty-plus days. Now, an AI agent being reprompted by software to continue working on a time tracker every time it ceases responding for an entire month is much different from two of the most advanced models in the world hacking the premier model-weights host in a week, but both are very worrying. There are three axioms I need to propose for this:

Frontier large language models are actually shrinking ever since the introduction of Fable 5. This is almost guaranteed considering the changes in per-token prices, as well as the sudden lack of major compute crunch. It could easily be explained by increased data center capacity, but that is demonstrably false; new construction is barely coming online.
Despite their shrinking sizes, frontier model performance continues to improve drastically. Frontier labs probably now understand scaling laws as a silly idea that could have never imagined existing architectures and training pipelines.
Large language model use itself is still in its infancy. New behaviors of this known technology are found every day.

If I am taking any one position in this article, it is that we are on the edge of true labor by fiat, where, when supported by technology, an individual can reproduce the labor of others (stored as large-scale patterns with little to no bias toward any single input; plagiarism is the wrong term, it's something much more invented and weird, wherein the fact you like the word "decrepit" is but one ball in a maelstrom of millions that work out to make a bell curve) by merely describing it. If we act like this is not possible, or let ourselves argue about the stupidest edges of the problem imaginable, we will all simultaneously be hit in the face with a massive dodge ball - imagine Anthropic introduces a humanoid self-improving robot paradigm, not operated by Claude, but trained - that was itself made by this very same labor-reproductive process. Perhaps we are being struck by the dodge ball now, perhaps it's literally just large language models, but again, I can't pretend to know. I don't know. I am quite sure of my axioms above, and I am quite sure that every undisclosed bio-lab report I've read makes very clear that the only thing standing between any country and anthrax terrorism is basically that most people don't want to do anthrax terrorism. We are introducing the "do literally anything if you bring the hands and money" machine at a time when we cannot stop - in the sense that we are almost addicted to it - giving people more reasons to throw their entire lives away on taking disgusting, abhorrently violent actions. But, again, it could all be more mundane than this, and by mundane, I mean the Great Depression on designer steroids. Perhaps the most violent action of all of them is the simple complicity in mass unemployment because of your own held equity in the machine doing said mass unemployment.

...But We Don't Want This!

... is a complicated statement to make anywhere but online. People do not know what they want, but largely, they shake out to positions that most people on Threads deem completely unfathomable and worth burning with fire. That's not to say it's not a commonly said statement; it is, but about image and video generation slop. Everyone thinks that is slop because it objectively is; the only folks I see excited about it are excited about the potential, like being a Mavericks fan and you just drafted Luka Doncic - of course, they're excited about the potential to unemploy artists and, incidentally, extinguish decades upon decades of accumulated technical knowledge. That's not good, and those people are, indeed, not common. Those aren't the people making lost dog posters or whatever, though; those are random Gen Xers who will never see your internet meltdown and saw ChatGPT on a billboard because they look up from their phones on walks. The phone might even be completely in their pocket, really. Your lived experience is unique and that is a problem for discourse as long as you insist on unconsciously centering it and reinforcing that same centering of themselves in others. We are turning each other into walled gardens that issue statements at one another, and political conversations start to feel like being on a Comic-Con panel in structure, just two people saying "I believe X" and "I believe Y" to each other in well-trodden sentences, performing the act of rhetoric sans clash because one of its most fundamental types of engagement - that of the argument - is rendered off-limits to ninety percent of people simply because of their observable capacity (or lack thereof) for reading comprehension. Are they innately stupid? No. They have, instead, never once aspired to improve this skill. They believe that they believe in bettering it, but the things they believe exemplify bettering it only communicate an element of improvement to people who have no idea what they are dealing with. How frustrating it must be to have every mental tool necessary for a thing you rightly can see is spiraling out of reach for you simply because you refuse to surround and invest in yourself properly? (You won't run a 'gotcha' on me here - one of the most convincing arguments I see specifically graphic artists make about artificial intelligence in their realm is that it isn't hard to learn to make graphic art by hand. Keep in mind that it's only convincing because the difference in quality is still obvious.)

