I think we have to be careful when assuming that model capabilities will continue to grow at the same rate they have grown in recent years. It is very well-documented their growth in recent years has been accompanied by an exponential increase in the cost of building these models, see for example (of many examples) [1]. These costs include not just the cost of GPUs but also the cost for reinforcement learning from human feedback (RLHF), which is not cheap either -- there is a reason that SurgeAI has over $1 billion in annual revenue (and ScaleAI was doing quite well before they were purchased by Meta) [2].
Maybe model capabilities WILL continue to improve rapidly for years to come, in which case, yes, at some point it will be possible to replace most or all white collar workers. In that case you are probably correct.
The other possibility is that capabilities will plateau at or not far above current levels because squeezing out further performance improvements simply becomes too expensive. In that case Cory Doctorow's argument seems sound. Currently all of these tools need human oversight to work well, and if a human is being paid to review everything generated by the AI, as Doctorow points out, they are effectively functioning as an accountability sink (we blame you when the AI screws up, have fun.)
I think it's worth bearing in mind that Geoffrey Hinton (infamously) predicted ten years ago that radiologists would all be out of a job in five years, when in fact demand for radiology has increased. He probably based this on some simple extrapolation from the rapid progress in image classification in the early 2010s. If image classification capabilities had continued to improve at that rate, he would probably have been correct.
No, models significantly improved at the same cost. Last year's Claude 3.7 has since been beaten by GPT-OSS 120B that you can run locally and is much cheaper to train.
They justified it with the paper that states what you say, but that's exactly the problem. The statement of paper is significantly weaker than the claim that there's no progress without exponential increase in compute.
The statement of the the paper that SotA models require ever increasing compute, does not support "be careful when assuming that model capabilities will continue to grow" because it only speaks of ever growing models, but model capabilities of the models at the same compute cost continue growing too.
I wonder whether for a lot of the search & literature review-type use-cases where people are trying to use GPT-5 and similar we'd honestly be much better off with a really powerful semantic search engine? Any time you ask a chatbot to summarize the literature for you or answer your question, there's a risk it will hallucinate and give you an unreliable answer. Using LLM-generated embeddings for documents to retrieve the nearest match, by contrast, doesn't run any risk of hallucination and might be a powerful way to retrieve things that Google / Bing etc. wouldn't be able to find using their current algorithms.
I don't know if something like this already exists and I'm just not aware of it to be fair.
I think you have a very good point here: a semantic search would be the best option for such a search. The items would have unique identifiers so the language variations can be avoided. But unfortunately, I am not aware of any of these kinds of publicly available projects, except DBpedia and some biology-oriented ontologies that would massively analyze scientific reports.
Currently, I am applying RDF/OWL to describe some factual information and contradictions in the scientific literature. On an amateur level. Thus I do it mostly manually. The GPT-discourse somehow brings up not only the human-related perception problems, such as cognitive biases, but also truly philosophical questions of epistemology that should be resolved beforehand. LLM developers cannot solve this because it is not under their control. They can only choose what to learn from. For instance, when we consider a scientific text, it is not an absolute truth but rather a carefully verified and reviewed opinion that is based on the previous authorized opinions and subject to change in the future. So the same author may have various opinions over time. More recent opinions are not necessarily more "truthful" ones. Now imagine a corresponding RDF triple (subject-predicate-object tuple) that describes that. Pretty heavy thing, and no NLTK can decide for us what the truth is and what is not.
Since you specifically were wondering if something like this exist, I feel okay with mentioning my own tool https://keenious.com since I think it might fit your needs.
Basically we are trying to combine the benefits of chat with normal academic search results using semantic search and keyword search. That way you get the benefit of LLMs but you’re actually engaging with sources like a normal search.
