AI eats the world (Spring 26) [pdf](static1.squarespace.com) |
AI eats the world (Spring 26) [pdf](static1.squarespace.com) |
Nov 2024: Don’t dismiss this; it may be the next platform shift. But the actual questions are still unsettled: scaling, usefulness, deployment, and business model.
May 2025: The model layer is already showing signs of commoditization, so the important question shifts toward deployment: products, use cases, UX, errors, and enterprise adoption.
Nov 2025: The capital cycle has become the story: everyone is spending because missing the platform shift is worse than overbuilding, but there is still no clarity on product shape, moats, or value capture. That creates bubble-like dynamics.
May 2026: Provisional thesis: models look likely to become infrastructure, while value probably moves up-stack into apps, workflows, product, proprietary data/context, GTM, and new questions made possible by cheap automation. But he is still explicitly calling this provisional.
Some things that stand out:
* He’s really good with his historical analogies, especially looking at previous transformations like the early Internet and mobile; no surprise given that he has a history degree.
* he emphasizes over and over how we have still have no idea how all of this is going to work when the dust settles. I think that’s kind of a historian’s move as well. When you look at what people were saying during the early days of the web, for example, almost all of their predictions weren’t just wrong… in hindsight, given how the future played out, they were asking the wrong questions. The implication is that we are probably asking the wrong questions about AI too.
* Nonetheless his thesis about the commoditization of models is actually a fairly strong concrete prediction. i’m not sure if I agree with it entirely, but I do keep it in mind every time I look at the valuation of leading AI labs.
* he continually makes the point that a chat bot is barely a product and that AI labs have so far had very little success in delivering products above that layer… with the exception of coding agents, of course.
But I absolutely agree that in hindsight we are often asking the wrong questions about each new technology.
I keep seeing on HN that AI is a hype, and many here are anti AI (which I get, as a programmer AI made my job less interesting, and I'm even worried about losing it), but where has AI underdelivered?
Do you have an example of this? My (poor) memory remembers "it's going to change how people buy things", was the big deal at the time, and it seems like it was a great prediction.
Like, yes, the telecom bubble was a clear case of overbuilding and the AI data center "bubble" looks a lot like that... but this overlooks that the fiber capacity being laid back then far outstripped the demand, whereas all the compute providers today have been desperately crunched for capacity, despite investing almost a trillion in CapEx -- to the tune of almost a trillion dollars more of backlog -- for multiple quarters now.
Or yes, historically new technology has always created new jobs... but all those new jobs required a higher skill level along dimensions that current AI models are already good at, meaning we've never had a technological revolution quite like this.
Or yes, prior technological revolutions consigned incumbents to irrelevancy, primarily due to shifts in technical platforms... but then today's business leaders are 1) very well educated about what happened to their predecessors, 2) very paranoid about the same thing happening to them, and hence 3) are actively making moves to capitalize on the next platform shift.
I also think his dismissal of chatbots is a bit premature. It is precisely because chatbots operate via an extremely simple, flexible and natural modality, i.e. a conversation -- entirely unconstrained by the form factor necessitated by any app -- that their infinite use-cases have become unleashed.
My take is that the AI labs are actively exploiting this extreme flexibility to surface valuable use-cases -- one of the hardest parts of innovation -- at which point they can simply slap an agent on top of them. Which is, yet again, simply a chatbot, except one that can actually do useful things for you and hence can be charged for a lot more money.
Same thing happened on other places the open source offering became popular.
OpenAI / Google / Anthropic / XAI also have a ton of compute. That is the real moat.
The main thing that stands out to me on these graphs is just . how . early we still are - looking at industries like legal which in my mind are certainly going to be massively disrupted, and seeing the very low usage rates vs. tech (which still shows less than a quarter of tech people using AI daily) — we are in for a lot more change than we’ve seen so far.
On the technical side, one of the additional things I've had on my mind is the potential that these mega models are in fact hiding a ton of inefficiency.
The approach of simply shoving higher dimensionality and more parameters into largely tweaks to the current models has delivered results, but it feels like "mainframe" era of computing to me.
