agentcoops

I agree with your general point: there is a tendency to group not least these deals together without distinction as confirmation of a pre-existing belief that the AI bubble is soon to collapse. The latter certainly might be true, but we should be careful about too quickly deciding what evidence supports it. I am inclined to strictly separate: (a) the NVIDIA deal as a vendor subsidizing one of their largest clients, which perhaps signals market weakness; from (b) the AMD deal as a non-market-leading vendor trying to entice a new enterprise client, who will only be able to use AMD's chips for serious training after significant (and risky) collaboration to improve their product offering.

This distinction is important to me because I see more concrete evidence for a possible material drop in NVIDIA's business short- to medium-term than a collapse of the sector as a whole. The chips act clearly shut them out unexpectedly from their second largest market and now it seems likely that Chinese chips will be competitive for training sooner than expected. Indeed, if Chinese AI firms are suddenly able to obtain even a roughly approximate product at considerably less cost, OpenAI will suddenly find themselves at a cost if not compute disadvantage to their Eastern competitors. It isn't a surprise, then, that OpenAI is now looking to reduce their present vendor lock-in with NVIDIA.

Overall, it's not great for the latter if they lose access to their second largest market, suddenly have viable competitors in their home market, and either lose some of a major client's business or have to significantly reduce pricing to retain it.

Completely agreed: the way he articulates the problem is self-defeating. Apologies if this sounds absurdly abstract or hand-wavy, but I think the correct framing really has more to do with a sort of essential clash between law and software as technologies of social regulation.

Written law was something very important historically: from unwritten norms to codified representations that, however imperfect and provisional, were more or less accessible to all and changeable through whatever process, democratic or otherwise. Over the last decades, we -- and this could be taken rather literally as "the readers of hacker news" -- have been encoding so many aspects of the world into software in a way that doesn't clearly coincide with the legal norms of any particular let alone every country.

On the one hand, software is clearly "better" than law in at least the sense that the former eliminates the necessary ambiguity of the latter: the interpretation or "implementation details" of even just a particular law are always disputed. Perhaps a particular implementation of cross-border financial transactions, say, or of personal identity doesn't in fact reflect what the developers or product managers intended, but if that is identified it can be changed.

Yet, on the other hand, it is certainly true that, from the perspective of regulators let alone the masses, the resulting situation is much closer to pre-law. Nobody has any idea how or why they were suddenly banned from Instagram or their PayPal account closed, let alone how money actually moves around the world when they send a friend funds through Transferwise. Certainly, if we don't even know how things are working there is no process by which it could be decided that things should work differently, let alone a process by which software would concretely be made to work differently.

Indeed, I am skeptical that law as such will ever be able to regulate software: even just considering the problem in terms of a single country/legal system, how does one actually guarantee that the ever-changing corpus of code complies with the ever-changing and essentially ambiguous body of regulations? One of course sees this with the EU as the "avant-garde" of the struggle to regulate software. They pursue either these incredibly general wars on "cookies" that don't solve real problems, or endlessly deliberate when it's already too late about how to handle AI, or produce something relatively well-conceived like GDPR where enforcement is then incredibly unclear if not impossible.

TLDR I have no idea what the solution is, but I think the intrinsic problem of law and software is incredibly important to take seriously as software eats the world. At the very least, it's not just a problem of "competition" since, as you note, monopoly is at least one sufficient condition [1] of eventually rendering the way software regulates the world transparent and open to change.

[1] Undoubtedly, there are other ways this could possibly transpire through open source etcetc -- however, even in that case there has to be a guarantee that particular software defines the operations of a particular domain, i.e. that there exists software through which one can understand that domain and hypothetically change, which is in some sense just a "public" monopoly.

Very much agreed. I think the dominance of the chat interface to LLMs has materially impaired the general usefulness of these tools — the sooner it goes away the better. It’s almost impossible to explain to a non-engineer how the illusion of a continuous conversation is crafted through context management and why past moments in a conversation might fall out of memory. My general advice to non-technical friends is to create a new conversation for each prompt so that they can get a more deterministic sense of how to formulate instructions and which are successful.

I was really hoping Apple would make some innovations on the UX side, but they certainly haven’t yet.

Completely agreed. I note in a comment below that my only aim is to distinguish between deals that provide new evidence of a bubble and those that do not provide any further evidence towards that conclusion. I think the Oracle and especially the recent NVIDIA deals provide such evidence, while this AMD deal does not.