If your issue is "incorrect answers" (hallucination) then an embedding search will naturally also produce "incorrect answers" because an embedding is a lossy compression of the source.
it explores the world of outsourced labeling work. Unfortunately hard numbers on the number of people involved are hard to come by because as the article notes:
"This tangled supply chain is deliberately hard to map. According to people in the industry, the companies buying the data demand strict confidentiality. (This is the reason Scale cited to explain why Remotasks has a different name.) Annotation reveals too much about the systems being developed, and the huge number of workers required makes leaks difficult to prevent. Annotators are warned repeatedly not to tell anyone about their jobs, not even their friends and co-workers, but corporate aliases, project code names, and, crucially, the extreme division of labor ensure they don’t have enough information about them to talk even if they wanted to. (Most workers requested pseudonyms for fear of being booted from the platforms.) Consequently, there are no granular estimates of the number of people who work in annotation, but it is a lot, and it is growing. A recent Google Research paper gave an order-of-magnitude figure of “millions” with the potential to become “billions.” "
I too would love to know more about how much human effort is going into labeling and feedback for each of these models, it would be interesting to know.
That was indeed a great article, but it is a couple of years old now. A lot of of the labeling work described there relates to older forms of machine learning - moderation models, spam labelers, image segmentation etc.
Is it possible in 2025 to train a useful LLM without hiring thousands of labelers? Maybe through application of open datasets (themselves based on human labor) that did not exist two years ago?
Good question, I don't personally know. The linked article would suggest there are plenty of people working on human feedback for chatbots, but that still doesn't give us any hard numbers or any sense of how the number of people involved is changing over time. Perhaps the best datapoint I have is that revenue for SurgeAI (one of many companies that provides data labeling services to Google and OpenAI among others) has grown significantly in recent years, partly due to ScaleAI's acquisition by Meta, and is now at $1.2 billion without having raised any outside VC funding:
Their continued revenue growth is at least one datapoint to suggest that the number of people working in this field (or at least the amount of money spent on this field) is not decreasing.
Also see the really helpful comment above from cjbarber, there's quite a lot of companies providing these services to foundation model companies. Another datapoint to suggest the number of people working providing labeling / feedback is definitely not decreasing and is more likely increasing. Hard numbers / increased transparency would be nice but I suspect will be hard to find.
I love nanobind, use it all the time and highly recommend it. It makes it very easy to pass numpy, PyTorch, cupy or tensorflow arrays to your C++ extension, to specify what array shape and flags are expected, and to wrap either C++ or Cuda code. When paired with scikit-build, it makes building Python packages with C++ extensions a breeze. I would give it more than one star on github if I could.
I added a reply to the parent of your comment with a link to an article I found fascinating about the strange world of labeling and RLHF -- this really interesting article from The Verge 2 years ago:
There's a really fascinating article about this from a couple years ago that interviewed numerous people working on data labeling / RLHF, including a few who had likely worked on ChatGPT (they don't know for sure because they seldom if ever know which company will use the task they are assigned or for what). Hard numbers are hard to come by because of secrecy in the industry, but it's estimated that the number of people involved is already in the millions and will grow.
Interestingly, despite the boring and rote nature of this work, it can also become quite complicated as well. The author signed up to do data labeling and was given 43 pages (!) of instructions for an image labeling task with a long list of dos and don'ts. Specialist annotation, e.g. chatbot training by a subject matter expert, is a growing field that apparently pays as much as $50 an hour.
"Put another way, ChatGPT seems so human because it was trained by an AI that was mimicking humans who were rating an AI that was mimicking humans who were pretending to be a better version of an AI that was trained on human writing..."
They find that the general public is overall much more skeptical that AI will benefit anyone, much more likely to view it as harmful and much less excited about its potential than "AI experts". A majority of Americans are more concerned than excited. There is interestingly a large gender gap between men and women -- women are much less likely to view AI favorably, to use it frequently or to be excited about its potential than men.