Throwing reams of annotated human content and forcing the machine to globally draw associations from it feels clumsy. Just as people are able to learn structured knowledge via rule-systems that are successively elaborated with extensions and situational contradictions, I feel like there's probably a much more compact representational model that can be reached by adapting the current technical foundations (transformers, attention, etc.) to work well with generated examples from rule-systems, that then gets used as a base layer to augment the "high level" models that process unstructured data.
The risk for the behemoth datacenter might be similar to the risk in the early computing era of building compute centers right before the PC revolution took off.
If it turns out that there exists some more compact and efficient representation for this intelligence (which IMHO is likely given that we are still in the first generation of this technology), the datacenters may end up decaying mausoleums of old tech that has no relevance to a distributed intelligence future.
That's the big technical unknown unknown for me. How much efficiency juice is there left to squeeze, and what does that mean for a distributed landscape vs a centralized datacenter based landscape.
Pretty much all of the stuff that was suggested back then or earlier: Shopping, advertising, video conferencing, collaboration, software distribution, media consumption, banking, finance and of course communication overall.
Most of these ideas weren't exactly new in 1997, but go back to services like CompuServe and even Douglas Engelbart's Mother of All Demos. The bottlenecks were bandwidth and personal computer performance (both of which were then predictably following Moore's law), not human imagination.
A few examples that a lot of people correctly extrapolated from: NLS (1968), PictureTel (1987) and later LiveShare, IndyCam (1993), CUSeeMee (1995), RealAudio (1995), RealVideo (1997).
Perhaps the core business problem with LLM:s isn't finding a product-market fit, but that our imaginations have been running wild with expectations on "AI" since at least the 1950s, and now we have something that quacks - but doesn't quite walk - like a duck.
A better comparison is actually AWS/Azure/Google Cloud/NeoClouds to AT&T and Verizon. The data centers follow a standard (CUDA/PyTorch/etc.) while OpenAI and Anthropic are becoming more like iOS and Android. Both the clouds and telecoms had to spend a ton of capex to build out infrastructure first.
Because of what I think is a poor comparison, the the next few slides make the wrong conclusions. For example, it thinks that models will be a commodity like 5G data. I disagree. I think frontier models are a classic duopoly/monopoly scenario. The smarter the model, the more it gets used, the more revenue it generates, the more compute the company can buy, the smarter the next model and so on. It's a flywheel effect. This is similar to advanced chip nodes like TSMC where your current node has to make enough money to pay for the next node. TSMC owns something like 95%+ of all of the most advanced node market. Back in the 80s and 90s, you had dozens of chip fab companies. Today, there are only 3. There should only be 1 but national security saved Intel and Samsung fabs.
There is evidence that the Chinese models are falling further behind, not gaining. Consolidation will likely happen soon because many unprofitable open source labs will have to merge and focus on revenue generation.
* Hardware era (pre 1995s) -> IBM, Intel, Microsoft, Apple
* Internet era (1994-2001) -> Amazon, Google, Meta, Salesforce
* Mobile era (iPhone+ era) -> Uber, Mobile Games, Youtube, Snapchat, Tiktok, Airbnb
* Cloud era (AWS+ era) -> AWS, GCP, Azure, Snowflake, Databricks and bunch of other data & database startups
AI era (ChatGPT+ era) -> Change is inevitable
Edit: I hadn't seen the original presentation yet. I see that Evans already divided the eras like I suggest above.
Change might be inevitable, but I'm not sure your list shows or proves that.
AI era will get its own winners, but there will be some new big players as a result of this era I think
I don't think that was implied at all, just that the context of the web is what allowed those companies to pop up.
Meta, née Facebook, wasn’t started until 2004.
Models were always going to be the commodity, just like the most popular and viable use cases at present are less job-replacement than "let's analyze huge data sets for patterns we're missing, and adjust accordingly" or "probabilistically generate deterministic software for me for X function/task". One-offs simply aren't profitable when models are interchangeable commodities, hence that brief attempt to pivot to "pay by outcome" before giddily embracing the classic consumption-based-billing playbook.
Not an uncommon event - not only did this happen to many companies who were big in the original internet boom (e.g. Sun Microsystems, as well as all the Boo, Pets.com etc), it also happened to the railway boom of the previous century, and even the Channel Tunnel.
I always considered jokingly that I am "selling" my intelligence when I work for a company. This clarifies that my perception wasn't far off.
