Hacker Newsnew | past | comments | ask | show | jobs | submit | achompas's commentslogin

One thing I’m observing in these comments is a willingness of folks to project their own predictions onto Ed’s statements when validating their plausibility. Eg. “I think he’s wrong about the timing but I do expect AI companies to go to zero.”

You can do that, but then you’re no longer discussing his predictions. You’re discussing your predictions, and your own positioning.

Those differ from Dan’s essay, which engages with the literal text of Ed’s numerous predictions during 2024 and 2025 which are demonstrably invalidated by their measurable outcomes.


>You can do that, but then you’re no longer discussing his predictions. You’re discussing your predictions, and your own positioning.

Depends if you care about the "prediction" part or if you care about the assessment of the situation (regardless of date).

If someone in 2000 said "the subprime mortgages market is a bubble and will blow no later than 2003", they got the prediction wrong, but their assessment would be right.


Predicting things 5 years too early is often as useless as not predicting anything.

You can say "AI will be able to _____" and be right 99.9 times out of 100, but the question is when.

You can say "The AI market will go to 0" and be at least directionally right eventually.

But none of it matters if you get the timing wrong.


>Predicting things 5 years too early is often as useless as not predicting anything.

It's not the predicting that's important to me, it's the assessing.

I don't care if it will do X in 2 or 5 years.

I do care if something IS "X", and we should correct course to not get the bad outcome associated with X (whenever it might come) or not.


> I don't care if it will do X in 2 or 5 years.

Yes, but for should care if it's 2 years vs 50 years. That's why timeline are important for predictions.

Anyone can predict lots of things that given enough time will eventually be true. "The sun will burn out", "humans as a species will no longer exist", "the US will collapse". That's useless. But those become meaningful if they will happen within the next three years.

It's the reason the saying "the market can be irrational longer than your can stay solvent" has such important meaning.


> Predicting things 5 years too early is often as useless as not predicting anything.

It depends on what type of thing you are predicting and why. Even something as simple as "Enron will implode" in 1995 has very different levels of usefulness if you are planning to short a stock versus figuring out what stock to buy and hold in your retirement account.


It would matter if we had politicians who acted before bubbles collapse instead of bailing out the perpetrators after the fact.

That will successfully prevent 200% of market multipliers: bubbles and just plain sudden rapid value creation.

It's like cutting out the prostate from all men at age 30 to prevent nearly 100% of prostate cancer.


>That will successfully prevent 200% of market multipliers: bubbles and just plain sudden rapid value creation.

So a win-win. Market multipliers rarely end up well, bubbles or not.

>It's like cutting out the prostate from all men at age 30 to prevent nearly 100% of prostate cancer.

It's more like cutting steroids from all men, and letting them build strengh naturally over time by training.


If actions are extreme, perhaps. But that doesn't have to be the case.

We have made efficient frictionless markets a holy cow. The appropriate response is not to eliminate them.

The better part of "rapid value creation" is not the rapidity.


Take a break from LLMs.

>Predicting things 5 years too early is often as useless as not predicting anything.

I dont think Zitron has predicted anything 5 years too early. 1 or 2 at the outside is probably the limit.

I remember during the original internet boom the people who were given the most shit were the ones who predicted the bubble popping a year or two early.

Zitron fits quite neatly in this category: most of the predictions of his Ive seen that are "wrong" are things which could still happen and things where he gave wrong timelines.

Realistically identifying a bubble and the stupidity associated with it just requires an ability to do the research and see through bullshit. identifying even roughly when it will pop really needs a crystal ball. Mass human delusions do not collapse in a predictable way.


My recollection is a couple years before it popped everybody (except tech CEOs going public or investing tens of millions in dog food delivery infrastructure) was in disbelief it hadn't happened yet

yeah I also remember a lot of talk about the bubble for 18 months to 2 years before it actually popped.

i think we're in that (quite long) stage with AI right now where most halfway smart people "know" but they're still invested due to the greater fool theory and still talking their book.


I'm invested because I "know" it'll pop and I also "know" it'll eventually work and revolutionize everything.

What do I have to lose? I'm just going to get decrepit and die anyway if it doesn't work out.


The subprime market changed quite a bit in size and how much was securitized in the run up to 2007, so not sure a prediction in 2000 for a 2003 event would have been easily transferred to what happened later. Would really come down to what specifically the prediction was based on for it to be a bubble in 2000.

Excellent point. We can look at the reasons for the prediction as well as the result in assessing its value. It seems hard to do in practice unless you are evaluated by someone who makes better predictions, but it would be interesting.

> If someone in 2000 said "the subprime mortgages market is a bubble and will blow no later than 2003", they got the prediction wrong, but their assessment would be right.

It doesn't make sense to split it though. Their assessment is that it's a bubble AND that it will blow no later than 2003. It's a single statement.

And it matters, because if all you're doing is saying there's an AI bubble then it's harder to prove you wrong but you also don't stand out and won't get a lot of credit for it. A very large number of people are saying the same thing as you, so who cares.

People like Zitron stand out because they go further than others and make detailed statements. Which happen to be wrong.


Zitron predicted the downfall of Oracle as someone mentioned below. He also predicted that the overhyped data center construction plans (Project Stargate, repeated vague Nvidia pledges) would not materialize.