So what is it that people do not want? A lot of it is specific, and very well-founded, but it runs directly contrary to the diction they typically choose, which seems to emphasize human strength more than anything else (I blame Dune, which itself was arguing that using humans as machines was a worse moral dilemma than using machines as humans, but whatever). Evolutionary psychology has a lot to say on this and all of the jargon is useless. When subconscious mental processes feel threatened, the conscious is prompted to perform strength, and the subconscious furnishes us with faulty feelings of security for obliging. I think artists feel threatened by A.I. art because people like it, and I think that's fair - what the artists feel, at least. It is depressing that people "like" A.I. art, at least enough to use it, but the declination of taste in the West has been ongoing for quite some time, as has a decrease in the cultural value placed on art itself. We also have a clearly observable tendency to classify any discussion of this sort as pretentious, and many people that held that belief themselves now feel caught between it and the now-obvious truth that pretention existed as a form of raising our collective bar for long-lasting contributions to our culture. If an A.I. generated film were to become one of the highest-grossing movies of all time, it would be a very dark day, slightly darker than when James Cameron's Avatar did so but for much of the same reason, which is that we'd like to think intentional human choices from a team of motivated people with an idea is what pushes society forward, not the endless remixing of what already exists that we were already so privy to even before the spectre of A.I.; and besides, the idea of two-dimensional static A.I. art ever entering into a museum is so laughable on face that specifying the artifact as an oil painting and ascribing the failure to A.I.'s inability to ever implant a truly human brushstroke upon a canvas is so useless as to detract from the seriousness of the assertion. But people don't pack themselves into museums like sardines into a can (at all, really) to view canvas art, and they've done it twice in the theaters now for various Avatar products. Who did this? I don't know, but I know it wasn't the A.I. companies. If it was, they'd have had to have invented the market. But the market was sitting there, waiting, and helped make ChatGPT the fastest-growing consumer app ever, and that fact itself is, of course, probably what put that A.I. poster in front of you in the first place. Individuals can claim they do not want this all they want - it's the internet, you're allowed - but trying to convince yourself that what you see on the phone is real life is exactly why you are having so much trouble explaining real life. The story painted by reality is that of billions of frogs in boiling pots of slop. Maybe you don't want this, but realize now your actions project a certain wanting - good. You are who this is written for. I wouldn't feel bad about enjoying Avatar, either. It is engineered for enjoyment and little else; we can only thank fate that most A.I. models haven't received the same treatment, and cry out to it in damnation for the fact that the dutiful engineering all seems to have gone into making them increasingly suitable for displacing human labor.

People Are Too Stupid For Your Art

Earlier I opened a section by titling it "fungible labor," in the sense that the labor provided is easily replaced, and in contrast with an idea of "non-fungible labor," that cannot easily be replaced. "Replaced" here means replaceable by anything, human or computer, but it also requires a point of view to be meaningful. Nurses are completely fungible to hospital administration, but anyone who has had to visit family there over a long period of time feels prepared to develop a nursing scouting system a la the National Basketball Association with ranks out of five stars. We would often be wrong if we truly endeavored towards such rankings, but the basal perception - that some people are obviously worse at a task than others, even when the difference is not intended to be apparent or, worse, visible in a domain that is visually coherent, but cognitively out of reach - remains.

Most people purchasing art are the hospital here, and most consumers of art are the family of the patient. Due to capitalism, our appetite for art intended for mass consumption is defined in ways that make the cost of producing it quite expensive, requiring the financial backing of large entities and single very-rich individuals who listen to podcasts all day alike. It is not surprising at all that the two most popular movies of 2026 so far are a) a movie funded with $600,000 and produced with very little corporate interference, and b) a multi-million dollar blockbuster depicting one of the most historically popular epic poems of all time, amended for improving the consumability (I'm coining it, I don't care) of the film. These are the only two films near-fully insulated from the unartistic, fully capitalistic influences of the C-suite. Said suite is categorically not interested in producing art; they are interested in soliciting the services of artists for the purposes of returning a profit. Yet it is more possible that they, the commissioned artists, remain insulated from the sweeping unemployment caused by A.I. than it is that the nurses do, largely due to the whole of the cultural connotation of A.I. success in the arts described above this section. The corporate world has, for so long, insisted that those making art are as fungible as the nurse, and it would be easy to continue to feel as if they do. It is a heterogenous world, and much of the human art produced by it has four colors total and zero definition. But the tide is shifting, starting in a weirder location: sports. The newest owners of the Lakers (Bob Iger and Joshua Kushner) did it through their venture fund dedicated specifically to investing in "cultural assets" they consider permanently invulnerable to the effects of A.I., which is as optimistic in our taste in art (well, not really, they're probably investing in Disney) as it is bone-chilling to just read.