(Unless the job gets you burned out, in which case it was selling indeed.)
In a way this is like distilling (but it is not) but you can make better harness (tackle more edge cases, better tool/function definitions, sandbox handling, bash management, DB management, deployment management, etc.) but extracting what LLMs know into code.
Maybe I am wrong but I would like to see custom software for the last mile (tiny/small businesses) becoming a reality. AI would eat the world of software but costs would go down since you can extract value upstream from the LLMs and spread downstream through tighter coding agents.
I am building a coding agent that will not be small - it will be a lot of code, carefully mixed roles (mimic a software dev shop) with separate tools available to different roles. And all this code is generated by other coding agents. https://github.com/brainless/nocodo
I am a nobody from nowhere with 18 years of software engineering behind me. I do not care about revenue. I just want to see a regular business owner's workflow going live on their own VPS.
why is it multiplied by 13?
So the two best metrics are annualized recurring revenue (take last month * 12 or last 4 weeks * 13) and QoQ growth %.
There are two caveats:
- If the revenue is high quality (e.g. annual enterprise contracts, good NRR), then last 4 weeks * 13 is actually a conservative estimate as your company will likely continue to grow.
- But if the revenue is more volatile (e.g. consumption, token usage, bad NRR) then annualized recurring revenue can be used to hide worse performance because companies will juice revenue one month and report high "ARR"
This is a marketing Gish Gallop talk that pretends to invalidate counterarguments with a couple of fantasy graphs.
To quite Ilya Sutskever:
> I think it’s pretty likely the entire surface of the earth will be covered with solar panels and data centers.
Technology is meant to serve us not drive us into a hellscape lol
None of "us" gets to decide this. Only the very wealthy get to decide.
Here's a test to know if this is/will be true: look for a situation where the "needs" of AI (e.g. land, electricity, etc) conflict with the needs of people (e.g. land to live on, grow food on, electricity to light our homes).
Find a place where the needs of AI conflict with people, and observe who wins out.
Does the entity that owns the datacenter say, "oh sorry! I guess we're using too much electricity. No worries! We'll stop doing that" ...or does it say, "lol too bad, all the electricity belongs to us!"
Does the entity wanting to build a datacenter say, "oh sorry! We thought you'd be okay with us using this land. But if you're not that's okay, we wont build here" ...or does it say, "lol too bad, we own the government and they're seizing the land under eminent domain!"
(both of these scenarios have happened, btw)
I'm old so my computer career has gone: punch cards => calculators => command-line => GUI => touch screen => voice => chat. Chat seems to be the best blend of expressiveness and utility, with a dose of magic thrown in.
> "What happened the last time that everything changed?"
Honestly, I'm glad we hear more of the commoditization of AI, and I hope that the comparison of AI with water or electricity will become mainstream and that the states (as in nation states) will understand that sooner rather than later and act accordingly.
That's a great quote.
So yeah, the next stage is models that basically do what humans do: encode causal models of the world in a composable, symbolic form that can be falsified and refined through interventional experiments.
Yes, and: we concluded that enough of reality doesn't work like that. The formal reasoning space is very powerful, but all the stuff we're really interested in has enough ambiguity and generalisation in that you can't cover it with a "small" set of rules.
Maybe if you had a really large number of rules? And used matrix multiplication to make sure that you covered all the marginal interactions between every possible set of rules? And then had some means of looking back on both output and input to constrain it towards things that were relevant? Wait a minute ...
Unsupervised learning over domain rulesystems has the potential to let us define really well-defined, scoped models that behave a lot more deterministically and don't colour outside the lines, and reserve their weights for cleanly modeling the domain associations and relationships that matter.
I just asked codex the following question in the middle of my coding prompt:
What are you thoughts on the relative strengths of ewoks vs jawans?
• Ewoks are stronger in direct conflict. They are organized fighters, good at
ambushes, traps, terrain control, and coordinated attacks. On Endor, the beat
a technologically superior force by using preparation and local knowledge.
....
Mixture-of-Experts seems like an attempt to do this - the domain structure being extracted into specific sub-models that are presumably trained on particular domain-associated content - but it feels like this is once again the beginnings of what is possible.