As @stingraycharles notes above, the AMD stock went up a lot already and this "may finally enable AMD to get a foot in the door in the whole large scale AI market."

I understand that. I simply think there's some importance in at least trying to empirically tease out what represents: (1) a circular deal in which vendors facing the limits of growth are subsidizing vulnerable clients; versus (2) a risk-hedging deal in which a non-market leading vendor offers upside to a market leading client.

Again, I'm in no way denying either that (1) exists or that there's almost certainly an AI bubble -- I just think this difference is material.

For example, I would classify the recently proposed deal between NVIDIA and OpenAI as a case of (1), but this deal between AMD and OpenAI as (2). Namely, because I think it's clear that the chips act as well as recent advancements by Chinese manufacturers are threatening to NVIDIA's market-leading position and OpenAI investigating new vendors suggests they have suddenly become concerned with reducing cost of goods. Indeed, if both the leading Chinese firms and OpenAI were shown to be able to work with other vendors without sacrificing speed to market it would materially impact NVIDIA's stock price. AMD, on the other hand, is not trying to subsidize an existing client, but convince a market leader to take a risk.

The NVIDIA deal, then, suggests to me that certain limits have been reached in the industry, while the AMD deal does not provide me with any further evidence as to the existence of a bubble.

I’ve seen a lot of confusion about this type of deal recently: notably, it is often taken to imply something more or less shady is going on. I’m not sure when such arrangements became a thing, but I know equity stakes have long been an important part of enterprise SaaS deals. The reasoning is relatively straight-forward: if a large client commits to a vendor in a way that holds some risk to the former and will materially impact the latter’s business — and especially if the client’s support of said vendor will directly or indirectly benefit their competitors who might also use this vendor — an equity stake is a way to offset risk with upside.

You can see this play out in the history of OpenAI. NVIDIA supported them from an early stage and in exchange received OpenAI equity to offset the risk. Now from a position of relative strength, OpenAI has become concerned about vendor lock-in and so is rationally exploring AMD. Yet, because any such deal will materially impact AMD’s stock price and there is risk both of losing time trying to train with new chips as well as of benefiting competitors if they work with AMD to improve their hardware offerings/APIs, it is reasonable to ask for equity upside. So, for the same reasons (increase in stock price and enterprise client who will help improve their product offering) only without risk, is it understandable why AMD would want to offer equity on such favorable terms.

TLDR; My sense is that the sudden skepticism towards this relatively common enterprise deal structure seems to derive from the understandable interest in identifying signs of an AI bubble. Such a bubble may (and indeed almost certainly does) exist, but I don’t think this is evidence thereof.

----

EDIT: I'm just clarifying something I saw in a lot of responses. My only point is that it is important to try and empirically tease out what represents: (1) a circular deal in which vendors facing the limits of growth are subsidizing vulnerable clients; versus (2) a risk-hedging deal in which a non-market leading vendor offers upside to a market leading client.

I believe the recent Oracle and NVIDIA deals are cases of (1) that provide evidence of an AI bubble, but that this AMD deal is most likely a case of (2) that provides no further evidence.

I agree with you on the question of extrinsic values and do not envy people who are starting college right now, trying to make decisions about an extra-ordinarily unclear future. I recently became a father and I try to convince myself that in eighteen years we'll at least finally know whether it has all been hype or not.

However, on the intrinsic value of these new tools when developing habits and skills, I just think about the irreplaceable role that the (early) internet played in my own development and how incredible an LLM would have been. People are still always impressed whenever a precocious high schooler YouTubes his way to an MVP SaaS launch -- I hope and expect the first batch of LLM-accompanied youth to emerge will have set their sights higher.

(Almost) nobody who does a humanities PhD is doing so for a job. It's wrong, I think, to consider that simply idealism: there still do exist people who consider writing to be a vocation and that their life would be intolerable if it isn't what they pursued on a daily basis. Rationally -- and conscious of the "opportunity costs" -- such a one should seek the best apprenticeship possible, which is really what a humanities dissertation comes down to. I know many more people who pursued STEM PhDs more or less for a job -- and so, in my anecdotal experience, I would say the outcomes for friends who received their doctorates in the humanities are, measured by life satisfaction, greater than those who only at the end of it all realized STEM post-docs are miserable and that their academic programming skills aren't quite up to Silicon Valley standards. It's easy to forget at Hacker News that most life decisions these past few generations that didn't amount to getting an engineering job at a high-growth startup were much closer in outcome to a humanities PhD than retirement at 35.