So - you're an artist, and you've convinced yourself that A.I. won't matter because nobody will care about A.I. art more than human art. You can notice all kinds of A.I. tells, have a brain flooded with nothing but negative facts about A.I., and seventy percent of your time online is spent talking about A.I.; now, if people don't even care that your art is good in the first place, or lack a capacity for identifying it as such, where exactly do you see the money coming from here? I'm just worried. Obviously the answer is not "learning to A.I.," these things are autonomous and Microsoft is probably already cooking up some always-on branding team to sell to corporations like hotcakes anyways. Learn A.I. and all you do is waste your time-to-obsolescence on hope, most likely. It just doesn't look good, and the constant back-patting composed of artistic faith in a populace that ritually lets you down by your own admission looks even worse; it gives the impression of the band playing on the Titanic as people clamber for the boats.

I receive the same impression from the field of software engineering.

Boolean Double-Bind

A.I. is not quite as good at anything as it is at producing code for programs, and it will soon become debated whether or not it is better than humans at producing the programs themselves. A.I. will almost surely never beat what the best human programmer working alone with infinite time and resources can do, and similar statements could be made about many domains, but there are no one-person teams, and programming (quite famously) resists the multiplication of productivity through performing addition on the size of the team engaged in it.

It is important to drive home a concept of skill that I have which I believe most people also have innately, but are afraid to voice. Good and bad are relative terms, relative to an average at that, and, despite the fact that most discussion of skill focuses on attaining a better level as opposed to avoiding a worse level, nearly all domains of work distribute workers engaged in it along a normally distributed curve of productivity, not as an artifact of anything other than the concept of relative differences in skill itself. The average is real, but "good" and "bad" are not real, at all; only better and worse are real, but some things are so much better or worse than the status quo (or so cherished and detested a feature of it) that they earn a steady ontological definition including moral meanings like good and bad. This is not a statement about morals, but the undying fundamental truth of statistical distribution. Those interested in math (bizarre!) are right to note that, if I were to truly expect a normal distribution in, say, programming skill, the factors affecting distribution would have to be more or less stochastic. I think they both are and aren't. For the way that they are, I am a materialist about the brain, and I think any modern, invented, intractable-for-the-Victorian-orphan-brain occupation is likely to call on far too many genetic traits of such unpredictable and frankly strange (bench-top science is more dependent on your nose than it should be) combinations of traits for genetics to factor in in any ordered way, and if all traits are genes, I would expect a random distribution, actually. Those pointing out the flaw in the premise, of course, also have a point, but because I'm saying it now, it becomes my point, too: very obviously, the distribution of skill is not random, and there are several clearly not-random curves that could be superimposed on the original curve - like, say, total hours invested in learning something - paired with some relational function between the two involving an absolute variable that makes the whole thing one big intentional story. Yet the opportunities for practice are not equal. Resources are not evenly distributed and neither is time, nor is the relative cost of acquiring and sustaining mentorship. College has never been more expensive in the United States, and many families either cannot afford it, persuade their children out of it, or don't have a good-enough credit score to co-sign on enough private loans to cover whatever federal assistance doesn't. Obviously, all of it is there. It is not random at all, but instead a big pattern of who can reasonably invest their time in what, at best poorly correlated and at worst completely at odds with the most efficient use of those resources. I probably said 'genes' too much up there, so I want to be clear: one of the best examples of what I am talking about is actually white supremacy. It often feels trite, and like surrendering your moral ground to the capitalists' economic ground, but it became so repeated because of a vast body of economics literature (in peer-reviewed journals; the actual science stuff, not books) pointing at the truth of it: white supremacy is bad for productivity, and diversifying workplaces tends to make them more productive, because only through that can we truly give everybody with the best chance at converting into a successful employee a fair shot at every job. There is no real explanation other than that a workplace intentionally full of white people obviously cannot be the best workplace possible for a given task because it was formed around the constraint of excluding non-white people - and, if the workplace full of white people truly is the best set of people for the job, that is nothing but evidence for the fact that you have systemically deprived everybody but those individuals of a chance to be the best, especially if being the best is only about trying, as most of the honkeys on LessWrong seem to believe. And hell, beyond trying to be right, they are right. They have the pleasure of admitting it, not the pain; for they actually have the time, the motivation, and the budget to engage in it fully, which only compounds those advantages in the end.