If he got MSFT's cloud revenue growth wrong for this year, how much of that is selling shovels to OpenAI and how much is circular?


Why is Zitron’s repute evaluated entirely on the basis of failed predictions? Predictions are incredibly hard. AI enthusiasts and thought leaders have made so many demonstrably incorrect predictions it’s hard to keep track. Based on this metric, Altman and Amodei should never be taken seriously again.

Years back John C. Dvorak talked about predictions. He's was being ridiculed for his comment that there was "no evidence that computer users would want to use a mouse" (which was sort of true at the time). One of his point was that he had made a crazy amount of predictions on various topics, some came true, many didn't. People just remember the one you got right, and the ones you got horribly wrong.

Ed Zitron is just one AI crash away from being known as the guy who saw this coming. Everything else he has said can be completely wrong, he just needs to be somewhat correct on a minor crash.


Correctly predicting nine out of the last five recessions.

> Ed Zitron is just one AI crash away from being known as the guy who saw this coming. Everything else he has said can be completely wrong, he just needs to be somewhat correct on a minor crash.

That's the theory, but I feel like enough people have heard of him by now to be aware of the numbers game being played.


I don’t agree with your conclusion. It’s like Michael Burry, he’s gotten 10 of the last 2 economic crashes right.

How many people in the financial space recognized Burry's name before 2008?

Pretty much none, but he’s been saying the sky is falling for the last decade.

Well yes, that's the point. Burry had his success early and was not notable before that. Zitron has been outputting a continuous stream of bearish sentiment that hasn't come to fruition, from the moment anyone recognized the name.

And even then, people are starting to notice that Burry's gone a while without a noticeable success.


Have you ever listened to Zitron speak? He's not exactly the type to hedge his predictions behind careful language about how hard predictions are. He makes every one of these predictions with absolute confidence and conviction.

He speaks with all the conviction of a classic "Fire and Brimstone" preacher. Someone for whom the only metric is stirring emotion, not accuracy of statement.

Apparently that's necessary and sufficient for getting views. In these interesting times when the future is very uncertain we like to hear people speaking with certainty about the future. Especially people who appear to not be paid for it.

Can he predict the future? No. Can he pretend to be able to predict the future? Yes.


The first time I heard of Ed Zitron was 18 or so years ago when he wrote a review of the Darkfall MMORPG. He tore it to pieces, but as someone who played the game I could tell he had never actually played it, which is what the game developers also claimed when they reviewed his logged in session, that he did nothing for 30 minutes then logged out (I might be a bit off on the details, it has been awhile). He’s always been someone who wont let the truth get in the way of his success. I really hope people start valuing accuracy over confidence

So...Ellsworth Toohey?

So does everyone else, including prominent AI enthusiasts and tech CEOs. And most of those predictions are wrong, because predictions are really hard.

But this doesn't excuse Ed making a living from awful predictions, surely?

If you realise you're bad at something it seems immoral to keep charging people for it.


So Altman and Amodei should retire?

Altman and Amodei do not make a living out of that. They make a living out of being CEOs of OpenAI and Anthropic. Whether they are doing well in this position is a different question, but that has little to do with their public predictions.

Nah they do, Sam does nothing but lie in an effort to keep up share price. Selling people an impossible future is the job of OpenAI CEO. If tokens were profitable it'd be selling tokens, but they aren't so the only other option is selling ridiculous predictions to investors

What? The only reason their companies make so much money is BECAUSE they make their predictions about AI replacing every job soon. A couple months ago it seemed is was their ONLY job.

The only reason Anthropic and OpenAI make money is by saying AI will replace workers? Not from selling access to their models?

Neither company is remotely close to making a profit yet. To the extent people at those companies make money, it is unrealized capital appreciation driven by successively higher valuations during successive funding rounds.

In short, financially they are still running on hype. Quite a lot of operations are being paid out of capital, not revenues. So they have to keep selling a dream to keep raising capital. Next stop: IPO.


How do you think they convince people, esp. CEOs to buy access to their models?

I was thinking about this as I walked earlier.

People who "make predictions" are like psychics. Unless those predications are uncannily good – in which case we call that person an analyst – we can assume that the person is either wilfully lying, or that they're stupid.

Ed's predictions have so far failed to come true. I don't think he's stupid (but I might be wrong). That leaves me feeling like he's the psychic who knows exactly what they're doing. They're telling a gullible audience a story they want to hear because it's a nice way to earn a living.

Comparing him to the CEOs of companies who produce incredibly useful products used by millions is facile nonsense.


> they're telling a gullible audience a story they want to hear because it's a nice way to earn a living.

Sure sounds a bit like Altman though ultimately, the audience being VCs


The difference being that the story he's telling is "you'll make money" and he's usually right.

Ah that's why they're losing billions of dollars a month...

Citation needed

The difference is that Altman and Amodei are doing something else besides making predictions that turn out to be wrong. Most obviously, they're running tech companies that are changing the way entire industries operate. We put a lot of weight on that in evaluating them.

What else is Zitron doing that he could be evaluated on, besides making predictions that turn out to be wrong?


Predictions being hard is what makes them valuable as a check on insight and accuracy!

