19 responses to “AI developer warns the AI jobs apocalypse is closer than we realize”

1. Mark

   February 12, 2026 at 7:19 am

   I think we need to be careful about the timeline. I could be wrong, but I find it more plausible that this scenario plays out over 10-20 years than over 1-3 years. My reasoning is as follows:

   First, pretraining scaling has clearly tapped out. This was the method that drove performance improvements up to around GPT-4o: scaling up the amount of data and compute used to pretrain the models. These methods work by masked text prediction: given the beginning of a sentence, predict the rest of the sentence, with the sentences drawn from the internet. Up to around 2024, you could draw a simple power law curve relating amount of training data, amount of compute, and performance on various benchmarks. If you then extrapolate that curve based on what you expect to have in a year or two, it was possible to predict that simply extending this approach would lead to AGI shortly. However, the curve subsequently flattened sharply, as we can see from what a dud GPT-4.5 was. Given that human beings are often illogical, factually incorrect, or just plain stupid, it should not surprise us that, at some point, improving the model’s ability to imitate humans no longer improves performance on what we care about.

   The new paradigm has been about reinforcement learning (RL): give the model a problem, have it try a couple of times to solve it, then reinforce the best solution it comes up with. To do this, you need be able to evaluate the solutions. This is done either through human feedback (RLHF) or, for tasks where this is possible, automatic verification (RLVR). Initially, there was widespread expectation that this would unlock a new scaling regime, especially by combining RL with chain-of-thought and inference scaling. RL has clearly improved models significantly, but there is good reason to think that we are close to the limits of what it can do:

   First, there is significant evidence – though not incontrovertible – that RL only “sharpens” the output of the LLM. If, out of ten tries, the LLM *ever* gets the right answer, RL will ensure it *always* gets the right answer. But if the LLM *never* gets it right in ten tries, RL won’t help.

   Second, RLHF faces fundamental limits in scaling, because it needs humans to evaluate how the model did. This has a whole host of logistical challenges, especially when the problems are complex or difficult to solve. And RLVR only works on problems which can be automatically verified, which primarily means programming and math.

   So it’s very possible that I’m wrong, but my median expectation is that we are already past the exponential part of the logistic curve, and we’re now facing a period of diminishing returns in LLM performance, especially on tasks that cannot be automatically verified, like legal reasoning. Furthermore, for a company like OpenAI, their business case only makes sense if they can achieve AGI *soon*. So I would not be surprised if we see a period of retrenchment, analogous to the dot-com crash in 2000. The internet didn’t go away, but pets.com did.

   However, that does not mean performance will not improve, just that it will improve more slowly. Given the value of an LLM that can genuinely substitute for a lawyer, companies will keep working at it. With enough data and enough human feedback, I don’t see any reason they won’t succeed, I just think it will take a while.

   There will probably remain some tasks that will never be feasible for LLMs, given the limitations of the architecture. Extrapolation – generating genuinely out-of-distribution results – is the most obvious one. But even if I’m right about that, how many jobs need to create genuinely novel ideas, as opposed to remixing existing ones? Even among scientists, I suspect the answer is “not very many”. So, as a member in good standing, I’m very concerned about the fate of the professional middle class over the next few decades…

   Loading…

   Reply
   1. Ian Jensen

      February 12, 2026 at 11:08 am

      This more measured view seems sensible to me as well and is supported at least somewhat here: https://arxiv.org/pdf/2602.06176

      I’m no AI expert and not qualified to discuss “AGI” or whether LLMs reason or not, but the taxonomy of persistent failures in the link suggests to me that we aren’t seeing tremendous improvements in what the authors call “fundamental” failures–failures caused by the way LLMs work. Indeed, much of the work to address these failures seems to me more like workarounds and patches than it does addressing architectural (if the metaphor can be excused) issues with LLMs. And this is not for lack of trying!

      IMO, improvements will continue but won’t be enough to make LLMs reliable, credible, and accurate *enough* to really take over workplaces, at least in most cases.

      Will we see job losses? Yes and we have seen them. So-called job creators love nothing more than cutting jobs. Will certain kinds of work be heavily impacted? Probably also yes, and maybe already.