Take quantum computers for example, a lot of the time people will compare that to the dawn of classical computing, with claims such as "we can't know yet what we'll be able to achieve, we have to build it first!". Except that even the first classical computers were built with goals and applications in mind. Turing's was to decrypt Nazi codes, for example. Instead, when asking a quantum computing company what they're trying to achieve, they'll gesture vaguely at "chemistry, finance, ecology".
I agree that there are a lot of overhyped technologies though. Quantum computing has been in the works for decades now, with little to show for it in the popular perception.
I think the problem is a little bit more subtle:
To finance a lot of innovations, better also some intermediate step towards the far goal should already be very useful, otherwise the company that builds it will go bankrupt.
If this is not the case, it's typically not commercially viable, some product category is typically basic research (which is very important, but it typically means that the commercial potential will only come up in some future).
There do exist problems where a quantum computer gives an extreme advantage in the sense that we have no idea how a fast classical algorithm could look like. So, the only viable approaches for these problems are:
1. work on a huge algorithmic breakthrough (to be able to solve these problems fast on a classical computer)
2. build a quantum computer
What are these problems?
They are basically all special cases of the abelian hidden subgroup problem:
> https://en.wikipedia.org/w/index.php?title=Hidden_subgroup_p...
In particular cf. the table at the end of this Wikipedia article:
> https://en.wikipedia.org/w/index.php?title=Hidden_subgroup_p...
If you do have such a problem to solve, 1 and 2 are the only viable approaches.
So, there do exist goals and applications for which a quantum computer is insanely useful (assuming no huge algorithmic breakthrough happens).
The questions are thus:
- Is the abelian hidden subgroup problem sufficient for being able to carry a whole potential industry?
- (To come back to my introduction) What use does a quantum computer that is only capable of solving very small instances of this problem have for the user?
I'm aware that telecoms networks are standardised (I was once a telecoms analyst), but that isn't a precondition for a commodity.
It turns out that fabs follow Rock's Law which is that the capital cost to build a new fab doubles every 4 years. This means it will quickly get rid of the less competitive players. This is not dissimilar to the LLM scaling laws where you need a magnitude more compute to get unlock a new tier of intelligence.
Today, Anthropic and OpenAI are clearly in the lead for models and then there is everyone else. Google is a close 3rd. No one else is challenging them anymore in SOTA models. Some models might beat them in one or two benchmarks but none can compete overall. I expect this gap to grow bigger as models cost more and more to train.
1. Ex-ByteDance Engineer said so recently: https://www.businessinsider.com/ex-bytedance-engineer-says-c...
2. A benchmark concludes so: https://www.nist.gov/news-events/news/2026/05/caisi-evaluati...
3. Generally, people say Chinese models bench better than they perform
I think it seems to make sense given that China does not have access to Blackwell while in the past, they at least had access to gimped H200s.
(365/7)/12 = 4.3452…
Who made Ilya Sutskever, or any other LLM-bonehead the Grand Prophet of Humanity? Why the fuck is his opinion on that relevant? Of course he will shill for data centers.
no disagree, specifics matter.. There are a dozen well-defined LLM application subject areas that are regularly tested.. one overall grade IMO lacks important detail.. To go a bit abstract, it is ironic that "oversimplification" in the discussion of these complex machines mirrors the effects on information of the automations themselves.. constantly simplifying, substituting and diluting real meaning
I suppose a concrete example in 1997 would be that a lot of companies thought the future of e-commerce was setting up a store on AOL, that people would use while sitting down at a desktop PC. Obviously it didn’t turn out quite that way. Furthermore, the Internet enabled new kinds of ways to buy things that weren’t even envisioned in the pre-Internet pre-smartphone world: think Airbnb and Uber.
Predictions are hard, especially about the future. Most predictions reflect the worldview and biases of the time in which they are made: think about all the vintage sci-fi from the 60s 70s and 80s that actually reads or looks kind of retro now. Similarly, our predictions of the future will look kind of retro and strange to someone living in the 2030s or 2040s. If studying history has any lesson to teach us, it’s really just this: that the past is an alien world with alien moods of thinking, and that our moment in time will look similarly alien to people in the future who choose to look back and analyze it closely.
This isn’t an argument that we should stop trying to make predictions. We need to, but it is an argument for humility, and also for questioning all your assumptions that you might be importing.