Ok, but it also means we shouldn’t immediately discredit people who make incorrect short-term predictions, because those are really hard and no individual person consistently gets them right. You can worship at the altar of short-term prediction if you really want, but if you want to be logically consistent, that entails never taking anybody seriously ever.

And we're used to insulting and making fun of the prominent members of those circles too, see: "Scam Altman", the large variety of jokes about Musk's timelines, comments about Dario waking up in a cold sweat whenever a powerful new open weight model drops, the "AGI achieved" meme etc.

Hyperbolic as those descriptions are, it’s not like they’re not sourced from reality. Sam Altman’s reputation in SV is well-established; Musk promised the roadster what, a decade ago? These guys preach a promised land of milk and honey and so many hear lap it up like dogs.

I ask again, have you ever heard Zitron speak? The way he presents himself has absolutely nothing in common with any tech CEO or AI enthusiast I have ever seen.

Both can be bad, and I think people take issue with intentionally projecting false or irresponsible confidence (lying).

Zitron twists and misreports a lot of facts where they should know better.

CEO hypemen knowingly conflate Ambitions for certainty


I think it's fair to say predictions that Altman and Amodei make should never be taken at face value, as well as Zitron. That's fine. But that doesn't have any bearing on Dan Luu's claims. This feels like an example of what he talks about in the article when saying that people respond to his claims by pointing at something entirely different. That is to say, whether AI enthusiasts make silly predictions doesn't mean that these companies aren't going to be profitable, or that any of Zitron's predictions are any good either.

What else would his repute be based on? Maybe telling retrospective truth? Per the article, he does terribly by that metric as well.

I thought in the post Ed was being evaluated based on all his predictions, and it turned out all of them were wrong.

Most people will read the post as going over all the falsifiable predictions and none of them panning out, since after the chronological prediction list it says "After this point, most further predictions that I saw were either non-falsifiable or resolve in the future".

If Dan is reading maybe he can clarify?


The list is chronological; as it nears the present, the predictions actually refer to the future, and are thus not yet resolved.

Yes it's chronological but no it doesn't have unresolved items nearer the present:

The last 5 ones in the list read "Wrong" and seem reasonable. The first thing after the list is "After this point, most further predictions that I saw were either non-falsifiable or resolve in the future".


Isn't that pretty much his whole thing, confidently telling us what will happen? Of course his reputation should suffer if his predictions are wrong, as should those of the others whose correctness : confidence ratio is too low. (But if their reputation/power/relevance mainly comes from things other than their punditry, we can't really stop 'taking them seriously' altogether.)

Because Zitron is specifically making a name for himself as a critic making bearish predictions; Altman and Amodei may have made unreasonably bullish predictions, but they've also done other relevant things like e.g. being involved in the actual development of the models.

Here's two possible set of predictions:

- a list of predictions that are entirely wrong, from A-to-Z, and are not even resembling what ends up happening

- a list of predictions that are wrong, but where the underlying points are in fact interesting and have some predictive value, and it's just the "last step" that is wrong

For example, one person might say "oh it's raining in Dallas therefore I should buy some TI stock". And we'll say for sake of argument that they say that even though it's nice and sunny in Dallas at the moment.

Another person says "Oh its raining a lot in Idaho and that is going to increase potato yields and therefore I will buy McDonalds stocks cuz fries will be cheaper". In this hypothetical it turns out McDonalds buys all their potatoes from ... Kansas or something instead (and it's a specific kind of potato in a completely separate market)... but Idaho potato yields _did in fact go up_.

An even more straightforward point: the iphone 3GS comes out in 2010, people are very hyped, someone looks at how RIM _still_ hasn't gotten its shit together and declares "RIM isn't going to to be able to stay profitable 18 months from now, they're gonna have their lunch eaten".

Turns out that RIM still made a healthy profit in 2010. and 2011. And 2012. 2013 was their first loss in a while... and then it wasn't until 2014 that they really got kicked in the face.

The prediction was early, overestimated how long of a tail RIM would experience, but how wrong was it? Was the prediction of some utility?

I'm saying this... it would be helpful if _some_ more AI companies flamed out. In some sense he does himself no favors by focusing on the corps with the biggest war chest instead of the various AI companies that spend a bunch to go nowhere fast and then have just disappeared.


> The prediction was early, overestimated how long of a tail RIM would experience, but how wrong was it? Was the prediction of some utility?

Wrong enough that the utility is seriously diminished. Predicting a specific quantity dropping to a specific level at a specific date is a lot more valuable than saying “those guys are cooked”.

And even if some minuscule utility existed: why should predictors be so coddled by their observers? We should be demanding more rigour from predictors rather than looking for new and creative ways to forgive them for their folly.


I think this is a legit read about the utility.

> why should predictors be so coddled by their observers?

I suppose that in the RIM example the predictor correctly identified causes and concerns. The timelines are off. But highlighting the causes are interesting to me, as a datapoint.

If someone says "hey there's a problem here", you can somewhat independently look at the problem and validate its seriousness on your end.

Some people in this thread are like "well the value is in the prediction", but from my perspective I'm thinking "the value is in the causal analysis". Because if you merely think the timelines are wrong you can easily just change numbers around.

Obviously Zitron is not "always right", and isn't getting all causal analysis right IMO. I do think he's bringing up things though.