      Loading…

      Reply
2. Philosopher

   February 12, 2026 at 9:00 am

   Like Mark, I suspect there is a substantial risk of overdone hype in Schumer’s piece. At least, one hopes so. Here’s Gary Marcus on the hype: https://garymarcus.substack.com/p/about-that-matt-shumer-post-that

   Loading…

   Reply
3. Mark Engleson

   February 12, 2026 at 9:12 am

   I can’t speak to the legal field, as it isn’t where I work, but I can talk about AI. I jumped on a thread with David Lowery on Twitter — he’s a musician (Camper Van Beethoven, Cracker), as well as an academic who teaches at University of Georgia.

   My profession is technical writing, and I work at the Washington Metropolitan Transit Authority. We automate some of our trains’ functions, but human operators aren’t going away. A week or two ago, we had a jumper. Even if you had all the tech set up for monitoring, you can’t train AI to detect this. AI requires *massive* amounts of data to train and, thankfully, this is a rare event. I won’t say that our human staff catches all these, but they do sometimes, which is something that AI can’t.

   This is to say, human capabilities of pattern recognition are better than AI. We are able to detect patterns from far less evidence and experience. When it comes to detecting low-frequency, high-impact events, like subway jumpers, humans have a clear edge.

   There are also general misunderstandings of automation. My bread and butter is writing Standard Operating Procedures, which someone suggested allow AIs to take over jobs. This just isn’t the case. There’s a difference between SOPs and the kind of the code you need to write to automate machines. I haven’t done it in a long time, but I have dabbled in programming, and they’re different beasts.

   When it comes to something like creating music, you have to have a lot of technical knowledge to guide the AI if you want anything out of it. I’m a huge music fan, and I even write about it, but I lack the vocabulary and jargon that professionals know. I wouldn’t be able to compose a competent melody with an AI tool because I don’t know how to guide the tool to do its job.

   Loading…

   Reply
   1. Mark

      February 12, 2026 at 4:01 pm

      My day job is in computer vision. We can’t automate jumper detection now, but I would give 50-50 odds that we will be able to in 3-5 years. Take a look at open-vocabulary object detectors, especially SAM 3 and DINO-X. They use massive datasets to train the model to detect novel, previously unseen target classes from a text description. The performance isn’t anywhere close to good enough yet for safety-critical applications, but it’s improving, and we don’t yet know where it will top out. Actually, the main thing slowing this work down is that LLMs have sucked a lot of the talent and resources out of the computer vision space.

      Loading…

      Reply
      1. Mark Engleson

         February 12, 2026 at 4:13 pm

         Where are you going to get the massive dataset? This is a very rare event.

         Loading…

         Reply
      2. Mark

         February 12, 2026 at 5:55 pm

         The open-vocab detectors work very differently; they’re not trained like a conventional computer vision algorithm. During training, at every iteration, you sample a completely new *class*, and you give the detector a text prompt to detect that class. Typical datasets have several tens of thousands of classes. Then, during test time, it generalizes to test prompts for new classes it did not see during training. These detectors require much more data than conventional detectors to train, but you can train it without *any* exemplars of the classes you want it to find, and you can reuse it every time there’s something new you want found by just changing the prompt.

         Loading…

         Reply
4. Ed

   February 12, 2026 at 1:00 pm

   A large and powerful union representing blue-collar workers — one of the last of a dying breed — currently has a pay guarantee clause in their contract. Workers are guaranteed the base amount of wages for five days of work. How it works is this. A worker must present him or herself at the dispatch hall where jobs are assigned on a daily basis, and if no job becomes available for them that day, that counts as a “roll-over” day; this means the worker can come back the next day and move up in the line, thus having priority for that day’s dispatch. Union members are required to be paid for all “roll-over” days, provided that they either work, or present themselves as available to work, for the full Monday through Friday work week.

   This is relevant to the current discussion because much of the industry this union represents is being automated, and has been for several years. So we already have a situation where robots are taking jobs, but workers are reaping the profits anyway. (The caveats are that these guaranteed wages come from a limited fund; the system was set up during the 1970s to keep workers afloat during periods when work is slow due to economic factors. So it’s not like the union could survive full automation under its existing contract. Also, workers are only entitled to the base amount of pay, whereas many of the more specialized jobs pay considerably more than the base amount.)

