In a previous generation, the enabler of all our computer tech innovation was the incredible pace of compute growth due to Moore's Law, which was also "top-down" from very well-funded companies since designing and building cutting edge chips was (and still is) very, very expensive. The hype was insane, and decisions about what chip features to build were made largely on the basis of existing vendor relationships. Those companies benefited, but so did the rest of us. History rhymes.

This is very pessimistic take. Where else do you think the innovation would come from? Take cloud for example - where did the innovation come from? It was from the top. I have no idea how you came to the conclusion that this implies monetization is going to be awful.

How do you know models are expensive to run? They have gone down in price repeatedly in the last 2 years. Why do you assume it has to run in the cloud when open source models can perform well?

> The hype is insane, and so usage is being pushed by C-suite folks who have no idea whether it's actually benefiting someone "on the ground" and decisions around which AI to use are often being made on the basis of existing vendor relationships

There are hundreds of millions of chatgpt users weekly. They didn't need a C suite to push the usage.

> All the innovation is coming "top-down" from very well funded companies - many of them tech incumbents

What I always thought was exceptional is that it turns out it wasn't the incumbents who have the obvious advantage.

Take away the fact that everyone involved is already at the top 0.00001% echelon of the space (Sam Altman and everyone involved with the creation of OpenAI), but if you had asked me 10 years ago who will have the leg up creating advanced AI I would have said all the big companies hoarding data.

Turns out just having that data wasn't a starting requirement for the generation of models we have now.

A lot of the top players in the space are not the giant companies with unlimited resources.

Of course this isn't the web or web 2.0 era where to start something huge the starting capital was comparatively tiny, but it's interesting to see that the space allows for brand new companies to come out and be competitive against Google and Meta.

> All the innovation is coming "top-down" from very well funded companies - many of them tech incumbents

The model leaders here are OpenAI and Anthropic, two new companies. In the programming space, the next leaders are Qwen and DeepSeek. The one incumbent is Google who trails all four for my workloads.

In the DevTools space, a new startup, Cursor, has muscled in on Microsoft's space.

This is all capital heavy, yes, because models are capital heavy to build. But the Innovator's Dilemma persists. Startups lead the way.

This is only if you ignore the growing open source models. I'm running Qwen3-30B at home and it works great for most of the use cases I have. I think we're going to find that the optimizations coming from companies out of China are going to continue making local LLMs easier for folks to run.

They've gotta hope they get to cheap AGI, though.

Any stall in progress either on chips or smartness/FLOP means there's a lot of surplus previous generation gear that can hang and commoditize it all out to open models.

Just like how the "dot com bust" brought about an ISP renaissance on all the surplus, cheap-but-slightly-off-leading-edge gear.

IMO that's the opportunity for a vibrant AI ecosystem.

Of course, if they get to cheap AGI, we're cooked: both from vendors having so much control and the destabilization that will come to labor markets, etc.

> The whole LLM era is horrible. All the innovation is coming "top-down" from very well funded companies

Wouldn't it be the same for the hardware companies? Not everyone could build CPUs as Intel/Motorola/IBM did, not everyone could build mainframes like IBM did, and not everyone could build smart phones like Apple or Samsung did. I'd assume it boils down the value of the LLMs instead of who has the moat. Of course, personally I really wish everyone can participate in the innovation like the internet era, like training and serving large models on a laptop. I guess that day will come, like PC over mainframes, but just not now.

Nevertheless, prices for LLM at any given level of performance have gone down precipitously over the past few years. Regardless of how bad it seems the decisions being made are, the decision making process both is making an extreme amount of money for those in the AI companies, and providing extremely cheap and high quality intelligence for those using their offerings.

What's the counterfactual? Where would the world be today? Certainly the present is not an optimal allocation of resources, uncertainty and hysteris make it impossible. But where do you think we'd be instead? Are you assuming all of those dollars would be going to research otherwise? They wouldn't; if not for hype "ai" LLMs, research funding would be at 2017+/- 25% levels. Also think of how many researchers are funded and PhDs are trained because of this awful LLM era. Certainly their skills transfer. (Not that brute forcing with shit tons of compute is standard "research funding").

And for the record I really wish more money was being thrown outside of LLM.

Come to the counterrevolution; we have cookies : )

How can you dismiss the value of the tech so blatantly? Have you used Opus for general questions and coding?

> no idea whether it's actually benefiting someone "on the ground"

I really don't get it. Before, we were farmers plowing by hand, and now we are using tractors.

I do totally agree with your sentiment that it's still a horrible development though! Before Claude Code, I ran everything offline, all FOSS, owned all my machines, servers etc. Now I'm a subscription user. Zero control, zero privacy. That is the downside of it all.