If the causal analysis is the real value, then it could be presented without a prediction, or at least without a specific prediction.

In 2024, if Ed Zitron presented all his analysis of AI companies without the specific “model capabilities have peaked this year” style statements that Dan Luu scrutinized, that would be more honest and provide the value you are looking for.

Similarly for the RIM example nobody would have to claim that RIM would fail in a particular year, or that “RIM is cooked”, they could present their data about the present without any projection.


> Predicting a specific quantity dropping to a specific level at a specific date is a lot more valuable than saying “those guys are cooked”.

It's also way harder, and the added value isn't that great, if you're not interested in playing the stock market.

As an example, explaining why the 2008 crisis was structurally bound to happen is probably more valuable to a policymaker than knowing whether it would start in august or september.


The added value other than the stock market would be in making major life decisions.

In the case of the 2008 crisis, specific knowledge of the timing could matter for buying or selling a house.

In the case of technological predictions, specific knowledge of the timing could affect whether you want to take a university program in a given field or join a certain company.

If a predictor can’t help with picking stocks or making big decisions, then what are these predictions for, entertainment? Actually an HN comment on Dan Luu’s post about futurists really did suggest that futurists are practically entertainers.


> If a predictor can’t help with picking stocks or making big decisions, then what are these predictions for, entertainment?

Not sure how you constructed that specific sentence from my message.

My point was that most big decisions aren't that sensitive to small timing changes, like stock market can be. Your examples highlight that point: few would change their university choices depending on the knowledge that a future event will happen one year earlier or later.


> Not sure how you constructed that specific sentence from my message.

Fair enough, I didn’t.

> most big decisions aren't that sensitive to small timing changes

That makes sense for a single year at least.


If you're starting a 4 year degree, whether the bubble bursts tomorrow or 6 years from now isn't that important: what's important is whether you want to start a degree in the bubble's field.

Likewise if you're signing a 30 year mortgage, knowing vaguely that the housing market is precarious is still valuable even if you don't know if it'll collapse tomorrow or in 6 years: you'll be on the hook either way if you sign today


RIM is dying now -> useful for stock trading

RIM is dying in the next four years -> useful for life planning.

Anyway, i just wanna get on record that i predict an ai bubble pop event in the next 12 months.


> Anyway, i just wanna get on record that i predict an ai bubble pop event in the next 12 months.

If it happens on Sep 2nd 2027 you'd be wrong though.


Next 12 months: Oct 2026, Nov 2026, Dec 2026, Jan 2027, Feb 2027, Mar 2027, Apr 2027, May 2027, Jun 2027, Jul 2027, Aug 2027, Sep 2027.

So, if we're being pedantic, I'm not wrong until Oct 1st 2027 or later. But I'm also wrong if it happens this month ;)


> Anyway, i just wanna get on record that i predict an ai bubble pop event in the next 12 months.

Certainly possible, but I feel like you're very much going out on a limb predicting anything that soon. The market can stay irrational for a surprisingly long time if there's enough money floating around.

But I would be stunned if we don't have an AI bubble pop sometime in the next decade.


Utility for investors i guess sure, but for people that aren't investors, it mattered much less when RIM went under than that they still did.

> why should predictors be so coddled by their observers?

Because prediction is hard and no one is so good at it that they won't make errors like that.


They’ll never get better if not held accountable.

They haven't got better when held accountable, but you dissuaded further predictions.

Being early is the same as being wrong. Being lucky is the same as being right.

Nice saying. If in the prediction "given [intricate analysis] the whole [shebang] goes [bust] at [date]", only the [date] turns out to be wrong, I'd say it is still a valuable prediction however, though it was wrong.

Problem is even his “intricate analysis” was wrong: for example he was saying the models wouldn’t keep improving… in 2024.

Have they improved? Is there evidence of that? Got a task you were doing in 2024 and 2026 and the results of each?

There is plenty of evidence that they have improved in all benchmarks and also in my private experience. But have they improved in the things they still fail at? No, they still fail at them. You need only one example of failure to prove that it still fails. They still fail a lot on many real world tasks.

So, depending on what you ask, they may have not improved even a tiny bit.


I am a very light user, so this is my feeling from reading about other people's experience; I wouldn't say that they plateau'd but up to 4.5/4.8 the gains in the models felt exponential while since then they feel more linear and the big improvements are coming less from the models and more from everything around it (harnesses, agentic development, skills...).

So, while I don't feel like there has not been improvement, it really feels like there is a limit that will be reached sooner than later (and for sure before any AGI).


Oh my goodness. Is this really a good-faith question?

Since almost everything eventually goes bust, I wouldn't agree that this was particularly valuable, unless the analysis proved out in other ways. It's like the saying that some economists have predicted 10 of the last 4 recessions. The analysis that leads to that prediction may or may not be valuable, but is only valuable if it predicts something, because otherwise how can you know it has any accuracy at all?


Ok, I was thinking that probably they were saying that he was accidentally right, but missed the timing. It's not that. It isn't that he was accidentally right in a certain scenario also. It's that "I reinterpret the prediction to make it fit my own vision of the world"... that's not how prediction works. Hell that's not how anything would work.