   Why couldn’t an arrangement like this work for every industry? Is a lack of worker solidarity the main obstacle? This of course gets us into questions about UBI, which we’ll probably a lot more about in the coming years, if the AI job apocalypse is truly upon us.

   (I didn’t name the union in question because its membership tends to be rather touchy about making the details of their contract public — which itself says something interesting about the health of the wider economy. I’ll just say that it represents a work force on the US west coast.)

   By the way, in case philosophers are wondering about their own job security, I asked Chatgpt whether it can produce original philosophy. FWIW, here is its one-paragraph answer:

   “Original philosophy” can mean different things depending on the standard we apply. At a minimal level, it might refer to novel recombination: bringing together existing ideas, arguments, or traditions in ways that have not previously been explicitly articulated. A stronger sense of originality would involve generating genuinely new arguments within established debates—introducing a previously unrecognized objection, distinction, or line of defense that advances the discussion. The strongest and rarest form of originality would be conceptual innovation that reshapes a field altogether, reframing its central questions or altering its methodological assumptions, as in major philosophical turning points. AI can clearly achieve the first of these, since recombination and synthesis are well within its capacities. It may sometimes approximate the second, producing arguments that appear new within a debate. The third, however—deep conceptual reorientation that transforms a discipline—remains far more doubtful.

   Loading…

   Reply
   1. Jeremie

      February 12, 2026 at 2:27 pm

      The Union to which Ed refers represents employees in an industry that is structurally integral to the US economy. As in, when those members stop working, the entire economy starts to shut down. And the work can’t be outsourced to a foreign country. An arrangement like that won’t work for every industry because most industries aren’t nearly as important, and don’t have the biggest physical moat on the planet to protect them.

      With respect to “what comes next,” one might expect to see a value premium attach to jobs that cannot be automated by AI, such as those in the industry Ed references and other skilled trades, and a rotation into these jobs. Advanced economies are experiencing demographic decline, which in the long-term may mean there are enough of these jobs for everyone who needs them, although the short to mid-term prospects for victims of mass white-collar job displacement is alarming.

      Loading…

      Reply
      1. Ed

         February 12, 2026 at 6:03 pm

         The threat to the industry in question has primarily been automation, and certainly AI will impact it as well. Similar story for other industries; it’s just happening later.

         However, it is amusing to imagine laid-off white-collar workers looking at their still-employed blue-collar brethren and, like Marlon Brando, saying “I coulda been a contender.”

         https://youtu.be/1B7usdUAjFE?si=M9UoiGzJM-K2FnfJ

         Loading…

         Reply
5. Charles Pigden

   February 12, 2026 at 7:24 pm

   Just to reinforce Brian’s point that if AI can substitute for enough jobs then it might induce, rather belatedly, a Marxian crisis of capitalism Marxists believe (or used believe) that capitalism is doomed. Two chief scenarios for capitalism’s eventual demise have been suggested, one involving the alleged tendency of of the rate profit to a fall and the other involving an overproduction/underconsumption crisis (or series of crises). The first presupposes the labour theory value which is false since value as Marx conceived it simply doesn’t exist. (Furthermore it is refuted by the very possibility of commodities being produced without active labour input.) The second relates to a collective action problem, internal to capitalism as an economic system ,which might reasonably be described as a ‘contradiction’. Other things being equal it is in the interest of companies as individual enterprises to cut costs by keeping down wages and automating production wherever possible. But it is also in their collective interests to operate in an environment where there are plenty of people – often the employees of other companies – with the financial wherewithal to pay for their products. But each company pursuing its own economic interest pursues a policy of wage-reduction and automation leading to a situation in which profit becomes impossible since there aren’t enough people with enough money to buy what the companies produce. Thus there is poverty in the midst of plenty. See for example, Engels Socialism: Utopian and Scientific, and Edwards & Leiter (2025) Marx chapter 4. Edwards and Leiter make the point that this crisis, if it occurs, might not presage a socialist utopia but an Elysium society – named after the Blomkind SF movie –  in which most people live in poverty and the super-rich minority live in abundance. Hitherto the Marx/Engels prophecy has not been fulfilled, so there is clearly a flaw in Engels’ argument. But if AI-based automation does away with the need for human labour across a wide range of industries, and no new trades or professions arise to soak up the resulting unemployment, we might arrive at a situation in which most people don’t have a job and an underconsumption crisis is triggered, because the unemployed won’t have sufficient funds to buy the abundant products of industry. If this scenario threatens, is there a solution that is compatible with even a radically reformed capitalism? Perhaps not. Suppose our basic model of society is one which most goods and services are produced for profit by privately owned companies,, those profits being sustained by sale of goods and services to the employees of other companies. Such a society can sustain large numbers of people who are NOT working in the capitalist sector either because they are not working at all (at least not for pay) or because they are employees of the state. Thus Left Social Democracy (including perhaps UBI) is possible within a capitalist framework. But this only works by taxing both companies and their employees. If there are hardly any employees and most companies become unprofitable then there is nothing to tax. Again we have the problem of poverty in the midst of abundance. To solve this problem it might be necessary to move away from capitalism altogether, towards some kind of socialism. Or the problem might not be solved , in which case Elysium beckons. .