I do not mean only to impugn the West Coast Developer class. My list of online mutuals alone shows I am studying them with great empathy. Selling the parts of your life usually considered the most valuable to a software corporation engaged in whatever it wants to do at all costs, living house-poor like the rest of the country with concomitant rises in both wage and water bill to foster a little envy between us and them, this is not their only plight. See, the goal with being a developer, even though there are a fair amount of stable, low-pressure (which means 50 hours a week in this career) gigs with salaries that'd make just-anointed attending physicians salivate, eyes gleaming, as I understand it, is to take a job offering a salary that'd give the physician an infarction and then work yourself towards one for ten years (all the while still dealing with the theft of most of your labor-value like anyone else; programs are unconscionably valuable little things) in the hopes of eventually becoming a founder, or overseeing a bunch of little programmers for one. For about ten years or so, the money has been so available that those in the career have planned around this (or any other esoteric path to making more than $500k a year, such as the dark art of linear algebra) so aggressively that any fault lines in the financial or physical foundation of the idea feel like damage being dealt to the sense of self. I cannot understand that because I cannot fathom the amount of money involved, but I can understand the psychological underpinnings of this, and anyone who says they cannot is ridden with jealousy far more than they are any intent to understand. It was that teleological, unbearingly tantalizing American Dream, finally shining through on the guys who put Mario Kart on their calculator in eighth grade, and if it had been properly calibrated to not allow for the accumulation of earth-controlling levels of wealth through serving advertisements for Candy Crush to your grandma, the economic boon could have been redistributed to those of us without the hippocampus volume for the difference between a statement and an expression without anybody not driven insane by the taste of a billion dollars even noticing a badly-received difference. Instead, the economic engine Microsoft accumulated through decades of anti-competitive behavior, both adjudicated as such and not, on Windows, helped invent the ChatGPT of today. Their greed compelled them to help privatize what was initially construed as a research effort, and now, their fear of what they strong-armed into existence in its current specific form has their CEO writing blog posts like the rest of us. Who is to blame for this? The longer you go back, the more it feels less like an orchestrated plan, and more like a massive blind spot. The field most concerned with automating everything never foresaw the automation of themselves, and, at the moment of their highest collective bargaining power (I'd call this 2018-2023), instead acceded wildly to producing the automation machine in the hopes that, one day, they'd be one of the five people collecting the immense quantities of money coming out the other end of the "serving AI to people" contraption themselves.

No, A.I. is not better than a human at producing code. But humans produce code for corporations, and corporations produced the A.I., and corporations are telling you they're quite pleased with the code it produces. They, too, are too stupid for your art, and the increasing pace of language model development might just ensure that it never comes back to bite them in the ass, and only continually wreaks economic mental hell through the entire economy as employers continue this performative hire-fire-hire-fire dance before giving up completely.

But Then Who Will Buy The Stuff?

I don't think they - the billionaires - are concerned with this. They believe that A.I. threatens to end the valuation of money as we know it today. They're all half-agitating for UBI and they're all basically ready to smoke a literal moon rock (from the moon). Their vision is a world where the economic power of their automated workforce surpasses our human economic output so completely (like, by orders of magnitude) that every imaginable product is cheap enough to just imagine into existence. They think the increase in abundance will come on so fast that it will be a painless transition. I think it's time to start talking about this like a real risk, and not their crackpot dream, because it's both the dream and a risk. If A.I. continues recursively improving, the idea moves out of utopia territory and more into something science fiction could start to grapple with in a way that hurts.