There was a brief period with the dramatic expansion of the university system following World War II during which the need for bodies to teach introductory classes to auditoriums of uninterested students briefly matched the organic production rate of scholars. This period is certainly over. However, I'm not sure that's a bad thing for the humanities. In fact, it's only a matter of centuries in which formalized PhD programs were considered a prerequisite to becoming a researcher at all -- and not even in all Western countries during that time. In Italy, for example, the highest degree was a "laurea" until the 1980s, which was the product of only a five-six year bachelor's program. Humanistic research was largely published by presses outside of the university and so those who for whatever reason wanted to be scholars found a way to support their life, often editorial positions or teaching in high school, and simply got to work, struggling to make their research of interest enough to be published. This system did not at all negatively impact research outcomes and, measured by the numerous Italian works from this period that are still being translated, perhaps even improved them.

TLDR I'm not happy with the context in which the most recent changes are being made to the university, but I think it will be a net good if scholarship in the humanities becomes less sequestered from society -- and especially if many of those who might have sought to teach at the university level instead decide to teach in high schools.

Triples are fantastic for information retrieval, but I think if there's any takeaway from the unexpected success of LLMs it's that AI researchers historically undervalued language as such. Early symbolic approaches to AI retrospectively appear torn between reverence towards and hatred of language: on the one hand, a sensible skeptical doubt that language is within reach of software systems; on the other, a belief in the inadequacy of language in the unambiguous representation of knowledge. This paper just seems to confirm that, at least at the training level, the "problem of hallucinations" is not to be resolved by regression back to the various proposals to separate knowledge from its linguistic representation.

Again, this isn't to demonize symbolic AI or to say the answer isn't in the fusion of LLMs with knowledge graphs etc, but I think we now at least know that language is certainly within reach of software and that linguistic representations of knowledge are information-dense in ways we didn't previously anticipate.

Same. I note in-advance that I'm not sure whether you yourself are referring to use of LLM tools in your research or rather the results of your own domain specific application of deep learning etc -- here, I assume the former.

I feel like the common refrain of most LLM success stories over the past year is that these tools are of significantly greater help to specialists with "skin in the game", so to speak, than they are to complete amateurs. I think a lot of complaints about hallucinations reflect the experience of people who aren't working at the edge of a field where they've read all the existing literature and there simply aren't other places to turn for further leads. At the frontier, moreover, the probability that there exists a paper or book that covers the exact combination of topics that interests you is actually rather low; peer discussions are terrific, but everyone is time-starved.

Thus I find the synthetic ability of LLMs to tie together one's own field of focus with those you've never thought about or are less familiar with to be of incomparable utility. On top of that, the ability to help formulate potential hypotheses and leads -- where of course you the researcher are ultimately going to carry out the investigation or, in the best case, attempt to replicate results. Conversely, when I'm uncertain of my own conclusions, I often find myself feeding the best LLM I have access to the data I reasoned from to see whether it independently gets to the same place. I'm not concerned about hallucinations because I know there's nobody but me ultimately responsible for error -- and, at the fringe of knowledge, even a total fabrication can inspire a new (correct) approach to the matter at hand.

I think if I had to succinctly describe my own experience it would be that I never get stuck any more for days, weeks, months without even a hint of where to turn next.

Related, there's an ancient Palantir blog post (2010!) that always stuck in my memory about a chess tournament that allowed computers, grandmasters, amateurs and any combination of the above to enter [0]. At that time, the winning combination turned out to be amateurs with the best workflow for interfacing with machine. The moral of the story is probably still true (workflow is everything), but I think these new tools for the first time are really biased towards experts, i.e. the best workflow now is no longer "content neutral" but always emerges from a particular domain.

[0] https://web.archive.org/web/20120916051031/http://www.palant...

To be clear, I'm not 'missing' that -- I'm saying it's part of a trend since the 1990s ("globalization") that, for various reasons, people believed stopped in 2016/2020.

I don't disagree, but this has been the trend since the 1990s. This was the whole question of "globalization" and all the debates about NAFTA etc. The only anomaly was that people came to believe (in 2020? maybe even 2016...) that "globalization was over" and that now we saw the reverse trend. That seems to be leading people to now believe that this offshoring is something new, when the most likely interpretation is that it never went away.