Here's an actual prediction I made about a year ago: LLMs have to demonstrate that they make productivity gains that explain the costs or economics will make this problem solve itself, via higher energy cost and loss of business/productivity.

That prediction is unbound in the time horizon but it's bound by conditions that explain the triggers and how they will behave. Such prediction is useful. Hell, I could even make a prediction on why the timeline can't be bound while making a prediction on the timeline: I predict that in the next 3-5 years this will have to solve itself, because there's a limit on how much money irrational actors can pour onto something that have limited value. 7-10 years is way too much. I at least hope their coffers are that deep... if this drags on long enough, at some point people are going to want a change


This is an article that only cites Zitron’s opinions about subjective model quality. You’re tisk-tisking people to be objective about a bunch of words saying that the author’s opinions are better opinions than the guy he’s talking about OP

Some guy wrote that he’s grumpy because he couldn’t sleep and decided to dunk on an internet personality he doesn’t like, it’s not the ceremonial placement of the ur-kilogram


> This is an article that only cites Zitron’s opinions about subjective model quality. You’re tisk-tisking people to be objective about a bunch of words saying that the author’s opinions are better opinions than the guy he’s talking about OP

Oh it's not just the author's opinions. They're the opinions of a bunch of LLMs he checked, too. Much better.


What I observe is that what people don’t seem to grasp is that Zitron isn’t an AI sage. He’s simply someone who figured out that he can get a lot of attention by taking a contrarian stance when it comes to AI.

He can be 100% wrong about AI, but people will still read or listen to his next prediction. At this point, it’s mostly entertainment rather than a source of solid predictions.

He has one primary objective and that's to keep Ed Zitron in people’s minds by any means necessary.

It happens in sports, politics and, with Ed Zitron, AI.


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:

> August 2025 https://www.wheresyoured.at/how-to-argue-with-an-ai-booster/ : "These models have clearly hit a wall where training is hitting diminishing returns"

> Wrong

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.


>> August 2025 https://www.wheresyoured.at/how-to-argue-with-an-ai-booster/ : "These models have clearly hit a wall where training is hitting diminishing returns"

>> Wrong

>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.

>> Those differ from Dan’s essay, which engages with the literal text of Ed’s numerous predictions during 2024 and 2025 which are demonstrably invalidated by their measurable outcomes.

Dan Luu did not engage on anything more, than a disorganized wall of text, ranted like a teenager using toxic personal attacks, while obsessing over calendar errors and a placeholder in a spreadsheet. If this is what passes here for a smart engineer...Lets analyze his post in a more logical and analytical way:

- His entire argument is based on the naive logic that because LLM execution speeds or benchmarks marginally improved over the last 24 months, the entire trillion dollar investment cycle is justified. A short window of venture subsidized chip buying...tells you absolutely nothing about the multi decade debt structures, physical infrastructure depreciation, and power grid constraints that dictate whether a capital heavy business model survives.

- While he whines about Zitron numbers, fails to provide a single! macro level equation to address the real financial threat. NYU finance professor Aswath Damodaran for example, explicitly warned that the current AI build out is an asset heavy, debt funded run up backed by private capital markets. Unlike the dotcom boom which was equity funded and contained to tech shareholders today AI infrastructure burdens companies with a massive $80 billion in CapEx per gigawatt, meaning a monetization correction will trigger widespread systemic debt distress and loan defaults across the real economy.

"Aswath Damodaran: Big Tech Has No Idea How AI Pays Off" - https://news.ycombinator.com/item?id=49229981

- Luu and this HN crowd, today in a mob mood...completely ignore the highly unstable plumbing of the sector growth metrics. Patrick Boyle is a quantitative finance professor and former hedge fund manager, and has meticulously mapped out the mutual dependence the entire AI boom. Big Tech companies are pouring massive venture pools into AI startups, which are then contractually bound to hand that cash right back to the hyperscalers to buy cloud compute. Analysts have identified more than $800 billion in these arrangements:

"Why Wall Street is ignoring big tech's debt" - https://news.ycombinator.com/item?id=49230630

- The worst of Luu logical failure, is ignoring ( on purpose? ) were Zitron numbers come from! They come from some very disciplined institutions, which Luu completely ignores. Citigroup quantitative analysts project cumulative global AI CapEx hitting $9 Trillion through 2030, with maximum global AI revenues ( not profit...) covering less than 30% of that expenditure.

- To break even on the physical infrastructure currently under construction, the AI sector needs to generate over $2 Trillion in annual end user revenue by 2030. Total actual revenue generated across the ENTIRE global AI sector today sits at a fraction, around $150 billion.

- Anthropic in a hysterical push, to make it to public markets, before the bubble bursts, recently claimed their addressable market is 30 trillion... the whole of US economy. Are we getting a post from Luu on that? This of course this ignores that MIT Professor and Nobel Laureate, Daron Acemoglu, mathematically proved that while 20% of all labor tasks are exposed to AI, only about 5% can be automated profitably due to upfront enterprise systems integration and the high financial burden of constant human in the loop verification.

"A new look at the economics of AI" - https://mitsloan.mit.edu/ideas-made-to-matter/a-new-look-eco...