   Of course all this depends upon the number of jobs that AI can replace and whether the system will generate new trades and professions. I have an economist friend who hopes ,rather optimistically, that Bullshit Jobs will somehow save the day. But I do think that an AI-induced Marxian crisis of capitalism is the DEEP threat that we face, though this isn’t to say that things could be pretty bad, with large numbers of people being thrown out of work, without bringing down the economic system.

   Loading…

   Reply
6. Austin

   February 13, 2026 at 9:06 am

   Software engineer here wanting to add some context. The claim that Claude, or any other current LLM, is capable of making a meaningful contribution to an application with “no corrections needed” is both an exaggeration and a misrepresentation. Many of my colleagues and I are in a constant position of having to tell other colleagues that their AI generated work is not fit for addition to our codebase. Given my experience, which is anecdotal but comprises dozens of cases, I haven’t seen a single one that required no correction. I don’t believe that it’s possible beyond trivial cases; and, if a case was rendered trivial, a human was behind the rendering.

   Software is not a uniform construct. On paper it might just seem like a language is a language, but remember that these LLMs are fancy plagiarism machines. There are languages and patterns that the LLM’s are WORSE at because of more limited public samples. Nevermind the fact that they also straight-up don’t know how to use the right tool for the job https://arxiv.org/html/2503.17181v1 . That the word “intelligence” is applied to something with this type of problem makes my eyes roll fast enough to power every house on my block. These things do not learn in the way that we learn. Hand-waving at this fact is especially ridiculous where worse outcomes are concerned.

   There is a real threat that current attitudes will lead to a bottleneck of junior hires, or the erasure of more junior roles altogether. AI is patently not in a position to replaces these jobs over the long-term: experienced engineers are needed to steer the ship, and that’s being charitable. The current gambit I see unfolding in my workplace and in the industry as a whole is best conceived of as a race: can the people who own the models and the data centers ACTUALLY replace us while they still have a generation that can help shepherd their own obsolescence? Right now they’re losing the race despite a veneer of impressive gains, and the longer it goes on, the more of a generational gap in skills may be created.

   Loading…

   Reply
   1. Keith Douglas

      February 13, 2026 at 2:04 pm

      (Cyber security professional, trained originally in philosophy and logic, here)

      Indeed; let me echo Austin’s remarks). It is clear that people who claim that AI is suitable for “coding” either don’t get that 80-90% of software dev is not “programming” in the narrow sense, or have an axe to grind.

      I am, however, worried, that the subjective “I think it can” will take over, and we’ll have some sort of industry crisis.

      Loading…

      Reply
      1. Runa

         February 13, 2026 at 3:47 pm

         I suspect people underestimate the impact of the “I think it can” that Keith Douglas mentions. If middle managers (or e.g. university administrators) believe AI can do a person’s job, then, regardless of what the job-holding human might say about it, and regardless of that person’s level of expertise or the rarity of their expertise, AI will have the job. Any resistance will be considered to be self-serving, and thus responded to in kind. How many middle managers will be looking to downstream system-wide effects as opposed to approval from those in their immediate environment and status and profits they can report? Few are incentivized to do so; few are habituated to thinking in this way.

         Loading…

         Reply
7. Benjamin B.