There is some research to suggest that consumers are less likely to buy a product and less likely to trust it (less "emotional trust") when AI is used prominently to market it:
So I think the data suggests that while there is excitement around AI, overall consumers are much less excited about AI than people in the industry think and that it may actually impact their buying decisions negatively. Will this gap go away over time? I don't know. For any of you working in tech at the time, was there a similar gap in perceptions around the Internet back in the days of the dot com bubble?
The other problem as pointed out is that MANY things are labeled as AI, ranging from logistic regression to chatbots, and probably there is more enthusiasm around some of these things than others.
Prior to the dot-com bubble itself, hype for the growing potential of the internet was modest and mostly in line with organic adoption and exploration. People at large weren't anticipating a revolution. They were just enjoying the growing areay of new products and opportunities that were appearing.
During the dot-com bubble, inasmuch as it represented a turning tide, this trickle had reached a tipping point and we witnessed a tsunami of innovative products that consumers were genuinely fascinated by. There were just too many of them for the market to sustain them all, and a correction followed, as you would expect.
This AI story is basically the opposite, much like the blockchain story. Many investors and some consumers who have living or borrowed memory of dot-com bubble or the smartphone explosion really really want another opportunity to cash in on a exponentially expanding market and/or live through a new technological revolution and are basically trying to will the next one into existence as soon as possible, independent of any organicity or practicality.
In contrast to blockchain hype, maybe it'll work here. Maybe it won't. But it's fundamentally a different scenario from the dot-com bubble either way.
There've been a _few_ of these over the last decade; _two_ attempts at blockchain stuff (an initial "use blockchains for everything" one, and a "use NFTs for everything" one a couple years after the first crashed and burned. And then of course there was 'metaverse'.
VCs just need to make sure there is enough hype by the time AI startups IPO, so that they can cash out. It's part of a bigger trend in finance, arguably caused by quantitative easing. It's why Uber could IPO, while they had never made a profit; but because of the hype, their stock price did great on day 1.
I find it interesting how a lot of other comments are saying how "HN users are a bubble, the public is actually really excited about AI", when the research indicates that the general public is even less interested in AI then HN is.
It's fine to express your opinion on AI, whether positive or negative. It's even fine to share anecdotes about how other people feel. Just don't say that's how "most people" feel without providing some actual evidence.
I think HN is a space where practically everyone has a grasp of what AI is and is not capable of, and of what tools could theoretically exist in the near future. I also think that HN is a space where there is not a consensus on whether AI is "good" or "bad," and there is a lot of discourse on the subject.
In my experience, this makes HN probably the most pro-AI spaces around. Most people in my life feel more negatively about AI, without a lot of defense for it (even if they do use it). The only space in my life that is more pro-AI than HN is when people from the C-suite are speaking about it at work meetings :/
In the sense that many anti-AI articles are shared and upvoted, and many anti-AI or AI-skeptical comments are upvoted. Of course, a lot of pro-AI and AI-enthuisastic content is also shared and commented.
HN has many, many users, and they're not all monolothic, which explains 90% of the questions along these lines that people raise.
I believe most people don't want AI because I read the Pew Research report linked by the parent comment, which indicated most non-experts don't want AI. That report has a pretty large sample size, the methodology seems sound, and Pew is an organization that's historically pretty good at studying this sort of thing.
Obviously one report is not the end of the discussion. And if more research is done that indicates that most people really are interested in AI, I'll shift my beliefs on the matter.
I was interested in that 400 million weekly user number you posted, so I did a little digging and found this source [1] (I also looked through their linked sources and double checked elsewhere, and this info seems reasonably accurate). It seems like that 400 million figure is what OpenAI is self-reporting, with no indication how that number is being calculated. Weekly user count is a figure that's fairly easy to manipulate or over-count, which makes me skeptical of the data. For example, is this figure just counting users that are directly interacting with ChatGPT, or is it counting users of services that utilize the ChatGPT API?