- Dismissing the AI bubble thesis, because you found a spreadsheet typo in a newsletter, and ignoring the other voices who are aligned with Zitron core premise, means you are also dismissing the research of a Nobel Laureate in economics, the Dean of Valuation, veteran hedge fund managers, Barclays, S&P Global, and Citigroup. Arguing that "the models are hitting benchmarks" while ignoring that the physical balance sheets and enterprise budgets cannot support a multi trillion dollar infrastructure build out, is exactly the type of Dunning Kruger this corner excels at....

Ed Zitron is correct, despite the clumsiness or unpleasantness of his message delivery, and this community reaction, will be an historical record of the AI bubble crowd madness.

It took years to take down Maddoff, and more to take down Bear Stearns. It will take maybe 5 - 10 years of "Ed Zitron is wrong posts here" until Anthropic and OpenAI have to be bailed out by the US government, but the day of reckoning will come. The end of this universe is all tax payers will own a piece of AI and will pay for it with increased interest rates for the next 25 years...


> the naive logic that because LLM execution speeds or benchmarks marginally improved over the last 24 months, (...)

"Marginally improved"? Are you really going to sit there tell me that an appropriate way to sum up the difference between the AI we had access to in Sep 2024 and the AI we have access to now, is "the benchmarks marginally improved"?


What do you mean by "take down Bear Stearns"?

Btw., projections are just that - projections and I am not sure Acemoglu proved things mathematical (as in a mathematical proof) but rather within the context of a model/assumptions.

That financial markets/innovation can outpace the actual innovation is also not some new insight, but that alone doesn't necessarily make for a useful prediction.


Bear Stearns was the first bank to collapse in the 2007-2008 subprime mortgage crisis, also known as the housing bubble.

That bubble was also manufactured by reckless financial engineers.

And those who warned early were ridiculed:

https://markets.businessinsider.com/news/stocks/who-is-nouri...

"When he spoke of an impending housing crash at the International Monetary Fund that year, the audience chuckled, the New York Times reported."

'"He sounded like a madman in 2006," IMF economist Prakash Loungani told the Times, after inviting Roubini to the IMF conference that year. "He was a prophet when he returned in 2007."'


I am fully aware of the GFC, but not sure what "take down" should mean there in relation to Bear.

Not everyone who spoke about house price risk was ridiculed, btw.


Catchy and ironic phrasing of "the financial fraudsters ruining Bear Stearns".

It seems unambiguous in this context.


Were there some criminal convictions?

The post is titled “How accurate have Ed Zitron's AI skeptic predictions been?” not “How accurate will Ed Zitron’s AI skeptic predictions be in the future?” or “Is the entire AI industry build out justified?”

If Patrick Boyle, Aswath Damodaran, and Daron Acemoglu have more accurate reporting and predictions about the upcoming decline of the AI industry, maybe those are voices who should be elevated over Zitron.


> It took years to take down Maddoff, and more to take down Bear Stearns. It will take maybe 5 - 10 years of "Ed Zitron is wrong posts here" until Anthropic and OpenAI have to be bailed out by the US government, but the day of reckoning will come.

Unfortunately, the market can stay irrational (far) longer than you can remain solvent.

In any case... I doubt Anthropic, OpenAI and xAI have any kind of moat that can justify a bailout. There is nothing truly unique either of these three possess, and certainly not against the free competition mostly from China or from Facebook that anyone can self-host.

Who will get the bailouts instead is the pension funds and other investment vehicles that have been force-fed crap AI stock like foie gras geese.


On interest rates I guess this is one of the concerns:

"AI “definitely is, in the short and medium run, a force that increases both natural rates and potentially price pressures,” Arellano said. But other shifting pieces of the U.S. economy appear to be significantly offsetting the effect of AI investment, for now. If accelerating AI investment were to outpace the residential slowdown—or if rates were to fall and residential investment rebound—spiking aggregate investment would mean strong demand and even more upward pressure on rates."

https://www.minneapolisfed.org/article/2026/how-is-ai-influe...


> Who will get the bailouts instead is the pension funds and other investment vehicles

It's the same thing in the end. Bailouts don't come from outer space, we all pay for delusions of few.


If Ed Zitron was merely saying that there is a AI bubble on the markets that will ultimately collapse even if we're not exactly sure when and how, then such prediction would be less interesting but also much harder to disprove.

But that's not what he's saying. He's making very specific claims that are indeed proven wrong. You can't honestly say he's correct, and the burst of an AI bubble will not be a reckoning.


You maybe right, but why don’t you put your prediction in Metacalculus or a prediction market. Thats what they are for.

On the flip side, I also see many people taking this thread as an opportunity to shit on Zitron as a person, and not "discussing his prediction". Incompetent/ blow hard/ dishonest are ones I remember off the top of my head.

AI by itself, is surprisingly polarized; Ed Zitron even more so.


Not sure why you see it as such;

Incompetent and dishonest are characteristics that follow from this professional work, adequately describing an individual who continues to make poor predictions, analyses and false statements refuted by past events.

Far from “shitting on” Zitron


I mean if his personal work is incompetent or dishonest...?

One of the easy ways to evaluate this is: how has he taken being incredibly wrong about his extremely confident predictions over and over and over?

Typically with people like this, they completely shrug off being super wrong. It's barely even a blip on their radar, and even bringing it up is a good way to get them to immediately attack you to deflect attention from how bad their predictions or assertions were.