For example, now it may seem that the models are becoming mere infrastructure, and the value moves up to apps and data. But if the models of tomorrow become able to write the apps themselves, then the value moves back. I won't need to pay some to write me a wrapper for the LLM, if the LLM will be able to write the same wrapper, maybe even better because it will be customized for my needs. The app providers are currently profiting from the gap between "what a software company can do using the AI" and "what the AI can do unaided", but that gap is going to shrink, possibly to zero.
But it still has huge gaps in quality. And from time to time, it shows me that it doesn’t really understand things. You might point out that how is that any different from your mediocre engineer. But for most people skilled enough, you can easily know the difference when someone doesn’t really know something.
With AI, you discover this after reading several pages being dumped on you by people being “more productive” with AI.
But I feel like people are more hyped about what the AI will be able to do soon rather than what it can do now.
I think AI does understand things (depending on your definition), how else could we communicate and ask it a question if it didn't? I mean we're quite far from Eliza here.
And yes, often their answer would be so wrong that we think it is impossible that AI understands anything, but this jagged intelligence doesn't prove, at least to me, that there isn't some understanding. At what point do we say that AI understands things? What if we can reduce 99% of those dumb failures, would we then say than AI understands?
No, they just say you are using the wrong model or something.
If it's a coworker dumping reviews of crap code on you at work, the incentive is to blanket approve everything because otherwise you're just the grumpy old man who is resisting innovation. No matter that the code makes no sense at all and the tests aren't actually testing what they should test.
Other than the stock market (which seems decoupled from reality at the moment), where has AI delivered?
The only use case where I see anything resembling AI delivering on it's promises is software, and my personal experience with that is that everything that comes out of the teams using AI is destructively broken. (Where they used to be able to deliver software that worked, even if it wasn't ideal, now they reliably make things worse and their stuff doesn't work when used.)
I think one of the things that the usage data shows us is that chatbots absolutely do not have infinite use cases - most users only use them a day or two a week or less.
But I also do disagree with the take that usage patterns indicate a fundamental shortage of use-cases. Yes, everyone reports WAU instead of DAU because WAU numbers look much more impressive, but I think the extreme shortage of compute plays a major role in this. I suspect all the AI labs are deliberately holding back from pushing AI adoption too much because of this. (Google execs have even made comments internally to this effect.) Note that even at such low frequency of usage all the model providers are desperately strapped for compute, which means there is insanely high demand from some quarters.
One way how capacity limitations could impact adoption is that the free-tier models are not as good as the frontier ones, so the free users come away less impressed with AI capabilities, leading to lower regular usage. This problem is larger than it appears, because it can take a long time to figure out how to get AI to work for your use-case, and people simply have not experimented nearly enough, partially due to first impressions. On the other hand, most companies seem to be OK with huge tokenmaxxing bills!
It seems to me the AI players are all playing a delicate balancing game across three fundamental dimensions: adoption, monetization, capacity. That is, they are simultaneously 1) pushing free / cheap AI usage as much as possible to hook users, capture market share and suss out new use-cases, while 2) carefully allocating token quotas for the most lucrative use-cases to satisfy investors, and 3) balancing available compute between those two competing priorities. I suspect as the compute bottleneck is alleviated and frontier models become more accessible cheaply, we'll see way higher DAU numbers.
The industrial revolution didn't seem to require any particular special skill at all. Just anyone who was willing to tend to a machine all day. (Maybe that's a parallel...)
Specifically, the 3 dimensions along with new jobs required new skills were: a) cognitive, b) technical or c) social skills. I guess tending to a machine was a mix of a) and b), because even if the controls were straightforward, it probably required some understanding of the underlying mechanisms.
Don't be vague.
> AI data centers crowd us out.
> the entire surface of the earth will be covered with solar panels and data centers.
You'll think otherwise the first time you're a victim of a zero-day ewok.
Seriously though, while coding models may not need to know about ewoks, their contextual knowledge of things beyond just writing code almost certainly makes them better coding models.
It could be difficult to constrain the training corpus "just right" so that you eliminate all the irrelevant subjects like ewoks but retain enough so that the model doesn't turn into an idiot savant capable of churning out correct code but incapable of understanding what you really want.