None of the above, however, begins to touch the question of whether AI has, is or will introduce an entirely new dynamic into the equation. I'm not sure myself of the answer to this question, but I believe it's a very important one and shouldn't be dismissed too quickly -- certainly not by the continued presence of an undeniable tendency of labor markets that has long been observed.

Of course it’s not black and white: it’s the very fact that there are so many possible explanations of the current changes that makes pretending data supports a given explanation more than another when it doesn’t so misleading. There’s no doubt that offshoring and H1-B hiring is an important trend — as it has been since the 1990s, NAFTA etc. We don’t need this article to tell us that. The difficult question is whether AI is actually introducing a new dynamic into the situation. The article suggests it isn’t. All I’m arguing is that, whether or not the conclusion is ultimately true, the data presented does not shed light on this important question.

Apologies if my presentation is inadequate, but I’m not suggesting that stock price is a proxy for headcount. I’m simply arguing that you have to look at the two together in the context of a particular firm — for exactly the reason you note, ie NVIDIA has a lower head count per share etc. It’s the ratio that matters and is the best proxy we have for inferring the cause of a shift in hiring. Lower growth of headcount and stock price going down suggests the cause of the former is reduced expectation of return, ie particular industry or general economic slowdown; lower growth of headcount and stock price remaining constant or rising would be more suggestive of (not causal proof of course) of an AI-related effect.

If we don’t have the data we shouldn’t argue that the data we do have says what it doesn’t. I actually agree with you and the author that trying to figure out what’s going on in labor markets right now is incredibly important — but for that very reason we have to be careful not to accept too easy conclusions. There’s a lot of information we do have about firm expenditure on labor within public filings — I’ll actually try to pull some figures this weekend, can let you know when I do if you’re interested. I think the interesting question on the AI side is whether it allows more firms to present a ratio like NVIDIA, ie reducing expenditure on labor that is considered a “cost center” in a particular industry.

I’m not describing what’s happening right now — I’m saying the data shown in the article is insufficient to ground what it claims. The testimony you mention better supports the argument, but also is insufficient to show that this shift to H1-B is not in addition to a reduction of overall projected headcount due to AI.

I think this interpretation of the data rests on a misunderstanding of how a firm conceives of headcount. You saw this a lot during Covid layoffs when the press tried to point out the contradiction between hiring X number of people at the same time as firing Y number of people. Wasn't the firm just staying in place? Yet, in fact, the firm had planned to hire Z people that year, a number much greater than X, greater still than X+Y, and likely greater than Z_t-1, i.e. the number hired the previous year.

A firm's headcount is a dynamic value that is reflective of a rate of growth: its headcount is going up by x% each year, ideally linear growth or at least not a higher rate of growth than revenues or free cash flow. It requires an accounting perspective on the inflows and outflows of an ongoing process and not just a small slice of data. If you just look at that firm's hiring and firing in a particular year, subtracting the one from the other, you would appear to have shown that the firm was more or less simply replacing the other with the one. Yet, the main question for actually measuring the (firm and investor expected) impact of AI is whether that rate of headcount growth has changed without the stock price going down. That is, is it now proposed by management and accepted by investors that the firm will henceforth require less human labor in order to deliver on the present expected value of future free cash flow. You cannot infer anything about this question from the data shown.

There certainly is a separate question of whether a firm has changed its source of new hires, but that also isn't reflected in the data, which just shows approved visas against layoffs. For this, you would need to look at net hires year after year and prove that a growing percent thereof are H1-Bs.

There's an ambiguity in the title, reflected in some comments below. It can be understood either as the claim that "in a particular human being, to be intelligent as measured by IQ means that you are more likely to be autistic", suggesting for example a trade-off between social and general intelligence; or the claim that "the evolution of the human brain and so human intelligence as such, which characterizes both those of low and high IQ, entailed those genetic shifts that made autism a possibility for our species but not other primates." The paper argues a form of the latter.

Completely agreed. Whether it be AI or otherwise, I consider anything that gives me more time to focus on figuring out the right problem to solve or iterating on possible solutions to be good.

Yet every time that someone here earnestly testifies to whatever slight but real use they’ve found of AI, an army of commentators appears ready to gaslight them into doubting themselves, always citing that study meant to have proven that any apparent usefulness of AI is an illusion.