If you're constantly making predictions on Topic X, and said predictions are consistently, wildly wrong, and you never actually grapple with that or acknowledge how wrong you were in the past, then that's, at the very least, intellectually dishonest.

But by all means, someone link us to his blog posts where he goes over his wrong predictions without excuses or deflections. I'd be happy to change my mind.


AI companies will never go to zero because AI is part of the Military Industrial Complex now. All the money is coming from the military and government for surveillance and power and war.

Speaking of wrong predictions... the entire US military budget would have to be spent on OpenAI/Anthropic to keep the bubble from bursting.

Every conspiracy theory has an escape hatch

Ad-hoc hypothesizing ("escape hatches") are dangerous but not quite invalid. Eg, they failed to find gravitational waves until they did, and you could have viewed building yet another more sensitive detector as a similar exercise in refining a hypothesis that you keep receiving contrary evidence for. Sometimes you really did just underestimate how difficult your hypothesis was to demonstrate. Maybe Meta will be destroyed in 2027, or whatever.

The problem with conspiracy theories is more that they have a ratchet-like quality where counter evidence reaffirms the theory in your view and you can only ever get more confident. We should have been increasingly skeptical of gravitational waves to some degree as we failed to demonstrate them, even though we didn't abandon the hypothesis and it ultimately prevailed. But if you adopt a wrong idea, and people try to demonstrate that to you, and you take that effort they're putting forward as a sign that you are correct and they must be hiding something from you, it will be very difficult for you to realize your mistake.

So, as long as you are less certain than you were before, I don't think rolling your prediction over into the future is necessarily conspiratorial or a mistake.


It’s the moon-leopards in league with the bee-people. Only an IDIOT wouldn’t know this, it’s OBVIOUS.

I actually think you’ve got it backwards

I think one of Zitron’s problems is that his moral righteousness has blinded him to how embarrassingly incorrect he is about the AI space.

It’s similarly hubristic with the benefit of shielding from his adversarial framing.


I see this so much. "This shouldn't work, therefore it won't work." It's like the world is imagined to run on a moral causal framework rather than uncaring quantum mechanics.

I have watched many of his screeds; I think he undercuts his thesis with excess bile.

His core observation is that the unit economics of openAI and anthropic don't actually yield enough profit to pay off the huge debts that these two companies have incurred, and that as a result all the debt they've taken on will have to be written off which will trigger "the hyperscalers" to themselves suffer huge losses (likely wounding google and microsoft and destroying oracle).

He is rather more negative about the utility of LLMs than lots of other people (myself included); but his overall view of

  1. they're not completely trustworthy
  2. they're really expensive to train
  3. it isn't obvious that anyone's willing to pay the full freight for the resulting product
  4. lots of large orgs "that should know better" have gone far down the LLM "AI" road because they're looking for "the next big thing" when they should be pivoting to "mature, stable" companies instead of "hypergrowth" companies.
  5. there's lots of debt and obligations and no obvious way for all of it to be paid off from revenues from openai / anthropic.
seems pretty reasonable. There seem to have been lots of bets placed on the hope that this stuff will continue scaling as it has in the past, if it is given more compute and data, and that bet is one that has yet to have demonstrated itself as correct.

Elsewhere, people have pointed out that many of the "FAANG" companies have shed lots of people and driven lots of profits, largely on the back of internal use of LLM tools. That doesn't necessarily contradict the skepticism that anthropic and openAI will succeed, and given all their obligations, if they fail it'll be a big mess.

But again, big Z doesn't do himself any favors when he rants about "failsons" or whatever.


Seems to me it's a reasonable position that runs face first into the irrational market.

Like they're digging a gigantic hole and Ed's up the top saying if you keep digging the hole will collapse (+ a whole lot of unnecessary swearing), and then a bunch of people jump into the hole to brace it and say "nuh uh, see we can keep digging" but really it's just postponing the inevitable and increasing the number of people who will be destroyed when it all crashes down


But I think anyone who evaluates their claims understands that they’re talking their own books.

Zitron’s problems are more subtle. He benefits financially from his own claims, while also touting his impartiality and capacity for objective thought.


A person savvy enough to understand the indirect financial benefit to Dario promising that Claude is so dangerously smart it must be regulated understands Zitron’s schtick

But we’re talking about Gemini, which is one of Dan’s point: the surface area is so large that any prediction is almost laughable. Zitron’s prediction here is laughable in fact because he claims both (1) Pichai is jamming AI everywhere and (2) the 500M number is impossible. Those statements are immediately contradictory to anyone who understands the scale of Google’s products.

A prediction definitionally includes an outcome, and it _can_ include a timeline. (if you’re serious about your forecasting, then you generally include a timeline; forecasting is non-actionable if it does not include a timeline).

You might choose a different timeline for Ed’s outcomes. Okay: those are now your predictions, not his.

Zitron has done us the favor of including timelines with his predictions, so that Dan can invalidate almost all of them.


The problem isn’t the SDK but the API usage.

Users will say this-or-that about choice etc etc. It’s about subsidized tokens. Otherwise th users (and OpenCode) would have stopped pushing the workarounds months ago.


> The problem isn’t the SDK but the API usage.