   February 14, 2026 at 2:01 am

   I haven’t seen anyone point this out yet, but the author has an “interesting” history when it comes to making grand pronouncements about AI: https://venturebeat.com/ai/reflection-70b-model-maker-breaks-silence-amid-fraud-accusations

   It’s scary that this context is entirely absent in the reporting on this; and equally scary that most people digesting these news seem to either fail to find this themselves or consider it irrelevant.

   Loading…

   Reply
   1. PS

      February 16, 2026 at 6:46 pm

      Litigator here (and former Leiter student), but with only a few years under my belt.

      Benjamin B. — This is exactly the sort of care that we should have w.r.t. AI claims. I’ve seen a lot of these sorts of spiels from AI boosters over the last two years (which prove effective because (a) most of the audience doesn’t work in the space, and (b) gains all of the persuasive power of a seeming “confession against interest”).

      Other than Pascal wager-like disposition for those not heavily invested in the industry, caution seems the correct demeanor when reading about AI development. I will remark there isn’t enough reference to the financing of the industry, at least state-side, which appears to be a bit more concrete and tempers some of the “rosy” (or doomsday) outlooks about future projections. For the balance of sanity, people should add more skeptical tech journalists like Ed Zitron to their ordinary diet of optimistic accounts of the market.

      ***
      On a personal note, I’ll add that I’d be very, very concerned for the former student/now seasoned litigator that Brian mentions, who implies that these LLMs are already as good as a first-year associate. This is an implication that is completely at odds with my experience.

      Working for a “big law” firm, I’m familiar with AI tools specifically developed for legal application (e.g., Co-counsel, Harvey) and have to (try to) use them on a regular basis. While I haven’t used a non-enterprise version of Claude (although I would assume because of client-confidentiality, the aforementioned litigator isn’t doing so either), in my experience the usefulness of these tools for litigation-related tasks can be significant but is VERY limited. I’ll skip the useful tasks in the interest of space (I don’t really think that they are specific to law anyways), and just provide one example that illustrates their limitations. To put it euphemistically, the legal analyses generated by these tools are *deficient* enough that I feel very comfortable with saying that no attorney should rely on them for anything other than a first foray into an unfamiliar area of law (definitely some steps down from a good treatise, but generally more accessible) and as a source of potentially useful (but mostly not) case cites. Importantly, for all of the improvements yielded by better and better models over the past 3 years, I have not noticed a marked change in the quality for this use case (and, of course, there are others). Btw, while hallucinations aren’t the only problem with the output for legal research tasks, some of the hallucinations in this area are sufficiently bad to warrant a first-year being placed on a PIP if identified in the work-product that reached the senior associate or partner.

      Of course, these AI models and their applications to law (or specifically litigation) could dramatically improve in the near- or mid-term. But that almost is tautological. Simply put, it is by no means clear to me based on my ongoing experience that it will be so. And, with more confidence, it certainly has not already come to pass. Worry about AI disruption of the legal market in the near future may still be justified due to Pascal-wager reasons mentioned above or firm’s and major firm client’s (incorrectly) viewing AI as ready to replace entry-level attorneys. But–again in my experience–that justification is not yet licensed by the actual, current iteration of the underlying technology.

      Loading…

      Reply
8. Leiter: Will AI Soon Cause a Massive Contraction in Law Schools? • TaxProf Blog

   February 15, 2026 at 10:00 am

   […] Brian Leiter (Chicago), AI Developer Warns the AI Jobs Apocalypse Is Closer Than We Realize: […]

   Loading…

   Reply
9. another software engineer

   February 17, 2026 at 11:28 am

   I’m a software engineer who works at “AI adjacent” startups, and I think this article is a bit dramatic, but generally pretty spot on in direction. I wanted to organize my thoughts on it for fun, and ended up with a long post here!

   The main thing is, what one person can do will be so much greater in the future than in the past. It’s not, as many in this thread focused on, that automata will do everything without correction or non-experts will do anything without much effort. But that motivated humans using more powerful tools will do more.

   I don’t think people who are not pro software engineers can now create serious products themselves. Nor do I think that will be true even when models are 1000x better. The models will be able to manage the code then, but there will be a number of decisions which simply require expertise, -liability-, and effort spent tracking contexts which would dilute focus from other important contexts like sales, product design, etc.