At this point, even just considering the domain of programming, there’s more than enough testimony to the contrary. This doesn’t say anything about whether there’s an AI bubble or overhype or anything about its social function or future. But, as you note, it means these cardboard cutout critiques of AI need to at least start from where we are.

Thanks for the tips! It's really interesting going back to that point of indistinction in these languages where even the proximity to old English is so strong. I've been hoping to try and read some of the old sagas in the process, but that's probably a bit too far back -- I'll have to see if there are some good century old translations thereof, as you suggest.

The phonetics, on the other hand, are presenting some challenges...

Yeah, language learning is such a wild thing. I have a five month old son now and I'm just getting to see the process of language acquisition firsthand. He'll be going to early school (eventually) in Danish, but probably won't end up living here forever, so curious if it'll stick with him through life or not.

What's your mother tongue out of curiosity?

Good luck with your studies!

I'm sure there are, but I can personally attest to the quality of those two. The French one in particular is just astonishingly well-executed that I would recommend looking at it if you were interested in techniques. The "magic" of it is in the composition: each chapter, you read a little description of a new grammatical concept, you work through various related sentences with help, and suddenly you're reading a whole French text composed of those past sentences and able to answer comprehension-related questions. It just builds and builds like that to great complexity.

What always makes learning to read easier is that time is completely in your control. The principle is pretty straightforward: if you have enough time and patience, you can read anything (with a dictionary, grammar book etc) and the more you read in that language the less time it starts to take. These books basically just bootstrap the process.

I mentioned it above, but the other way is if you have a book or article you really know well in your mother tongue that exists in a language you want to learn, just patiently try and read it in that language. I think programmers actually have a bit of an advantage in this, as it's really just pattern recognition -- and it isn't that different from trying to understand a program in a language you haven't worked with before.

I just had a kid and moved to DK from the US. I was so skeptical of the infant (and often mother!) in cart on front of bike that you see everywhere, but it's such a life-changer versus the ordeal of carseats etc. TLDR totally agreed.

I understand both your historical question and the more concrete practical one. Separating reading comprehension as a skill from all the other discrete functions of a language is very straight-forward [1] and, in fact, there are some good analog resources for this.

For French: Dandberg and Tatham, French for Reading

For German: Jannach, German for Reading Knowledge

I've used both and swear they're magic, especially if you're trying to learn to read in a scientific domain that you're already a specialist in (versus literature).

Once you've sort of "learned the game" it isn't very hard to do a similar process for other languages on your own. Then, my main recommendation is to take a text you're deeply familiar with in your native language or English that exists in X other language and just go ahead and start reading it with a dictionary. It starts slow, but progress is very very fast if you stick with it, especially compared to learning to speak or even just listen to a language.

For life reasons, I've found myself having to learn Danish, so I'll let you know if I figure out any good resources for Scandinavian languages.

[1] The only downside I've encountered is trying to later learn to speak a language I had been reading for a while where overcoming the sort of "fictitious phonetics" that existed in my head proved problematic.

The article hints at it, but my guess would be this investment is intended towards Intel foundry and getting it to a place where NVIDIA can eventually rely on them over TSMC — and the ownership largely to give them upside if/when Intel stock goes up on news of an NVIDIA contract etc. It isn’t that uncommon of an arrangement for enterprise deals of such a potential magnitude. Long-term, however, and without NVIDIA making the call that could definitely have the effect of leading to Intel divesting from directly competing in as many markets, ie Arc.

For context, I highly recommend the old Stratechery articles on the history of Intel foundry.

I have a similar relation to AI with programming -- and my sense is very many HN readers do as well, evidenced not least by the terrific experience report from antirez [1]. Yet it is rare to see such honest and open statements even here. Instead, HN is full of endless anti-AI submissions on the front page where the discussion underneath is just an echo chamber of ill-substantiated attacks on 'AI hype' and where anything else is down-voted.

It's what is, to me, so bizarre about the present moment: certainly investment is exceptionally high in AI (and of course use), but the dominant position in the media is precisely such a strange 'anti-AI hype' that positions itself as a brave minority position. Obviously, OpenAI/Altman have made some unfortunate statements in self-promotion, but otherwise I genuinely can't think of something I've read that expresses the position attacked by the anti-AI-ers -- even talk of 'AGI' etc comes from the AI-critical camp.

