Sure, but that's not the "just a solution in search of a problem" claim made in the article. I don't think anyone including Ted Nelson has ever claimed that Xanadu was sufficiently complete at any point.
I think the prices are still really modest if you look at them in perspective. One of the small, irreplaceable stock of surviving PDP-10s could be had for less than the cost of a mediocre used Ferrari.
In general I have about zero enthusiasm for trying to find defensible interpretations of things that Ed Zitron said, and I generally agree that the name of Zitron just largely needs to stop coming up in anti- and anti-anti-AI arguments since, it seems, he's just not a particularly insightful or reliable voice on the subject. That said, one or two of the specific assessments in Luu's article seem dubious as well, especially this one:
It was my understanding—and I'm no expert, so if someone does know better please correct me!—that indeed by the second half of 2025 training, and also post-training reinforcement-learning stuff, both hit seriously diminishing returns, and the thing that is continuing to scale well or pretty well is inference. See eg. https://www.tobyord.com/writing/mostly-inference-scaling . And in fact in the quoted and linked article https://www.wheresyoured.at/how-to-argue-with-an-ai-booster/ Zitron comes up with something which looks like a recognisable explanation of this:
> Because model developers hit a wall of diminishing returns, and the only way to make their models do more was to make them burn more tokens to generate a more accurate response (this is a very simple way of describing reasoning, a thing that OpenAI launched in September 2024 and others followed).
> As a result, all the "gains" from "powerful new models" come from burning more and more tokens.
AFAICT the other drivers of recent progress in LLMs have been: ploughing in lots and lots of specialised training data custom-made at piecework websites https://www.youtube.com/watch?v=4pG3SJQPAwk ; and work on harnesses and the like. AFAICT neither of those makes false the claim that "[t]hese models have clearly hit a wall where training is hitting diminishing returns" either. Similarly, even if some big new advance does cause training or post-training to start scaling like gangbusters again in 2027 or 2028 that wouldn't make the quoted statement clearly wrong: Zitron would clearly like you to infer that there won't be any further big advances soon in LLM training, but the quoted statement doesn't clearly make that claim. (Even if he had made that claim, and it did turn out to be wrong, it would be a relatively forgivable error, more on the "cloudy crystal ball" than "misstates currently known facts" end of the spectrum.)
So: it seems that Luu took a fairly specific, objectively judgeable claim from Ed Zitron; and that claim was ... correct?; and Luu instead rated it "Wrong" without further elaboration. It seems that Luu interpreted the quoted claim as saying something like "model progress has ceased"; but it seems that's not what that specific claim (as opposed to whatever other things Zitron has said at other times and places) said.
>It was my understanding—and I'm no expert, so if someone does know better please correct me!—that indeed by the second half of 2025 training, and also post-training reinforcement-learning stuff, both hit seriously diminishing returns, and the thing that is continuing to scale well or pretty well is inference.
I'm not an expert either, but while I do think for a bit it looked like ~all the improvement was inference-time scaling, it hasn't stayed that way. Mythos/Fable is likely a very large model (ex: it knows many things without searching) and this is probably part of its high level of capability, and the companies have started doing very large amounts of RL (which in OpenAI's case led to the HF attack).
> and the thing that is continuing to scale well or pretty well is inference
No, the models are just more intelligent. GPT 5.6 Sol can do more in fewer output tokens than any model from late 2025. Test-time compute isn't the only lever the labs have for scaling. This is among the two major things Ed has gotten laughably wrong in his technical predictions (that TTC was the last resort to make models better, and that synthetic data wouldn't help)
My understanding is that RLVR, synthetic data generation and a slew of other post-training techniques are what have driven many recent advances in models more so than manual data providers. The economics of that are for sure worse than just scaling pre-training but it is incorrect to think that test time inference scaling and manual data entry are the only ways in which models are advancing.
> Reals numerical systems are much more diverse than that. For instance, Babylon used a base 60 (hence our minutes and hours). But there were much less than 60 numeric symbols, since they had a symbol for 1 and 10. So 2 symbols for a base 60!
And except at the beginning, the Mac Minis of the past weren’t famous for being cheap, but rather for being disappointingly, unattractively expensive and, over time, apparently increasingly forgotten by Apple.
What connection if any is there between pre-1946 German work on magnetic amplifiers and magnetic alloys and the pre-1946 German leaps forward in audio tape recording onto Magnetophons?
> If you read the article, it's ideologically consistent - the main motivation in abandoning Fraktur was strategic (i.e. military), not ideological.
But that isn't quite what it says either. Pragmatic considerations may have given the decisive push, but it seems that Gothic script had always been in a precarious position on the faultline between the archaising and modernising instincts of Naziism. The specific tall story that was eventually presented about how it was Jewish was quite clearly a low-effort post-hoc justification, but it seems it was always quite easy to take an anti-Fraktur stance from a Nazi perspective: see A. Hiedler, op. cit. And even just going on the small extract provided from that Reichstag speech https://penelope.uchicago.edu/encyclopaedia_romana/luftwaffe... , it seems that Hitler's objections weren't just pragmatic: before he gives a practical reason to drop Fraktur, he attacks Gothic revivalism in general on aesthetic grounds, no small matter to the Nazis. My assumption is that his objection to Gothic revivalism was based partly on a general aesthetic instinct that too much revivalism is cosplay and naff; and partly on a disapproval of medieval Germany specifically as a model or claimed precursor for Nazi Germany, for various reasons. A possible third, type-specific, factor, mentioned in https://penelope.uchicago.edu/encyclopaedia_romana/luftwaffe... , is that compared to "Atiqua" "Fraktur" looks vaguely kinda more Hebrew-ish. That could plausibly have been the underlying sentiment behind the (apparently completely false and baseless https://worldcrunch.com/culture-society/nazi-typeface-fraktu... ) story spun about Jewish printers in Schwabach.
That's why I said "main motivation", i.e. what actually caused them to follow through. The initial idea wasn't ideological either - not following from ideology doesn't mean it's pragmatic (in fact I'd argue those two aspects are orthogonal) and that it wasn't born out of pragmatism only reaffirms my statement that the Nazis were opportunists, not pragmatists.
A lot of things boiled down to post-hoc ideological justifications for the personal preferences of Hitler and his ilk. Even the racism was incoherent - it's cliché to point this out but Hitler himself infamously looked nothing like an Aryan übermensch despite the Führer cult treating him akin to a divine figure. This is consistent with research indicating disgust as a strong motivating signal in right-wing politics.
Plenty in Ireland as well. The craze certainly seems to be more subdued here, though some of the posh delis in Dublin, in particular, seem to be riding the wave. ( https://www.reddit.com/r/CannedSardines/comments/1o9dhlg/any... , for any Dublin tin-hunters.)
Right, it's clear that nVidia is taking care and trying to position itself so that it can continue making sales if and when inference goes local. And it's in a much better intrinsic position to do that than the LLM SaaS vendors are: nVidia sells shovels to the army, but it also knows how to sell shovels to Walmart. Whether the financial relationships that Huang's got his company into will cause it problems if the market shifts is a different question, though.
reply