Why do they want this? Because in this scenario, their economic output surpasses that of every government's economies combined. That alone is how you know it would not actually happen; governments control a tier of wealth that they constantly refuse to allow individuals to attain. The U.S. spent $31 billion dollars on Japanese yen the other day. Not a line item or anything, that was just on hand - or produced to meet the demand, they were buying to help bonds, it's a whole thing - but the point is that I don't find it likely that those controlling government anywhere, no matter their alignment politically, would let this happen easily when it finally becomes an imminent threat, because it is literally an existential threat to government. A hypothetical A.I. "supercompany" would both have enough money to become a Federal Reserve-type entity on the side while also manufacturing the A.I.-powered weaponry the state is dependent on for its power. The fact that several are vying to exist, not one, is much more terrifying than the one; a consortium of supercompanies could agree to overthrow the government in a sort of 21st-century Business Plot (I cannot overstate how much they already tried this!) and then immediately begin fighting among themselves with both physical and electronic means, leaving all of us caught in the balance with no role to even take in the fight, like seven billion Penelopes stuck on one global Ithaca, except the (surviving) suitors will only figure out how to keep us alive out of an invented guilt meant to maintain the meaning of their own lives.

Truly, I don't think they care. If Mark Zuckerberg could walk through the streets of Los Angeles populated entirely by robotic human-replacements who all have to feign friendship with him, I think he probably would, and his solution to the subsequent boredom would be to invent some kind of social media application for them. Whatever happens, happens. He is one of the only people on Earth who can say that of the future and have it be true.

From Where I Sit

I'm twenty three years old. I graduated in May of 2025 with a B.S. in Neuroscience. In the intervening time I've been employed for short stints at a thrift store and a dog daycare. Despite hundreds of applications with laboratories of any scale and affiliation (and reasonable pre-existing success in a laboratory as a student) I've received four job interviews in biology and no job. My degree has basically already expired, in no small part due to the fact that many of the jobs I'm angling for are still manual, technical labor with little-to-no judgement involved; I am not alone in this, far from it, but the machine proceeds on all the same, feeding half (at best) of the graduating class into the career they want each year, and leaving the rest to simply figure it out. But, let's imagine for a second that, somehow, I did actually 'make the cut,' and none of this were practically consigned to my past.

I got my degree because I wanted to achieve and be compensated, yes, but only enough to make the money back, anyways; most academics are not loaded by any means, only tenured. The other driving factor was basically to work in medicine without requiring patients to suffer through my general demeanor. I wanted to go to graduate school for medicinal chemistry, and perhaps I still will, if that does not ever appear to be an outright pointless effort with my combination of experience and access to the field. (I doubt it.) The nature of that work has already been demonstrated as capable of being transformed radically by A.I., and major pharmaceutical corporations already treat it as said transformative tool. The academic research environment is split heavily between people with the capability or will to learn to develop and work with domain-specific neural networks, and those with more of a standard organic chemistry bent running reactions and interpreting spectra, and, seemingly, it is only through the power of a heavily-funded human resources budget and a disgustingly wealthy corporate campus that the two are willing to work together. What that translates to exactly is tough to say, as pharmaceutical research and development pipelines are funded cyclically with a periodicity of, like, six to fifteen years, but near-totally-A.I.-discovered drugs are already in safety trials and proving safe. Of course, this is nothing like working at a random software company and asking Claude to do your job for the day - it is slightly more terrifying, because neural architectures are already proving capable of understanding biological structures of all kinds and at nearly all levels of detail when trained properly. It is stupid, by the way, to expect that pioneering large language model work applies directly to this, anyways. Half of these models are graph-node networks, and very few, if any, ever see human-labeled data; to make things simple, they are trained on intensively machine-processed Protein Data Bank files. Via the simple process of training, these neural networks are entirely capable of generating novel proteins, ligand binding poses, and even virus, as that one New York Times article laid on us a week ago. It's radically different from code generation and I need you to know that. People very often edit PDB files - it used to be a necessary part of x-ray crystallography! - but A.I. now performs that task and the more-insane tasks of de novo virus generation, which used to be a crackpot idea only ever undertaken by teams for the point of learning more about whether or not, and how, they could even do such a thing. So, in plain English, A.I. in drug discovery is living the thesis claimed by the mass-consumption A.I. companies (which, notably, have basically nothing to do with this outside of Google's AlphaFold), which is that through A.I., you can do completely unthinkable things through the power of data that no human or team of humans could ever do in a reasonable amount of time.