In a sense, the world seems divided into three: obvious self-promotion from AI companies that nobody takes seriously, ever-increasingly fervent 'AI critique', and the people who, mostly silent, have found modern AI with all its warts to be an incomparably useful tool across various dimensions of their life and work. I hope the third camp becomes more vocal so that open conversations about the ways people have found AI to be useful or not can be the norm not the exception.

[1] https://antirez.com/news/154

I had to double-check my source and I realize the error was mine. “40% read at or above an eighth grade level” is the correct description of the state-level data I was looking at, which is distinct from the general metric used for literacy.

This led me into a bit of a rabbit hole trying to understand what in fact the official literacy rate is measured by if it’s so wildly different from - indeed almost double — the portion who can read at an eighth grade level.

The data is actually quite interesting. US National Center for Education statistics administer tests to assess “the ability to understand, evaluate, use, and engage with written texts to participate in society” and an individual falls into one of five categories. Official literacy definition considers above category one (“below basic”), but it is category three that maps approximately onto “eighth grade knowledge” (thus four as high school, five as post-graduate). The most interesting thing I found in the data is exploring that gap between two and three, ie states that have a high attainment of official literacy but then very low rates of the higher levels. California, for example, has the highest percentage of people below level two, but a relatively high percentage at level three and above — obviously I haven’t considered the data for long enough to conclude, but that suggests to me largely a question of immigration/non-English speaking populations. The state I’m from does better than California on attainment of level two, but significantly worse at attainment of three or above.

States where level 3+ > levels 1-2: District of Columbia, Washington, Minnesota, Oregon, Massachusetts, North Dakota / Utah / Colorado.

States with lowest level 3 (ie “eighth grade” equivalent) attainment: Mississippi 35%, Louisiana 35%, West Virginia 37%, New Mexico 39%, Nevada 39%, Alabama 39%, Arkansas 39%, Texas 40%, Tennessee 40%, Kentucky 41%.

TLDR that gap looks like an interesting way to separate issues with a state’s educational system from other questions. Whatever the best measure of literacy may be, it seems like the bar should be a bit higher than just “native speaker of the measured language.”

There are more divisions than just “data” vs “vibes.” After all, even in the natural sciences, the best data is useless without an explanation/hypothesis that can never just be reduced to the data. Precisely what is thrown out in the decline of reading is familiarity with the centuries of hard-won Enlightenment knowledge, especially concerning the stakes of education, which isn’t just vibes and that ought precisely drive our further data-driven insight into these questions.

Second best, however, I’d take the “vibes” of a random teacher over the religion-based decision making that seems to be on the rise in the US. “Data-driven” religiously motivated educational policy is the worst of all possible worlds.

I’m originally from a US state that currently sits at a 40% literacy rate, but I’ve lived for the last decade in various European countries. I say this only because, even if still anecdotal, I feel like I have a decent basis for comparison. Certainly there are educational disparities from center to periphery and across income brackets everywhere, but I have never lived somewhere that the division was as stark as the US.

France — with all its problems — ensures the same incredibly high standard of curriculum across the country and perhaps most importantly it is actually expected that top university performers who will become researchers teach at high school in the periphery. It’s even a nation-wide competition by discipline (look up the “aggregation”) to obtain these highly sought positions. The idea is something like you teach high school outside Paris while preparing your doctorate and then either return triumphant to the big research institutes or continue teaching in the provinces. Something like this in the US would have immeasurable impact, since probably one of the biggest issues is just convincing well-educated people to teach in rural areas.

I’m from a US state with a 40% adult literacy rate (=above eighth grade reading level). At least there, none of those three things are even close to the root causes. The average school in the US outside of the big cities, especially the farther you get from the coasts, is just not fit for purpose — and funding only seems to ever go down (not that throwing money at the problem alone would solve it).

Honestly — and I’m not being at all utopian/overvaluing the present state of the technology — I think AI is one of the few prospects for even just marginal improvement, especially since it’s accessible by phone. Much as I wish it wasn’t the case, it’s hard to even imagine all the things that would have to change (from funding, to legislation, undoing all the embarrassing “teaching the controversy” curriculum, to say nothing of staffing) for a “non-technical solution.”

For a similar project, I worked with GPT to create an extensive dataset of translations from a historical language. I could then use this both to evaluate base capacity of other models in the language, i.e. giving the model the task of translating the various passages and evaluating the results with GPT, as well as for fine-tuning.