Yes, i completely agree - which was my question. To my understanding OC was using CC API directly instead of using the SDK. My question is, why? Does the SDK prevent some functionality they needed?


I’m sorry but you’re demonstrably incorrect.

Listen, I want more open weight models in the world. They create entrepreneurial opportunities and support use cases which the foundation labs don’t want to support.

But open weight models are consistently three to six months behind on performance compared to closed models, as confirmed by both benchmarks and personal use. They’re closer on coding and much further away on non-coding tasks.

There are theories as to why these models lag, which I won’t get into. But anyone claiming open-weight models are close to closed-weight models is ignoring significant evidence to the contrary.


> three to six months behind on performance

Yeah, like I said - it's just a post-training difference. That's not a material difference, that's a difference of chrome and polish.


> I’m sorry but you’re demonstrably incorrect.

Please so demonstrate?


The onus isn’t on me. It’s on anyone contradicting findings by most benchmarks, because most of them show a clear advantage for Opus and GPT over OSS models.


So Big Claim No Demonstration? :-)


I mean just use them and compare, the gap is obvious.


I did, and I fixed Qwen's issues with trivial sampling and loop detection hacks.

If I can do this, then a company that wants to sell local models seriously could do it too.


> I did, and I fixed Qwen's issues with trivial sampling and loop detection hacks.

Wow, that's amazing! Care to share the changes? Would love to try them out.


It's not amazing at all.

What's amazing is that LLM technologies are so immature that even basic engineering diligence isn't being done. (Like detecting token loops, for example.)


This is both true and immaterial to Microsoft’s net profits.


Yep, well said and great, sharp explanation.

I think we can attribute a bunch of consternation here to drift between assumed and actual licensing terms.

The actual licensing terms for Claude Code expressly prohibit use of the product outside of the Claude Code harness. If you want Opus outside of CC, the API is available for your use anytime.

Some percentage of the community seems to assume their Claude Code subscription licenses allow free usage of CC across any product surface - including competing products like OpenCode. While this is a great way to save on API costs, the assumption is incorrect. In fact, it is *so* incorrect that Anthropic has encoded their licensing terms into their Terms of Service, and a result can take legal action against any violating parties.

We can have separate discussions about Anthropic’s use of the Common Crawl in pre-training, or whether foundation labs adhere to robots.txt conventions. But those don’t directly impact Anthropic’s right to bring litigation.

——

Outside of that I think angry users have their own stated preferences v revealed preferences here. They claim they want Opus on their terms, and Anthropic’s actions infringe on their user rights.

Angry folks: Opus is right there! You just need an API key! The reality is you want Opus in your devtools of choice at discounted rates. You could at least be honest about your consternation


> including competing products like OpenCode

I think that’s a bit more nuanced. The actual „product” is not the harness, which is free anyway, but the Claude subscription. In any scenario, that’s what the customer continues to pay for. I understand why Anthropic is doing that, but I feel no need to defend it. Just like I understand why Apple limits your app choices to AppStore, but I’m not going to go out of my way to defend their decision.


It's way more nuanced, because the subscription is older then Claude Code - and they only started to have a problem with third parties using it after Claude Code. (And not with the release, just some time after the release)


That makes perfect sense because that's when it became orders of magnitude more expensive to offer the service.


To me, that argument would only make sense if the subscription wasn't metered... But it is.


Sheriff spits to the ground. One harness. One horse. How we do it' fer now on.


>We can have separate discussions about Anthropic’s use of the Common Crawl in pre-training, or whether foundation labs adhere to robots.txt conventions. But those don’t directly impact Anthropic’s right to bring litigation.

Some of us don't care for Anthropic's "right to bring litigation" anymore than we care about some scumbag patent troll company doing things "within their legal rights".

We care for the morality of its conduct, the openess of its products, and the environment it creates.


I think this is disingenuous, people want to be able to use a tool that they pay for to do useful work on their own terms because they payed for it and don’t see the differential pricing model offered by Anthropic as legitimate.


Why would it not be legitimate?


[flagged]


It's very amusing to hear this particular argument being made to defend AI companies.

When the people want that, it's inconsequential.

When the corporations did that, it was their God-given right.


I don’t agree, what people want is very consequential, because those people are paying customers of a service, if they aren’t happy with it they have every right to complain.

People should be vocal about what they do and do not think is reasonable behavior by corporations and then act based on those opinions with their wallets. Lord knows we have precious few other ways of influencing corporate behavior.


No, it's important. People are allowed to discourage each other from buying a product that they consider subpar.


>What the people want is inconsequential here. The people also want to abolish copyright and freely share and download media too.

I already approved of the complaints against Anthropic here, you don't have to sell it this hard to me.

(Not to mention the blatant hypocrisy that their whole business is based on open copyright abuse - all that copyrighted training material, illegally obtained books and movies, etc).


They can't take any legal action outside of the US. In most other jurisdiction such Bullshit in the ToS would be void anyway


Yes, it is true that companies often litigate against customers who violate their Terms of Service. The TOS is put into place to protect the company’s interests from user abuse.

Paying customers of Claude Code don’t receive a free-use license for any desired application. They’re paying to use Claude Code. Anthropic can take steps to litigate usage outside of those terms, even if customers find that fact really annoying.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: