We just had a realization during a demo call the other day:
The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up. Not being dependent on LLMs for your fundamental product’s value will be a major advantage, at least in pricing.
Yup. Also regardless of price they need to spend more and more as the project collapses under the inevitable incidental complexity of 30k lines of code a day.
It's similar to how if you know what you're doing you can manage a simple VPS and scale a lot more cost effectively than something like vercel.
In a saturated market margins are everything. You can't necessarily afford to be giving all your margins to anthropic and vercel.
Look no further to be honest; look at older generation programming languages like COBOL and how sought-after good developers for that language are.
But I'm also afraid / certain that LLMs are able to figure out legacy code (as long as enough fits in their context window), so it's tenuous at best.
Also, funny you mentioned HTML / CSS because for a while (...in the 90's / 2000's) it looked like nobody needed to actually learn those because of tools like Dreamweaver / Frontpage.
The issue with COBOL code is that it’s hidden. It’s mostly internal systems so little code available for training. HTML, TypeScript, JavaScript, C, etc, are readily available, billions of code lines.
Well, on the 2nd paragraph, I have no illusion they’ll figure out more as they are being trained. I am more thinking of the custodians (as coders turn into that)
Say you are a good coder now, but you are becoming a custodian, checking the llm work will slowly erode your skills. Maybe if you got a good memory or an amazing skillset it might be some time, but if you don’t use it, you lose it.
How are COBOL developers "sought after"? That's an oft-repeated but woefully incorrect meme.
FAANG new grads make more. If the COBOL devs had upskilled throughout their career they'd be Senior Staff/Principal+ and making 5-10x more than they do today.
> The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up.
It's not that clear. Sure, hardware prices are going up due to the extremely tight supply, but AI models are also improving quickly to the point where a cheap mid-level model today does what the frontier model did a year ago. For the very largest models, I think the latter effect dominates quite easily.
>> The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up.
> It's not that clear. Sure, hardware prices are going up due to the extremely tight supply, but AI models are also improving quickly to the point where a cheap mid-level model today does what the frontier model did a year ago.
I agree; I got some coding value out of Qwen for $10/m (unlimited tokens); a nice harness (and some tight coding practices) lowers the distance between SOTA and 6mo second-tier models.
If I can get 80% of the way to Anthropic's or OpenAI's SOTA models using 10$/m with unlimited tokens, guess what I am going to do...
There's only so far engineers can optimise the underlying transformer technique, which is and always has been doing all the heavy lifting in the recent ai boom. It's going to take another genius to move this forward. We might see improvements here and there but the magnitudes of the data and vram requirements I don't think will change significantly
State space models are already being combined with transformers to form new hybrid models. The state-space part of the architecture is weaker in retrieving information from context (can't find a needle in the haystack as context gets longer, the details effectively get compressed away as everything has to fit in a fixed size) but computationally it's quite strong, O(N) not O(N^2).
I’ve read and heard from Semi Analysis and other best-in-class analysts that the amount of software optimizations possible up and down the stack is staggering…
How do you explain that capabilities being equal, the cost per token is going down dramatically?
Optimizations, like I said. They'll never hack away the massive memory requirements however, or the pre training... Imagine the memory requirements without the pre training step....this is just part and parcel of the transformer architecture.
And a lot of these improvements are really just classic automation or chaining together yet more transformer architectures, to fix issues the transformer architecture creates in the first place (hallucinations, limited context)
Inference prices droped like 90 percent in that time (a combination of cheaper models, implicit caching, service levels, different providers and other optimizations).
Quality went up. Quantity of results went up. Speed went up.
Service level that we provide to our clients went up massively and justfied better deals. Headcount went down.
The decline of independent thoughts for one. As people become reliant on LLMs to do their thinking for them and solve all problems that they stumble upon, they become a shell of their previous self.
There is no decline. Human assets were always too expensive to process some additional information. We are simply processing lot more of low signal data.
Actually some of our analysts are empowered by the tools at their disposal. Their jobs are safe and necessary. Others were let go.
Clients are happy to get fuller picture of their universe, which drives more informed decissions . Everybody wins.
You also have to look at how exposed your vendors are to cost increases as well.
Your company may have the resources to effectively shift to cheaper models without service degradation, but your AI tooling vendors might not. If you pay for 5 different AI-driven tools, that's 5 different ways your upstream costs may increase that you'll need to pass on to customers as well.
> Not being dependent on LLMs for your fundamental product’s value
I think more specifically not being dependent on someone else's LLM hardware. IMO having OSS models on dedicated hardware could still be plenty viable for many businesses, granted it'll be some time before future OSS reaches today's SOTA models in performance.
What's weird though is the bifurcation in pricing in the market: aka if your app can function on a non-frontier level AI you can use last years model at a fraction of the cost.
That'll be (part of) the big market correction, but also speaking broadly; as investor money dries up and said investors want to see results, many new businesses or products will realise they're not financially viable.
On a small scale that's a tragedy, but there's plenty of analysts that predict an economic crash and recession because there's trillions invested in this technology.
Indeed, it was clear from the beginning, "AI" companies want to become infrastructure and a critical dependency for businesses, so they can capture the market and charge whatever they want. They will have all the capital and data needed to eventually swallow those businesses too, or more likely sell it to anyone who wants the competitive advantage.
in fact I am betting opposite. frontier models are getting not THAT much better anymore at all, for common business needs at least. but the OSS models keep closing the gap. which means if trajectories hold there will be a near future moment probably where the big provider costs suddenly drop shaerply once the first viable local models consistently can take over tasks normally on reasonable hardware. Right now probably frontier providers rush for as much money as they possible can before LLMs become a true commodity for the 80% usecases outside of deep expert areas they will have an edge over as specialist juggernauts (iE a cybersecurity premium model).
So its all a house of cards now, and the moment the bubble bursts is when local open inference has closed the gap. looks like chinese and smaller players already go hard into this direction.
Local open inference can address hardware scarcity by repurposing the existing hardware that users need anyway for their other purposes. But since that hardware is a lot weaker than a proper datacenter setup, it will mostly be useful for running non-time-critical inference as a batch task.
Many users will also seek to go local as insurance against rug pulls from the proprietary models side (We're not quite sure if the third-party inference market will grow enough to provide robust competition), but ultimately if you want to make good utilization of your hardware as a single user you'll also be pushed towards mostly running long batch tasks, not realtime chat (except tiny models) or human-assisted coding.
Absolutely. Pricing exposure is the quiet story under all the waves of AI hype. Build for convenience → subsidise for dependence → meter for margin is a well-worn playbook, and AI-dependent companies are about to find out what phase three feels like.
Hyperscalers are spending a fortune so we think AI = API, but renting intelligence is a business model, not a technical inevitability.
How is that surprising? We've been taking that into account for any LLM related tooling for over a year now that we either can drop it, or have it designed in a way that we can switch to a selfhosted model when throwing money at hardware would pay for itself quickly.
It's just another instance of cloud dependency, and people should've learned something from that over the last two decades.
Not so much that it was surprising, rather that we looked at a competitor’s site and noticed that a) their prices went way up and b) their branding changed to be heavily AI-first.
So we thought, hmm, “wonder if they are increasing prices to deal with AI costs,” and then projected that into a future where costs go up.
We don’t have this dependence ourselves, so this seems to be a competitive advantage for us on pricing.
I wonder if it could be that they won't because the real mechanism is that AI wrapper pricing power is weak (switching costs near zero) but state of the art models makes it difficult to lower prices due to higher cost.
Also: AI dependance could be explicit AI API usage by the product itself, but also anything else, like: AI assisted coding, AI used by humans in other surrounding workflows, etc.
And I don't really mean new businesses that are entirely built around LLMs, rather existing ones that pivoted to be LLM-dependent – yet still have non-LLM-dependent competitors.
same as Uber… in the beginning everyone pretty much new that the cost of rides cannot possibly be that cheap and that it is subsudized. once you corner the market etc people just got used to “real” prices to the poibt that now there are often cheaper alternatives than Uber but people still Uber…
Its also quite interesting to read about Uber exploits their drivers and discriminating algorithms. Cory Doctorow mentioned it in his latest book, sadly cant link the direct sources.
Not really, the next move is to establish standards groups requiring the use of AI in product development. A mix of industry and governmental mandates. What you view are viewing as COGS instead becomes instead a barrier to entry.
China already operates like this. Low cost specialized models are the name of the game. Cheaper to train, easy to deploy.
The US has a problem of too much money leading to wasteful spending.
If we go back to the 80s/90s, remember OS/2 vs Windows. OS/2 had more resources, more money behind it, more developers, and they built a bigger system that took more resources to run.
Mac vs Lisa. Mac team had constraints, Lisa team didn't.
Though I do agree with you, I just came back from a trip to China (Shanghai more specifically) and while attending a couple AI events, the overwhelming majority of people there were using VPNs to access Claude code and codex :-/
On the Mac vs Lisa team, I generally agree but wasn't there a strong tension on budget vs revenue on Mac vs Apple II? And that Apple II had even more constrained budget per machine sold which led to the conflict between Mac and Apple II teams. (Apple II team: "We bring in all the revenue+profit, we offer color monitors, we serve businesses and schools at scale. Meanwhile, Steve's Mac pirate ship is a money pit that also mocks us as the boring Navy establishment when we are all one company!")
By the logic of constraints (on a unit basis), Apple II should have continued to dominate Mac sales through the early 90s but the opposite happened.
You can fine-tune a model, but there are also smaller models fine-tuned for specific work like structured output and tool calling. You can build automated workflows that are largely deterministic and only slot in these models where you specifically need an LLM to do a bit of inference. If frontier models are a sledgehammer, this approach is the scalpel.
A common example would be that people are moving tasks from their OpenClaw setup off of expensive Anthropic APIs onto cheaper models for simple tasks like tagging emails, summarizing articles, etc.
Combined with memory systems, internal APIs, or just good documentation, a lot of tasks don't actually require much compute.
It has been a very bad bet that hardware will not evolve to exceed the performance requirements of today's software tomorrow, just as it is a bad bet that tomorrow someone will rewrite today's software to be slower.
Eh, but then as hardware evolves, the software will also follow suit. We’ve had an explosion of compute performance and yet software is crawling for the same tasks we did a decade ago.
Better hardware ensures that software that is “finished” today will run at acceptable levels of performance in the future, and nothing more.
I think we won’t see software performance improve until real constraints are put on the teams writing it and leaders who prioritize performance as a North Star for their product roadmap. Good luck selling that to VCs though.
Harness is a big one, Claude Code still has trouble editing files with tabs. I wonder how many tokens per day are wasted on Claude attempting multiple times to edit a file.
As a recent example in AI space itself. China had scarce GPU resources, quite obvious why => DeepSeek training team had to invent some wheels and jump through some hoops => some of those methods have since become 'industry standard' and adopted by western labs who are now jumping through the same hoops despite enjoying massive computeresources, for the sake of added efficiency.
I'm having an hard time getting my mind to see this.
> Users should re-tune their prompts and harnesses accordingly.
I read this in the press release and my mind thought it meant test harness. Then there was a blog post about long running harnesses with a section about testing which lead me to a little more confusion.
Yes, the word 'harness' is consistently used in the context as a wrapper around the LLM model not as 'test harness'.
This field is chock full of people using terms incorrectly, defining new words for things that already had well known names, overloading terms already in use. E.g. shard vs partition. TUI which already meant "telephony user interface ". "Client" to mean "server" in blockchain.
Some people also call evaluations "tests". There are unexpected things that come along with new models, like the model in a workflow you'd set up suddenly starts calling a tool and never stops or decides to no longer call a particular tool, so running your existing evaluations to catch regressions like this and potentially updating the prompts is considered "testing" your prompts and harnesses.
It’s the tool that calls the model, give it access to the local file system, calls the actual tools and commands for the model, etc, and provide the initial system prompt.
Basically a clever wrapper around the Anthropic / OpenAI / whatever provider api or local inference calls.
pi vs. claude code vs. codex
These are all agent harnesses which run a model (in pi's case, any model) with a system prompt and their own default set of tools.
Whoever running and selling their own models with inference is invested into the last dime available in the market.
Those valuations are already ridiculously high be it Anthropic or OpenAI to the tune of couple of trillion dollars easily if combind.
All that investment is seeking return. Correct me if I'm wrong.
Developers and software companies are the only serious users because they (mostly) review output of these models out of both culture and necessity.
Anywhere else? Other fields? There these models aren't any useful or as useful while revenue from software companies by no means going to bring returns to the trillion dollar valuations. Correct me if I'm wrong.
To make the matter worst, there's a hole in the bucket in form of open weight models. When squeezed further, software companies would either deploy open weight models or would resort to writing code by hand because that's a very skilled and hardworking tribe they've been doing this all their lives, whole careers are built on that. Correct me if I'm wrong.
Eventually - ROI might not be what VCs expect and constant losses might lead to bankruptcies and all that build out of data centers all of sudden would be looking for someone to rent that compute capacity result of which would be dime a dozen open weight model providers with generous usage tiers to capitalize on that available compute capacity owners of which have gone bankrupt and can't use it any more wanting to liquidate it as much as possible to recoup as much investment as possible.
It feels like a repeat of the dot com infrastructure buildup that spurred the whole 2005 explosion in affordable hosting and new companies. This will probably leave us massive access to affordable compute in a couple of years.
No matter how low and reasonably Anthropic is valued, don't think $200 Max plans are going to recoup the investment + some return on top because size of the software industry is not that huge and profit margins for AI inference aren't very high either.
> because size of the software industry is not that huge
I onboarded marketing on a premium team Claude seat yesterday. And one of our sales vibecoded an internal tool in the last three weeks using Claude Code that they now use every day. I wouldn’t have imagined it a month ago. We still had to take care of deployment for him, but things are moving fast.
Seems like everybody an their mothers are using max plans these days. I wouldn't be surprised if LTV of each customer was big enough to justify spending.
Assuming there are 10 million developers and everyone is at $200 max plan, that would be $2 billion/month or $24 billion/year maximum.
Note - this is just the revenue not the profit. No salaries, no compute paid for. Just plain revenue. Profit would be way less.
But even that - if we take it to $24 billion/year and we take a 10x multiple, the company is barely valued at $240 billon dollar, lets be generous and make it double at $480 billion and then round it up to $500 billion for a nice round number.
Far far from the $800 billion valuation Anthropic is looking at.
Companies are spending far more than $200/month/developer. The $200 Max plan is a great value but you hit limits far too soon, and it also doesn't cover any of the other styles of integrations and tools that you can build and use to help your developers, like code review suggestions, which at the very least would come from additional Max plans, and not from the individual developers' plans.
Not counterproductive because people aren't just sitting back in the rest of the time while AI does work. They do more work. $3k per head on Claude is nothing if your devs get 2x more work done.
Do they review the code? Because in my experience using Claude Opus 4.6 generates code that would be buggy and the tests would be written agains that buggy code with wrong assumptions that certainly would pass with flying colors.
It is only when you look closed you get to know what the hell has happened!
And nor are they trying to. If you are spending $200 a month, you are a mere tolerated nuisance. The company I work for gives every developer a $5000 a month allowance to Claude and I think there are around 500-600 people eligible for it.
comments like these is probably while significant contingent of HN talks heavily about AI bubble. burning that kind of cash always ends up like the right move, godspeed :)
I work for a consulting company, we know exactly how much each billable person makes the company and whether the ROI is worth it.
We don’t have to measure “productivity gains”.