If you need to do programming do we really need trillions sized models? Other domains might be large or smaller, but there's no need for a model to 'know' everything and datacenter levels of hardware to run.
General chatbots might work better as larger models since you really don't know what people will also for, or alternatively we find a way to route the initial question to the appropriate model. Like MoE but without needing to load a gigantic model into memory first.
This is a common miss-conception. MoE LLMs are NOT trained with each expert receiving domain-associated data. It's just an unfortunate naming decision that stuck, and is commonly miss-understood by non practitioners.
MoE is an architecture change meant to lower the total compute for both training and serving an LLM. You basically have many smaller models (unfortunately called experts) and a router on top of them. The router "learns" which expert to activate for the next token generation, but that doesn't need to follow any semantic association. For the same math problem you could get experts 1 and 234 activate on the first token, 5 and 132 on the 2nd token and so on.
The issue that is hit immediately is we don't have a definition or test of understanding that AI doesn't clear easily. Then on top of that we can't even really be sure that we ourselves are understand things given all the tricks that our minds play with memory and perception. There is precious little evidence that the people around us understand things, they seem to be guessing. It is completely unclear if a Chinese room has or doesn't have a property if we rule out all the tests that check for it as not really counting. But all the tests we can do suggest it does understand, because engineers can implement Chinese rooms now and they even turn out to be more reliably artistic/capable of novel thinking/creative than humans. Anything that tests understanding they can do.
So long as there is demand, there are always going to be providers competing to offer it at a low cost. My understanding is that the median price on there is in the ballpark of what it costs to run the inference. This is very different from e.g. Opus, which you can basically only buy from Anthropic at the price they set.
https://bsky.app/profile/antirez.bsky.social/post/3mlzwmvlov...
It's much closer than you think. We're going to see specialized hardware in the next 24 months capable of running 2025-era frontier models. That's big.
Once and if the advancements with the AI models slow down, only then IMHO it will become feasible to design the specialized HW for general-purpose consumption and general-purpose workloads.
Even at 2-bit quantization, DS4 is probably on par with a 2024 frontier model. You can run that today on local hardware, and at a minimum, local models are going to keep pace over the next 12-24 months. Even if they don't close the gap with frontier models, they'll still play an important role in the overall pipeline for cost, speed and privacy reasons.
That's without even mentioning the additional capability that something like a Taalas chip churning out 17k tokens/sec could unlock.
Even if it were possible the LLMs are such a gold mine of user data. It's really hard to see that opportunity be passed up.
https://www.apple.com/shop/buy-mac/mac-studio
Same with the Mac mini. entirely removed from all store references
It feels great to finally have access to something local.
Even if this isn't true, comparing telecom bits to tokens is wrong. Bits are the same no matter what telecom transfers them. Tokens are not all the same. The quality varies.
We're already seeing a massive divide between frontier models and lesser models in growth rates. Anthropic is adding $10b - $15b every month in ARR. This figure likely dwarfs open source labs. This is all because its models are maybe 10-15% better.
The cost to inference a 1T param frontier model is the same as a 1T param open source model. Therefore, if the frontier model is even 10-15% better, it will gobble up the market over time.
Lastly, even though Claude Code and Codex are the biggest revenue drivers for Anthropic and OpenAI today, I don't believe this will be true 2-5 year from now. I believe selling their tokens via API will be their biggest. The sum of applications in the world will dwarf coding in market size. For example, biotech, finance, physics, engineering, robotics, sensor data, etc. This is why I think OpenAI and Anthropic are becoming more like iOS and Android than AT&T and Verizon. Applications will build on top of OpenAI and Anthropic just like iOS and Android.
[0]https://epoch.ai/blog/training-compute-of-frontier-ai-models...
OpenAI and Anthropic don't compete in the LLM commodity market. Hence, I had a problem with slide 22.
The reason mobile data had to standardize is because it’s a network and a network must have protocols. It’s useless without them.
Intelligence must have interfaces, and those can be standardized. Businesses will try to remain provider agnostic, which will also drive standardization via standard sales and marketing methods.
Separately, we are doing our best to standardize performances on benchmarks.
I don’t disagree that right now transport of standardized mobile data vs emulation of human intelligence is qualitatively different, but perhaps primarily because it is early in development, and our vantage point this time is relatively from within the network, instead of outside it.