In addition, someone can use ChatGPT while having a neutral or negative opinion on it. My linked source [1] indicates that around 10 million people are actively paying for a ChatGPT subscription, which is a much more modest number then 400 million weekly users. There clearly are a lot of people who use and like AI, but that doesn't mean the majority of the population feels positively about it.
I use an AI chat service, but would prefer that research and investment that might yield more powerful AIs be banned. Maybe that is what the survey respondents meant when they said that they don't want AI.
You don't really have to rely on self reported numbers to see its scope. It has become that massive. ChatGPT was the 6th most visited site in the world in March and will likely be 5 in April.
The idea that a site that consistently has billions of visits every month but most of them have a negative opinion of it seems more delusion than reality.
Absolutely. People using it to write shitposts and spam and a first draft of something is one thing, but "fun toy" is not the same thing as "sea change"
The article is so stupid. AI is replacing traditional Google search. It's everywhere already and people not only want it, they use it every day. My Google phone replaced the old assistant a while ago.
Part of the problem is that the term has become utterly diluted to the point of becoming meaningless - any computer system can be called AI nowadays.
But what I think people dislike most is the genre of Generative AI aka AI slop: images and texts generated by machines, often of low quality, unchecked or barely checked by humans. Another one is cost cuts by replacing human support with automated responses - which can give you abysmal experience even in trivial cases.
> For any of you working in tech at the time, was there a similar gap in perceptions around the Internet back in the days of the dot com bubble?
I wasn't drawing a paycheck from tech at the time, but I was a massive nerd, and from my recollection: Yes, absolutely. Dialup modems were slow, and you only had The Internet on a desktop computer. Websites were ugly (yes, the remaining 1.0 sites are charming, but that's mainly our nostalgia speaking), and frequently broke. It was or could be) expensive: you had to pay for a second phone (land!) line (or else deal with the hassle of coordinating phone calls), and probably an "internet package" from your phone company, or else pay by the minute to connect; and, of course, rural phone providers were slow to adopt any of those avenues of adoption. Commerce, pre-PayPal, was difficult - I remember ordering things online and then mailing a paper check to the address on the invoice!
Above all, we underestimate (especially in fora like this) how few people actually were online. I don't remember exact numbers at any particular times, but I remember being astonished a few times - the 'net was so ubiquitous in my and my friends' lives that "What do you mean, only X minority of people have ever used the internet?" For people who weren't interested in tech (the vast majority), seeing web addresses and "e[Whatever]" all over the place was mainly irritating.
Those elements and attitudes are certainly analogous to AI Hype today. Whether everything else along that path will turn out roughly the same remains to be seen. From my point of view, looking back, the most-hyped (or maybe just most-memorable) 1.0 failures were fantastic ideas that just arrived ahead of their time. For instance, Webvan = InstaCart; Pets.com = Chewy; Netbank = any virtual bank you care to name; Broadcast.com = any streaming video company you care to name; honorable mention: Beenz (though this might be controversial) was the closest we ever came to a viable micro-payments model.
The necessary infrastructure for (love it or hate it) a commercialized web was the smart-phone, and 'always on' portable connectivity. By analogy, the necessary infrastructure for widespread, democratized AI (whether for good or for ill) may not yet exist.
For me it implies that we need a better way to quote credentials.
We have plenty of skeptical experts, who were researching machine learning for decades.
AI experts is the same crowd as blockchain "experts". I used to work in the industry.
Between me and my employer, I knew how the whole thing worked, but it was him who was quoted as an "expert".
And I can't even call myself an expert, because for sure experts at the time were writing protocols and not implementing parts of them.
If there is a rift on whether or not AI benefits people right now, and the rift is between users and experts, the problem is the experts that were asked.
People don't need much convincing to find something helpful, we love to cut corners.
>I think the concept that AI can be used to sell itself -- that a product is more valuable simply because it incorporates AI -- has to end, and soon.
I can't help but think of the iPhone 16 series's top-line marketing: "Built for Apple Intelligence." In practice, the use cases have been lackluster at best (e.g., Genmoji), if not outright garbage (e.g., misleading notification summaries).