Actually, it's just like the mechanisation of farming! Collectivization in some countries was a nightmare for small land owners who cultivated the land (probably with animals). They went from that to a more efficient, government controlled collective farm, where they were just a farm worker, with the land reclaimed through land reform. That was an upgrade for the efficiency of farming, needing fewer humans for it. But a huge downgrade for the individual small-scale land owners.

We do seem to be hitting the top of the curve of diminishing returns. Forget AGI - they need a performance breakthrough in order to stop shoveling money into this cash furnace.

> Anecdotally moving from model to model I'm not seeing huge changes in many use cases.

Probably because you're doing things that are hitting mostly the "well-established" behaviors of these models — the ones that have been stable for at least a full model-generation now, that the AI bigcorps are currently happy keeping stable (since they achieved 100% on some previous benchmark for those behaviors, and changing them now would be a regression per those benchmarks.)

Meanwhile, the AI bigcorps are focusing on extending these models' capabilities at the edge/frontier, to get them to do things they can't currently do. (Mostly this is inside-baseball stuff to "make the model better as a tool for enhancing the model": ever-better domain-specific analysis capabilities, to "logic out" whether training data belongs in the training corpus for some fine-tune; and domain-specific synthesis capabilities, to procedurally generate unbounded amounts of useful fine-tuning corpus for specific tasks, ala AlphaZero playing unbounded amounts of Go games against itself to learn on.)

This means that the models are getting constantly bigger. And this is unsustainable. So, obviously, the goal here is to go through this as a transitionary bootstrap phase, to reach some goal that allows the size of the models to be reduced.

IMHO these models will mostly stay stable-looking for their established consumer-facing use-cases, while slowly expanding TAM "in the background" into new domain-specific use-cases (e.g. constructing novel math proofs in iterative cooperation with a prover) — until eventually, the sum of those added domain-specific capabilities will turn out to have all along doubled as a toolkit these companies were slowly building to "use models to analyze models" — allowing the AI bigcorps to apply models to the task of optimizing models down to something that run with positive-margin OpEx on whatever hardware that would be available at that time 5+ years down the line.

And then we'll see them turn to genuinely improving the model behavior for consumer use-cases again; because only at that point will they genuinely be making money by scaling consumer usage — rather than treating consumer usage purely as a marketing loss-leader paid for by the professional usage + ongoing capital investment that that consumer usage inspires.

The raw model scale is not increasing by much lately. AI companies are constrained by what fits in this generation of hardware, and waiting for the next generation to become available. Models that are much larger than the current frontier are still too expensive to train, and far too expensive to serve them en masse.

In the meanwhile, "better data", "better training methods" and "more training compute" are the main ways you can squeeze out more performance juice without increasing the scale. And there are obvious gains to be had there.

Its not clear to me that it needs to. If at the margins it can still provide an advantage in the market or national defense, then the spice must flow

I hope you're right.

> I can just pick an older model and often I can't tell the difference...

Or, as in the case of a leading North American LLM provider, I would love to be able to choose an older model but it chooses it for me instead.

> Anecdotally moving from model to model I'm not seeing huge changes in many use cases. I can just pick an older model and often I can't tell the difference...

Model specialization. For example a model with legal knowledge based on [private] sources not used until now.

Locally run video models that are just as good as today’s closed models are going to be the watershed moment

The companies doing foundational video models have stakeholders that don’t want to be associated with what people really want to generate

But they are pushing the space forward and the uncensored and unrestricted video model is coming

The scaling laws already predict diminishing in returns

It is being played like a winner-takes-it-all right now (it may or may not be such a market). So it is a game of being the one that is left standing, once the others fall off. In this kind of game, speeding more is done as a strategy to increase the chances of other competitors running out of cash or otherwise hitting a wall. Sustainability is the opposite of the goal being pursued... Whether one reaches "AGI" is not considered important either, as long as one can starve out most competitors.

And for the newcomers, the scale needs to be bigger than what the incumbents (Google and Microsoft) have as discretionary spending - which is at least a few billion per year. Because at that rate, those companies can sustain it forever and would be default winners. So I think yearly expenditure is going to be 20B year++

They are only getting deprecated this fast because the cost of training is in some sense sustainable. Once it is not, then they will no longer be deprecated so fast.

Imagine solving energy as a side effect of this compute race. There's finally a reason for big money to be invested into energy infrastructure and innovation to solve a problem that can't be solved with traditional approaches.

Continuing to carve out economies of scale in battery + photovoltaic for another ten doublings has plenty of positive externalities.