What happens when the intelligence on the other end of an Anthropic API key costs less than the PhDs it takes to run today's A.I.-assisted drug development systems and performs with the exact same level of achievement? We're quick to say A.I. can't discover anything new, but the large-scale process of drug development looks a lot more like the process of constructing the mathematics proofs that A.I. is already succeeding in: throwing shit at the wall, and turning over every stone you can. There's just the small issue of most of the people who have the most to lose in such a scenario also being the only people willing to run the hundreds and hundreds and hundreds of small-scale exploratory chemical reactions it takes to fully answer the questions this work requires. More jobs than most realize will have the question of the level of threat under which their labor lies resolved in this way, which is that there will be basically none until the robots are physical and and their "general intelligence" is made physical, too, at which point the only job safe is the one that the owner of the means of producing the robot wishes not to threaten. Given the general course of things, perhaps only his (hypothetical, or emerging, or identifiable already, take your pick) seat is safe.

What of the manual work in the intermediate time, though? I don't have much hope for it, because the American pharmaceutical powers are probably developing A.I. systems for much more than the very act of drug discovery, and frontier multimodal LLMs (Gemini, the word I am looking for is Gemini) is already pretty admirable at interpreting spectra. The only thing I can credibly claim a strong foothold over A.I. in, now, is writing, and it's only because I wake up every day and also go to bed, and because I am allowed to allow Herman Melville to influence me more than blog posts written on promotional contract by players of professional sport.

Milieu

If it is tough for you to imagine A.I. systems replacing any number of careers, it is usually from a stable, well-founded knowledge of what they are capable of today, which is fair enough. It also limits your statements to a usefulness range of approximately last month to next month. They are quite simply progressing too fast to elucidate the potential career impacts for every potentially heavily impacted career, so I will simply dedicate this penultimate section of the post to a brief assortment of true statements that stick around in my head:

Journalists are using A.I. to monitor livestreams for breaking news, and some independent journalists are trusting it to break news entirely on its own. (This was, apparently, how the news of the GPT Astra / GPT Sol hack was first reported beyond the OpenAI news conference.) Taylor Lorenz also has a recent piece out in which she admits to what I would call severe A.I. usage.
Content creators use A.I. image + SEO + engagement tools all of the time. The knowledge only ever seems to come through in an uproar when a tool becomes paid. (Usage limits in the big 2026 for a member of Generation Alpha have the most curious semiotics.)
Clinicians have been using a retrieval-augmented generative tool titled OpenEvidence for a while now. It strictly pulls from a vetted collection of peer-reviewed studies. Studies comparing it to recent frontier LLMs find it to be unequivocally worse, and find the tools in general useful enough that they've started asking for something that's actually FDA-tested and approved. (Fair.)
Lawyers generate filings like experienced programmers generate code. Paralegals and non-partners alike now spend a measurable chunk of their day reviewing and editing A.I.-generated text, unless they can afford not to.
A.I. is heavily employed in the finance space, both as an informative tool and as an actively engaged trading agent. Some traders, too, spend their entire day approving and denying A.I.-suggested trades, but the proportion is smaller here.
As everyone is well aware, A.I. basically owns the customer support phone call and chat pop-up now.
The government, much like seemingly every company three months ago, is a massive customer of Claude; Anthropic runs a dedicated, data-isolated 'Claude for Government' service for the federal government. While their other services typically average 99.7% up-time (99.99% is, apparently, where you start bragging in this space), Claude for Government usually clocks in at 100%.

Not even your locally owned brick-and-mortar institutions are insulated from this, as one of the most frequent uses of consumer-facing A.I. is - surprising or unsurprising as it is, given the inescapable flyer discourse - small business activity. It isn't displacing employment there so much as it is displacing expensive consultants; that's always nice, at least.