   I do think we are at the first stage of seeing the need for dedicated software engineers decrease. A team might hire one less person, since their current staff can take on much more. I have already personally experienced this smaller team decision. And I have been able to work in areas of software engineering I never would have bothered with before. A lot of what software engineers spent their time on traditionally will be increasingly handled by AI, while the “higher level” tasks get increased emphasis.

   I believe ultimately the software eng role will be defined by being responsible for the code, however it happens to be written in any particular case (compare order of justification to order of discovery in phil sci).

   I think this pattern will probably occur throughout the economy. A few less people needed in any given group, and those who have jobs take on a greater variety of labor. There will still be lots of specialists in various forms, but significantly less will be needed.

   My personal bet is that even as of this year the best models are good enough for a lot of work. But there will still be many years until most industries realize it, figure out new acceptable workflows, and the changes ripple through. Also, it will take many years to develop the human interfaces and computer protocols that enable working fluidly with AI. For example, in software engineering, the file based “PR review” process simply doesn’t scale for agent driven codebases. We will be developing new representations of software that can be used to understand and act on.

   —

   I wanted to share in more detail for those interested how my day to day work has changed in the past couple months. And I think this pattern will be followed throughout many corners of the economy.

   This is all moving so fast, so takes that are even a few months old might not be relevant. I and the article are talking about the new gpt-5.3-codex, or opus-4.6 which is comparable but more expensive and slower in my experience. Not any earlier model. For example, a year ago I would hardly trust sonnet-3.x with writing simple utility functions. And sonnet-4.x was useful for removing some drudgery and learning new things but would regularly trip itself up or generate aesthetically ugly but “ok” code.

   I’m finding AI can augment basically any part of the process, but the techniques for using it vary.

   For small bugs or enhancements, in a well designed codebase (with static types, automated tests, clear patterns, etc), you can give a decently defined spec (natural language instruction, with some concrete examples and tips) and it can go take care of it. You have to review the result, and it might be pretty good though likely you’ll want to give a few more instructions or make a few targeted edits. Yes, the recent change is that this is often way faster and -better- than doing it by hand.

   For bigger tasks, I now spend a lot of time having a conversation “planning” what needs to be done with the AI. One trick is asking it to interview you about your ideas to help articulate requirements, uncover issues, follow local conventions, and get new ideas. Once a decent spec is constructed, it implements a checklist of tasks. This is never the end of the matter, and the human has to step in and adjust things to preference, get ahead of issues with foresight, etc. Usually I watch its execution and intervene midway to correct bad paths it started going down or give more context (“no use this_type for x @location_of_type”).

   I imagine people who are good at using AI for writing have a similar workflow. Not just saying “write about this” then embarrassing themselves by publishing the output or just laughing at how bad or cliche it is. Instead it’s a sounding board as you iterate through very specific ideas and phrasings in a very tight loop. And the result is still very much yours.

   Loading…

   Reply
10. S.

    February 23, 2026 at 4:58 pm

    Tldr: there are serious arguments for why AI will be disruptive to all White collar work: I’m wondering if people are seriously considering this given the data that is coming in? And if so what they are telling their students?

    I’ve been following AI for a while now. I had a cursory interested in it as a teenager looking at some of the stuff from Ray kurzweil. I’m in a field of study that intersects a lot with cognitive science and issues surrounding cognition and consciousness, so I always found the stuff fascinating even though I thought some of the projections were far-fetched. I’ve come to the conclusion that a lot of the stuff he said ( along with other futurists) we’re in the main correct: and this last semester I finally implemented some AI modules until my intro class. While there have been have been some discussion on this board about AI and cheating and all of that; I don’t really think the existential dimensions of AI have been taken up here to the degree that is necessary ( maybe I’m wrong and haven’t done a thorough search). The fundamental thesis that I have here is that AI is going to transform society and higher education completely in the very proximate future.

    What I think we need to continue with here is the possibility of the obsolescence of the current educational paradigm of credentialing since I think AI will have the capacity to do most white college jobs (and in our capacity as instructors and professors at a university that is on main task which is the offload the training of these white collar jobs from the from the employers). What I want to outline the reasons why I think that this is not a far-fetched projection based upon some of the technical advancements of AI and certain a priory principles that follow straightforwardly from certain naturalistic philosophical premises.