The problem is that in the meantime, they're going to nuke our existing powergrid, created in the 1920's to 1950's to serve our population as it was in the 1970's, and for the most part not expanded since. All of the delta is in price-mediated "demand reduction" of existing users.

> So long as there is competition it’ll be available at marginal cost.

Most things are not perfect competition, so you get MR=MC not P=MC.

We're talking about massive capital costs. Another name for massive capital costs are "barriers to entry".

As a fun comparison, because I saw the population is more or less the same.

The GDP of the Netherlands is about $1.2 trillion with a population of 18 million people.

I understand that that’s not quite what’s meant with ‘small country’ but in both population and size it doesn’t necessarily seem accurate.

Country scale is weird because it has such a large range.

California (where Anthropic is headquartered) has over twice as many people as all of Somalia.

The state of California has a GDP of $4.1 Trillion. $13 billion is a rounding error at that scale.

Even the San Francisco Bay Area alone has around half as many people as Somalia.

Not only are the actual models rapidly devaluing, the hardware is too. Spend $1B on GPUs and next year there's a much better model out that's massively devalued your existing datacenter. These companies are building mountains of quicksand that they have to constantly pour more cash on else they be reduced to having no advantage rapidly.

Assuming the best case: we're going to need to turn this productivity into houses or lifestyle improvement, soon... or I'm just going out with Sasquatch

I feel like it's pretty settled that they are a little bit useful, as a faster search engine, or being able to automatically sort my emails. But the value is nowhere near justifying the investment.

Off software codegen alone it is beyond useful

Whatever happens if/when a flagship model leaks, the legal fallout would be very funny to watch. Lawyers desperately trying to thread the needle such that training on libgen is fair use, but training on leaked weights warrants the death penalty.

In this imaginary reality where LLMs just keep getting better and better, all that a leak means is that you will eat-up your capital until you release your next generation. And you will want to release it very quickly either way, and should have a problem for a few months at most.

And if LLMs don't keep getting qualitatively more capable every few months, that means that all this investment won't pay off and people will soon just use some open weights for everything.

You can't run Claude on your PC; you need servers. Companies that have that kind of hardware are not going to touch a pirated model. And the next model will be out in a few months anyway.

> Once the model is out, no one would pay for it, right?

Well who does the inference at the scale we're talking about here? That's (a key part of) the moat.

There’s the opportunity cost here of those resources (and not talking only about the money) not being spent on power generating that actually benefits the individual consumer.

gpt-oss-120b has cost OpenAI virtually all of my revenue, because I can pay Cerebras and Groq a fraction of what I was paying for o4-mini and get dramatically faster inference, for a model that passes my eval suite. This is to say, I think high-quality "open" models that are _good enough_ are a much bigger threat. Even more so since OpenRouter has essentially commoditized generation.

Each new commercial model needs to not just be better than the previous version, it needs to be significantly better than the SOTA open models for the bread-and-butter generation that I'm willing to pay the developer a premium to use their resources for generation.

Can you elaborate? The technology to build a human brain would cost billions in today’s dollars. Are you thinking moreso about energy efficiency?

Sentiment I have heard is manufactories do not want to increase die size because defects per die increases at the same time.

A couple of counter-arguments:

Labs can just step up the way they track signs of prompts meant for model distillation. Distillation requires a fairly large number of prompt/response tuples, and I am quite certain that all of the main labs have the capability to detect and impede that type of use if they put their backs into it.

Distillation doesn't make the compute moat irrelevant. You can get good results from distillation, but (intuitively, maybe I'm wrong here because I haven't done evals on this myself) you can't beat the upstream model in performance. That means that most (albeit obviously not all) customers will simply gravitate toward the better performing model if the cost/token ratio is aligned for them.

Are there always going to be smaller labs? Sure, yes. Is the compute mote real, and does it matter? Absolutely.

it's difficult for me to imagine this level of compute existing and sitting there idle somewhere; it just doesn't make sense.

So we can at least assume that whoever is deciding to move the capacity does so at some business risk elsewhere.

Not to mention water for cooling. Large data centers can use 1 million+ gallons per day.

A lot, but maybe a lot less than you expect.

You'd be competing with ASIC miners, which are 100x more cost effective per MH/s. You don't need 100,000GB of VRAM when mining GPU, therefore its waste.

SV playbook has been to make sustainable businesses. Uber makes profits, so do Google, Amazon and other big tech.

> LLMs are amazing but I'm not sure they justify the astronomical training capex, other than as a stepping stone.

They can just... stop training today and quickly recuperate the costs because inference is mostly profitable.