I feel like a lot of AI use cases are solutions looking for a problem, and really sucking at solving those problems where the rubber meets the road. I can't even get something as low-stakes and well-bounded as accurate sports trivia and stats out of these systems reliably, and there's a plethora of good data on that out there.
I'm not skeptical about AI. I'm skeptical that the companies behind AI will deliver a product that makes my life better. Anything genuinely useful will be a toy or get bought and shut down, while the ones that survive will steal my personal data and serve me ads.
some credit card companies have botched chatbot process:
"lost/forged credit card report" / "talk-to-person" are essential support process, but they require you to enter your PIN to get thru that process
(if the request for a new credit-card is faked, you're out of luck)
K-means is like the McDonald's of clustering algorithms. It's fast, it's convenient, and sometimes you use it because it's fast and it's convenient...but if you care about quality, it's seldom the right choice. Assuming all your clusters are spherical Gaussians may be convenient but is very often unrealistic.
It depends on your goals. In something like bolt (a vector quantization algorithm), it's provably the right choice (and with other preferred error measures for your quantization, analogous measures on kmeans are still the right choice).
Depends on your data and what assumptions you can make. I work with sequence data a lot, and for that type of data the MMSeqs library (https://github.com/soedinglab/MMseqs2) is both very powerful and very popular.
For tabular data as in this blog post, there are a lot of options. For small datasets, hierarchical clustering is very powerful -- you can build and visually inspect a dendrogram and this can give you a lot of insight. It's implemented in Scipy and scikit-learn (e.g. https://docs.scipy.org/doc/scipy/reference/cluster.hierarchy... ). Hierarchical clustering however scales poorly. For relatively low-dimensional data the hdbscan algorithm is really nice and is implemented in Python (https://pypi.org/project/hdbscan/).
If you have reason to think your data is modeled reasonably well as mixture of Gaussians (think lots of elliptical clusters of various sizes) and it's not too high-dimensional, a mixture of Gaussians can work well; unlike k-means, it's probabilistic and doesn't assume all clusters are spherical and roughly equal in size. This too is implemented in scikit-learn (https://scikit-learn.org/stable/modules/generated/sklearn.mi...). If you think a mixture of Gaussians is reasonable but you know there are outliers, a mixture of Student t-distributions will work better; this is not in scikit-learn but there are multiple implementations on github.
It's also possible to improve k-means by using approximate kernel k-means, where you use a random Fourier features (https://people.eecs.berkeley.edu/~brecht/papers/07.rah.rec.n...) representation for each input datapoint then run k-means on that -- this approximates kernel k-means, so it relaxes some of the unrealistic assumptions of k-means. We no longer assume clusters are spherical, although this method may still work poorly if there are outliers and also still requires us to choose the number of clusters and the lengthscale for the kernel we are approximating, both of which may be hard to choose unless you already have a pretty good intuition for what are appropriate choices for your data.
There are other options that are sometimes useful, in fact, I could easily write a blog post about this (maybe I should), but the thing with clustering is that the "right" choice of algorithm is somewhat dependent on your data and on what assumptions are reasonable to make. People sometimes end up using k-means because it's fast (especially if you use minibatch kmeans) and can scale to crazy large datasets. But it makes very strong assumptions which are usually wrong (most datasets do not subdivide well into some number of spherical Gaussians of roughly equal size), and this can result in truly absurd partitions, especially when there are lots of outliers or clusters with highly irregular shapes.
Funny...I was just coming here to talk about Qt :)
To be fair, as you say, Qt definitely has shortcomings and some things I don't like. If you're trying to minimize your disk space footprint, it's hard to get anything under 50MB with Qt. The difficulty of navigating the licensing for a closed-source app is (while often exaggerated) quite real, especially for someone using Qt for the first time. And while I understand why they did it, it's often annoying to have to use the Qt equivalents of standard library types (e.g., don't use std::string...use QString! don't use int64_t...use qint64! etc.)