    The main foundation of the argument that I’m going to make here is that there is no longer any paradigm shifting theoretical moves necessary for AI to be able to displace most White collar work: all this necessary now is more scaling and compute and adaptation by different organizations. All the main arguments against why AI won’t be able to perform these tasks can be grouped over a few arguments. There are more arguments than these but these are some of the main ones that come up and I think once you dismiss these it’s easier to see why the no paradigm shift position is very likely to be the case. 1. One of the first arguments that AI lacks the ability for true or genuine creativity as it pertains to scientific or theoretical inquiry. The essential argument is that AI has the capacity to recombine already existing knowledge but it doesn’t have the capacity to produce paradigm shifting knowledge on its own ( 99% of white college jobs don’t even need this in order to function but it’s an argument that I’ll consider nonetheless). My response to this idea is that this process can essentially be brute forced and that essentially all that even human creativity is is the recombination of existing information over a large enough problem set: The brute force creativity argument holds that the qualitative distinction between recombination and genuine creativity may not be fundamental. Given sufficient problem space size and iteration, massive combinatorial search can produce functionally novel outputs indistinguishable from human creativity. The AlphaGo and AlphaZero precedents demonstrate this empirically systems producing genuinely creative moves through architectures entirely unlike human cognition. The implication is that creativity is better understood as a quantitative threshold in combinatorial search space rather than a qualitatively distinct cognitive capacity requiring special mechanisms.

    2. Some people argue that AI cannot possibly replace White collar jobs because there is a set of tacit information that experts in a given field possess that is not directly entered into the training data of AI systems but tacit knowledge in humans probably emerges from accumulated exposure to cases plus feedback rather than from some special non-empirical source. If the same empirical substrate, sufficient instances plus feedback signals, generates tacit knowledge in humans, then AI systems exposed to vastly larger instance sets can develop functionally equivalent judgment through the same empirical process, compensating for reduced richness per instance through sheer scale of exposure. Essentially if tacit knowledge is empirical all the way down and simply a matter of empirical instances plus feedback then AI will be able to develop this kind of knowledge.

    3. Some people argue that AI will falter because it won’t be able to draw causal emphasis from real world context. But multimodal training data on video information renders this objection essentially moot.

    4. Some people argue that the hallucination issue will keep AI from being verifiable enough to be using most White collar work since while it is very efficient and intelligent it can make disastrous mistakes that human beings would ever make. There are two responses to this: on the one hand the hallucinations have been becoming more minimal as the models have developed: there is dats on on this that I can add to this post if needed. So even though this might be the case we could end up in a situation where AI produces most of the output but far fewer humans are employed to simply verify the AI from mistakes which would would still lead to less humans being employed since fewer of them would be needed to verify the AI output. But I think a stronger case can be made here. Essentially this problem will be tackled using a nested AI agent framework: Specialized verification agents checking reasoning agent outputs converts epistemic reliability from a property of individual systems into an emergent property of network architecture. This addresses the reliability objection without requiring paradigm shifts because verification is computationally easier than generation, allowing less capable specialized models to check more capable reasoning models. That is to say, frontier AI models that are better at novel reasoning will be connected up with verification AI agents which are more reliable in terms of their output and in this chain relationship the problem of hallucination will be solved by internal mechanism within an AI system itself. These are mostly the main arguments against the idea that AI will have the capacity to overtake all human intellectual work: I see no a priori arguments given a naturalistic philosophy for why these compounding effects will not have devastating results on White collar work given these arguments. AR agents are becoming more robust with stuff like open claw: the time frame during which the new models are released is becoming more and more truncated, the internal memos from many of the AI companies themselves report that most of their software code is being produced by the models themselves. And even the old projections from guys like Ray kurzweil made almost 25 years ago: projects that AI intelligence far surpass human intelligence by 2029. It seems that converging evidence is that human intelligence will be surpassed in the very near future. My question here is are people talking to their students or each other about this: not what AI can be used for in terms of cheating on homework but what the implications are for us training students who might not have any kind of white collar job waiting for them once they graduate? The speculations of AI have made me go back to Marx and the idea of the immiseration thesis ( but I’ll leave that aside). My question is have any of you been thinking about the question of AI as it pertains to the existential existence of the University system and how long we might have to exist as we currently do. I think we should take very seriously the many projections that are on the table about AI surpassing humans intellectually and what that means for our jobs and what we should be telling our students about their future.

    Loading…

    Reply