Why would a global taxi/delivery broker not be worth billions? Their most recent 10-Q says they broker 36 million rides or deliveries per day. Even profiting $1 on each of those would result in a company worth billions.

There are other coins that are less unprofitable to mine (see https://whattomine.com/gpus ) but it's probably still not worth it.

Before your edit I was going to answer, sadly no, they can't even be repurposed for Bitcoin mining.

> Barely 50 years ago computers used to cost a million dollars and were less powerful than your phone's SIM card.

Fifty years ago, we were starting to see the very beginning of workstations (not quite the personal computer of modern days), something like this: https://en.wikipedia.org/wiki/Xerox_Alto, which cost ~$100k in inflation-adjusted money.

Yeah, no hate for kung fu here, but maybe learning to better communicate together, act in ways that allows everyone to thrive in harmony and spread peace among all humanity might be a better thing to start incorporating, might not it?

Deepseek story was way overblown. Read the gpt-oss paper, the actual training run is not the only expense. You have multiple experimental training runs as well as failed training runs. + they were behind SOTA even then

your crystal ball needs calibration. this round alone was 14pct (183/13)... so the dillution was likely over 20pct.

It helps me to know that there are other people noticing this. Like Fox News, a lot of comments here probably make us dumber.

CC being better than Cursor didn't make sense to me until I realized Anthropic trains[0] it's models to use it's own built-in tools[1].

0 - https://x.com/thisritchie/status/1944038132665454841

1- https://docs.anthropic.com/en/docs/agents-and-tools/tool-use...

My gut feeling is that Claude Code being so popular is: - 60% just having a much better UX and having any amount of "taste", compared to Cursor - 39,9% being able to subsidize the raw token costs compared to what's being billed to Cursor - 0,1% some magical advantage by also training the model

Claude Code is just much _pleasant_ to use than most other tools, and I think people are overly discounting that aspect of it.

I'd rather use CC with slightly dumber model, than Cursor with a slightly better one; and I suspect I'm far from being the only one.

Their employers are paying that money. The jury is still out on how wisely that money is being spent.

Yeah, a collapse should only mean that training larger models become non viable right? Selling inference alone should still deliver profit.

Maybe in enterprise.

But in the consumer market segment, for most cases, its all about who is cheapest (free preferably) - aside from the few loonies who care about personality.

The true lasting economic benefits within enterprise are yet to play out. The trade off between faster code production vs poorer maintained code is yet to play out.

"This is a technology that could replace a sizable fraction of humamkind as a labor input."

And if it does? What happens when a sizable fraction of humamkind is hungry and can't find work? It usually doesn't turn out so well for the rich.

Hope it stays long enough to build up serious electricity generation, storage and distribution. Cause that has a lot of productive uses, and has historically been underdeveloped (in favor of fossile fuels). Though there will likely be a squeeze before we get there...

It'd be an interesting time for China to invade Taiwan.

Why is this downvoted when it's spot on.. if reality < expectations so much money is sitting on extremely quickly depreciating assets. It will be bad. Risk is to the downside.

I think you mean hypocritical.

To answer: no, and even if it was a “yes” it wouldn’t affect the argument I was making. I’ll explain.

I was wondering how long it would take for this kind of meta-critique would pop up. Meta critiques are interesting: some people use them as zingers, hoping to dismantle someone else’s entire position. But they almost never accomplish that because they are at a different level of argument: they aren’t engaging with the argument itself.

Meta-critiques are more like an argument against the person crafting the argument. In this sense, they function not unlike ad hominem attacks while sneakily remaining fair game.

Lastly, even if I was a hypocrite, it wouldn’t necessarily mean that I was wrong — it would simply make me inconsistent in the application of a principle.

I don’t interpret the above as mean-spirited comment, but it does miss the point of the example I gave; namely, people second-guessing a market (or information heavily influenced by markets, like a new funding round) tend to lose. (Of course there are examples in the other direction, but they are less common and do not deserve equal emphasis.)

In general, a market synthesizes more information than any one individual, and when they operate well it is unlikely for an individual is going to beat them.

This is a well known general pattern, so if someone wants to argue in the other direction, they need to be ready to offer very strong evidence and reasoning why the market is wrong — and even when they do, they’re still probably going to be wrong.

> Google also bought Motorola for 12 billion and Microsoft bought Nokia for 7 billion. Those weren't success cases.

A lot of that was patent acquisition rather than trying to run those businesses so it's hard to say a success or not.

I see what you are getting at, but it is important to understand the context for my example and the argument I’m making.