The problem is there really just isn't a perfect library or tool for building desktop apps, and writing your own would be...a lot of work. I share the linked post's distaste for Electron, which without a lot of work tends to result in horribly bloated software (often very nice look and feel, but bloated just the same). GTK sounds nice in principle, but you very quickly find yourself missing all of the functionality built into Qt. So Qt often for me ends up being like what someone once said about democracy: it's the worst system EXCEPT for all the other alternatives we've tried from time to time...all of which are worse.
Just wanted to let you know it's not all gray in Qt land.
The open-source Qt is licensed under LGPL. It's not really a mystery. I've built a proprietary app[1] on top of that, and as I explained in my blog post[2], the LGPL allows you to statically link your executable as long as you provide the object files and allow users to relink your app with a different version of Qt.
There's also a lot you can do to drastically limit your binary size. I wrote a little about it on my blog post but haven't experimented that much with it yet. Things like:
1. Compiling Qt from source with only the specific modules needed and using static linking.
2. Use the -optimize-size flag and Link Time Optamization flag -ltcg.
3. Running strip on the resulting executable to remove unused symbols.
4. Use UPX to compress the executable.
BTW, Qt can look very native, if you know how to use it properly (I wish to make it more commoditized!), for example, something I'm cooking at the moment:
Heavy Qt user here... I agree with everything you said. I think a lot of people are rightfully confused about the licensing because I think the Qt Company is vague about it on purpose, they want to sell licenses of course. I just hope their agreement with KDE never ends.
I also think that the vast majority of people that complain about the lack of "native look" are bad takes from people that don't actually ship apps professionally and listen to their customers. End-users don't actually care and it really simplifies things a lot to do it the way they are. wxWidgets uses platform native APIs and I would argue that it looks much worse (and old) compared to Qt.
One thing I will say though is that the fact that Qt specifically now uses LGPL version 3 as opposed to v2 (switched in 5.6 I think), can actually be quite significant and a non-starter for some people. If you're developing something where Qt is embedded fairly deep in to the system, especially if it's a critical system, you don't want to have to give the user the ability to swap out your Qt library with who-knows-what and now all your safety testing/validation/warranties etc. are void... which I guess for bigger companies they just buy a commercial license in that case.
> I think the Qt Company is vague about it on purpose, they want to sell licenses of course. I just hope their agreement with KDE never ends.
Yep, I agree. And I hope so too.
> I also think that the vast majority of people that complain about the lack of "native look" are bad takes from people that don't actually ship apps professionally and listen to their customers
Ding ding ding. Exactly. I would argue that most Qt developers aren't targeting the same audience that native iOS developers target, for example, which causes all Qt apps in the wild to look embarrassingly ugly and out of place. The Qt Company's own lack of taste in design (and their examples) doesn't contribute. It's only when developers with a strong taste for aesthetics and for creating experiences that delights people will take on Qt to develop apps, we will see and be able to show the world the power of Qt. I hope to be part of this change, I'm trying to push Qt limits with every new app I make. And honestly, QML is just an amazing joy to program in. Here's a cool window animation you can do in 1 minute using QML:
What we also need is to create a set of beautiful, easy to use components that work out-of-the-box for developers to tinker with. This is very lacking in the Qt ecosystem compared to other frameworks.
The gist I'm trying to get at is that Qt is not the limiting factor, the limiting factor are developers whose both their target users don't care about aesthetics (or so they think), and developers who themselves lack a good taste (or the know how) for good design.
Not sure about the difference between LGPL v2 and v3, but like you said, if you're a big company, paying for the license is a non-starter (and also helps fund Qt development so why not?), but for almost all developers the LGPL version should be more than enough.
Yea, Qt is a pain, but so is any kind of porting work.