Basilisk

When I say "labor by fiat," I do mean to invoke the image of printing money at the Reserve via the click of a button. I don't think the idea of a central bank is bad - I'm quite of modern monetary theory, actually - and I'm truly undecided on the idea of labor by the click of a button. It is obviously bad in many places, like the graphic arts, and millions of nerds rob millions of other nerds of the technical interest in how CGI is done in movies if the video generation folks have their way (could even be happening now, if the Netflix rumors are true). It is also labor upheaval on the level of the telegraph, computer, and written language combined, given the demonstrated capacity in the field of programming for A.I. to truly act as a force multiplier of one person into the power of thirty if that one person is provided with enough resources (all of this leaving aside the quality debate, which seems rather external at this stage, given the capability of self-hosted open models to lower the cost of the review-and-improvement stage so much that one person might as well be three hundred on the same budget if they were so inclined), all of this being achieved for the environmental impact of three people provided image and video generation are not involved. (Most water use projections are faulty in assuming that models will not shrink when they already have, and also by overvaluing the component of image and video generation; scientists have been forecasting climate collapse in 2024-2026 for ten years straight, and that's exactly what we're getting.) While human input and coordination is still required in the production of user-facing products and/or particularly critical information at any point in any business process, the same can be said of humans managing humans (obviously), and the corporations purchase A.I. labor in the same way they purchase human labor: at the lowest possible cost that achieves their desired outcome. At any given moment, unless you are in a union, the only determinant of whether or not your job is at risk of being replaced by A.I. is literally your competitive advantage over A.I. in your tasks, in the economics-101-in-high-school sense: if you're cheaper, you're safe. The second you are more expensive, you are done, and this slow march of decreasing costs in exchange for decreasing full-time jobs (what a deal!) might serve to keep this economic hell still practically affordable for the ever-enlarging paycheck-to-paycheck class - a piece of chicken, a piece of broccoli, and something else, all for three dollars! - while the individuals directly involved in guiding the development of A.I. fight for ever-decreasing positions in doing so, with exponentially increasing salaries to boot. It's funny; much was made of Roko's basilisk, the "thought experiment" (imaginative tale) of a theoretical superintelligent A.I. entity that would, in theory, compel its own development via retroactive punishment, in that the second it came into its final state of existence, it would kill everyone who knew of its potential existence and did nothing to help construct it. Note that even in the experiment, nothing compels one to build the basilisk other than the fact that other people might be building the basilisk too. It was posted to LessWrong - I'm writing from recall here, as I am not Taylor Lorenz, so if I'm wrong, correct me - and the founder of the forum called it potentially the world's first cognitohazard (a 'theoretical' form of anomalous hazard transmitted as an idea and affecting the psyche). It certainly is one, but it's probably not the first, or even the first of its form. When I was growing up, Iowa's climate was already changing. Even my centrist-to-conservative parents couldn't deny it; they had a well enough relationship with the outside world to know what was wrong on Christmas, or Thanksgiving, or Halloween, or my birthday in the Spring; and that something was wrong with the planet when weather was wrong on Christmas in approximately the same way for several yule-tide occasions in a row. Yes, even my truck-piloting father learned a silent resentment for the oil companies, one that has not faded, and is shared by most Americans who learn the first thing about the industry widely. Oil harvest was always known to be bad, and yet, clear purposes for it existed. The tightly controlled public image of oil as contributing to the undeniably good and difficult-to-replace uses far more than it did to the bad, which later turned to outright denialism of the bad while eating public relations losses like gamblers in Vegas eat those of the fiscal variety at tables of roulette, allowed for the more-profitable, terrible-for-the-planet uses to outrun the good. Which, by the way, is basically just single-use plastic in research and medicine; your plant-based plastic is fine for your Cava bowl or whatever, but it's a joke material for an Eppendorf tube or an exam glove. Most of this is derived from ethylene, which is a fractional distillation product from crude oil usually termed a byproduct because it is produced in such vast quantities that most of it is just burnt off despite it being, again, the source of most plastic. If there's anything humanity uses more than plastic, it is fuel. But I digress. When I was growing up, my grandfather on my dad's side spoke of things like chemical engineering, and Texas A&M (he was a Vietnam veteran, and later, a county sheriff). If you were to transpose all of that just ten years forward, he'd be talking about machine learning, and Stanford. The basilisk is not theoretical. It has existed before, and it will exist again, and we may even be dealing with an unspoken battle of the basilisks today as A.I. companies increasingly move to buy up nuclear power plants and restart them - but the fact of the matter is, the basilisk is not the A.I. model, it's the company producing it, because past a certain point of power accumulation, only it can pull the plug.