I’ve explained various points at length in other comments: (i) why I selected this example (simply to show that folk wisdom or common sense is less reliable than market-driven valuations) (ii) how a funding round is influenced by markets even though it isn’t directly driven by a classic full market mechanism.

Something I haven’t said yet would be a question: how can an outsider rigorously assess the error in a funding round or acquisition? To phrase the question a different way: what price or valuation would an oracle assign based on known information?

One might call this ex-ante rationality. Framing it this way helps remove hindsight bias; for example, a subsequent failure doesn’t necessarily mean it was mispriced (sp?) at the time.

You seem to think that one selected analogy proves that there’s no other explanation by which something is sensible?

Example: I could also make up dozens of reasons why something permitted by the laws of physics seems ridiculous.

You are just making up nonsense.

20x earnings is not that insane for a fast growing startup. Now Tesla and Palantir at 100x earnings is insane.

Do you think some arbitrary multiple defines a bubble?

Why WeWork and not Alibaba?

Answer: It's easy to pick and choose to prove one's point.

Softbank has been doing well lately by the way:

https://www.ebc.com/forex/softbank-stock-price-jumps-13-afte...

Most investors I've heard talk about the AI bubble have mentioned exactly that they know it is a bubble. They are just playing the game, because there is money to be made before that bubble bursts. And additionally, there is real value in these companies.

I would assume the majority of investors in AI are playing a game of estimating how much more these AI valuations can run before crashing, and whether that crash will matter in the long-run if the growth of these companies lives up to their estimates.

That sounds very cynical and knowing which is obviously great, but not super interesting. Do you think the investors are insane or stupid? Do you think this is a bubble and that it's about to burst? I'm interested to know why.

I see two of nineteen investors were also invested in FTX (Insight and Ontario teachers). With hindsight that's a bad investment although they probably recovered their money here so probably not their worst. Does this actually tell you they are stupid or insane?

I think that's one possible interpretation but another is that these funds choose to allocate a controlled portion of their capital toward high risk investments with the expectation that many will fail but some will pay off. It's far from clear that they are crazy or stupid.

... or really any SoftBank Vision Fund backed startup

In Jan, when deepseek launched, Dario Amodei had to disclose they spent about $10M to train the last generation of models (his arguments was deepseek was on the curve, not breaking it).

They earned $250M in May based on ARR, and about $400M in july. Model training is going to be amortized over multiple years anyway. I am not privy to how much they spent, not going to comment on that. GM was public news, and hence I got that.

Re Zitron's analysis, I don't find them to be reliable or compelling.

> what are people who make these investments even betting on?

They they achieve AGI or a close approximation, and end up wealthier than god.

That's basically the bet here. Invest in OpenAI and Anthropic, and hope one of them reached near AGI.

But if you're an investor who doesn't care about the long-term value of the firm, all you care about is maximizing your return on future sales of the shares of stock.

Doing proper intrinsic valuation with technology firms is nigh-on impossible to do.

Yeah. Electricity production that we also need for electrifying all of our transports and industry because of climate change. Good luck getting that cheaper.

That said innovation on the model side is more likely to come from a 10B-funded startup that still has some money to spare on the brightest researchers on top of giving them all the data and compute they want to play with.

But if we could drink a bottle of oil and become 10x smarter for 1 hour, it would be really cool. There just wasn't any use for that in the savannah, or indeed many bottles of oil.

Yeah these things take time to play out. So I always just say, the large populous of people will finally realise fantasy and reality have to converge at some point.

Government workers are funded by public money, so public pension monies are funded by public money ultimately.

Yes. So they indirectly owned some Anthropic through the FTX bankruptcy. I kinda wonder whether they somehow opted to keep their Anthropic stake when the FTX estate sold. Or maybe they bought it at some other time.

https://www.sec.gov/enforcement-litigation/litigation-releas...

A fascinating investor. I just finished re-reading Microserfs. The buzzwords may have changed between 1993 and 2025, but the human behaviors certainly have not.

For those curious, I found this to be a very entertaining retelling of events from Nortel's persepective: https://www.youtube.com/watch?v=I6xwMIUPHss

Proudly proclaiming on the Conversations With Tyler podcast that given a double or nothing bet with a 51% chance of success he'd keep playing forever.

Like Martin Shkreli, who made his investors whole with his gambling, but still went to jail.

Yes but investors being whole and profitable would almost certainly have not resulted in jail time. He probably would have even had enough unquestionably personal returns to pay back any misappropriated funds in a negotiated settlement, if they even come to light at all.

Let's not pretend there aren't multitudes out there doing similar things who never get caught. SBF was just more egregious and untimely w/ his actions.

Rogue