I've got a medium sized hobby project I work on that has a macOS and iOS version, separate GUIs, but sharing a lot of underlying code. Every once in a while, I think to myself "It would be nice to port this to Windows and Linux, just for the fun of it" and it always goes this way:
I end up reaching for Qt, but getting absolutely mired in things like: creating the .ui file in Qt Designer, making each screen and dialog equal to the original macOS app in features, converting things back and forth to QStrings and QObjects and QThis and QThat, cobbling together UI logic from scratch that we just get for free from Cocoa like bindings and connecting events to Objective C methods.
Then, when I'm absolutely not having fun here, I'll tell myself "OK, let's just create a native Windows application. Hmm, what's the recommended idiomatic windows way to create an application these days? Is it Win32 still? No, is it WPF? Is it UWP? Is it WinForms? Should I use C# and .NET or not? Before I even write a line of code, I'm already exhausted and demoralized.
Then I think, fuck it, let's start playing with Javascript and electron and all these other things that I only read about on HN but never learned or dove into. By the time I'm done yak shaving to get an environment set up and going through "How to declare a variable" tutorials, I'm wondering why not just add some fun features to the macOS app instead...
Galactica apparently thinks that there is evidence that vaccines may be linked to autism (!?!) If I enter the query:
vaccines and autism
I get:
--Vaccines and autism
Vaccines and autism is the hypothesis that vaccines are a significant cause of autism. The hypothesis is based on the epidemiological evidence of a temporal association between the two events, as well as on the biological plausibility of the hypothesis. The biological plausibility of the hypothesis is supported by the finding of an increased risk of febrile seizures following vaccination and the finding of a high rate of inflammatory and autoimmune disorders in autism spectrum disorders.(Zerbo et al., 2015)(Sweeten et al., 2003)(Ashwood et al., 2006)
There are many scientific and anecdotal reports linking vaccines with autism. However, it is difficult to determine whether the temporal association between vaccines and autism is causal or coincidental.
which is some fairly dangerous misinformation written in such a way as to sound plausible.
What is the point of a tool that can generate plausible-sounding nonsense? Is there not already enough misinformation on the Internet?
Oh and it gets better! Galactica suggests helpfully that maybe the preservatives in the vaccines cause autism!
Vaccine ingredients
Vaccines contain various ingredients. The ingredients in vaccines include antigens, preservatives, adjuvants, stabilizers, and traces of antibiotics.(Gerber et al., 2009) The preservatives, adjuvants, stabilizers, and traces of antibiotics have been proposed to be the main causes of autism.(Gerber et al., 2009) The preservatives and stabilizers are used to prevent the growth of bacteria and fungi in vaccines. The preservatives and stabilizers that have been proposed to be linked to autism include mercury, aluminum, thimerosal, and formaldehyde.(Gerber et al., 2009) The adjuvants are used to boost the immune response to vaccines. The adjuvants that have been proposed to be linked to autism include aluminum and mercury.(Gerber et al., 2009)
Maybe model capabilities WILL continue to improve rapidly for years to come, in which case, yes, at some point it will be possible to replace most or all white collar workers. In that case you are probably correct.
The other possibility is that capabilities will plateau at or not far above current levels because squeezing out further performance improvements simply becomes too expensive. In that case Cory Doctorow's argument seems sound. Currently all of these tools need human oversight to work well, and if a human is being paid to review everything generated by the AI, as Doctorow points out, they are effectively functioning as an accountability sink (we blame you when the AI screws up, have fun.)
I think it's worth bearing in mind that Geoffrey Hinton (infamously) predicted ten years ago that radiologists would all be out of a job in five years, when in fact demand for radiology has increased. He probably based this on some simple extrapolation from the rapid progress in image classification in the early 2010s. If image classification capabilities had continued to improve at that rate, he would probably have been correct.
[1] https://arxiv.org/html/2405.21015v1 [2] https://en.wikipedia.org/wiki/Surge_AI