I personally don’t come close to that and neither do I suspect most of us. Between using my $20 a month ChatGPT subscription with Codex and the amount of time I spend on Zoom calls “adding on to what Becky said” and “looking at things from the 1000 foot view”
On another note, the grunt work I use to delegate to a junior consultant, I now can get Claude to do in a fraction of the the time. They were making a lot more than the worse case of $72K a year fully allocated. But honestly most of that work is done with my $240 a year ChatGPT subscription + maybe $600 in Claude at the current prices/limits
There is a lot of demand still coming for sure but I think I'm more optimistic. Ready to eat my hat on this but
- higher prices will result in huge demand destruction too. Currently we're burning a lot of tokens just because they're cheap, but a lot of heavy users are going to spend the time moving flows over to Haiku or onprem micro models the moment pricing becomes a topic.
- data centers do not take that long to build, probably there are bottlenecks in weird places like transformers that will cause some hicups, but nvidia's new stuff is waay more efficient and the overall pipeline of stuff coming online is massive.
- probably we will see some more optimization at the harness level still for better caching, better mix of smaller models for some use, etc etc.
These companies have so much money and they at least anthropic and openai are playing winner takes it all stakes, with competition from the smaller players too. I think they're going to be feeding us for free to win favour for quite a while still.
I agree and I am amazed at how much money some individuals and also a friend's company burn on token costs. I get huge benefits from this tech just using gemini-cli and Antigravity a few times a week, briefly. I also currently invest about $15/month in GLM-5.1 running Hermes Agent on a small dedicated VPS - fantastically good value for getting stuff done and this requires little of my time besides planning what I need done.
I think the token burners are doing it wrong. I think that long term it is better to move a little slower, do most analysis and thinking myself, and just use AI when the benefits are large while taking little of my time and money to use the tools.
This is probably even the "fun" part of the whole picture. The purely dystopia starts when investment firms just silently grow bigger and bigger data centers like cancer. There will be no press releases, no papers, no chance anyone without billions will even know the details yet alone get access. One day we realise the worlds resources (maybe not as in the paperclip maximiser, but as in memory, energy, GPUs, water, locations) are consumed by trading models and the data centres are already guarded by robot armies. While we were distracted frighting with anthropic and openAI the real war was already over. Mythos is one sign in this direction but i also met a few people who were claiming to fund fairly large research and training operations just by internal models working on financial markets. I have no way to verify those claims but this happened 3 times now and the papers/research they were working on looked pretty solid and did not seem like they were running kimi openclaw on polymarket but actual models on some significant funds. Would be really interested if anyone here has some details on this reality. I would also not be surprised if this is a thing that people in SF just claim to sound dangerous and powerful.
one graph, One graph and the author is pinning an entire theory on it?
Infra is always limited, even at hyper scalers. This leads to a bunch of tools dfofr caching, profiling and generally getting performance up, not to mention binpacking and all sorts of other "obvious" things.
I think maybe infra is limited only at hyperscalers. For the rest of us it's just how much capacity to we want to rent from the hyperscalars.
It's kind of a recent cloud-native mindset, since back in the day when you ran your own hardware scaling and capacity was always top of mind. Looks like AI compute might be like that again, for the time being.
Well it's in the books. O(n^2) algorithms are bad in the long run, transformers algorithm has such complexity, so not a big surprise we hit the limits.
China is installing something like 500 GW of wind and solar per year now. Even if they're only able to build and otherwise access chips that have half the SoTA performance per watt, they will win.
A dollar is an entirely fictional unit and trillions of it can be manufactured at no cost, while watts are constrained by the laws of physics, photons/electrons, supply chain of electricity and all that fun stuff in the real world.
A dollar is still a useful unit as "the fraction of the economy that can be controlled by currency". It's true that printing a huge pile of it and throwing it at GPUs wouldn't instantly convert into more GPUs, but it would meaningfully represent that other things are being squeezed out to allocate more resources to GPU production even so. That such reallocation is inefficient, arguably immoral, and highly questionable in the long term versus other options wouldn't stop that from being ture.
>dollar is an entirely fictional unit and trillions of it can be manufactured at no cost
If the abstraction works better for you this way, call them interchangeable units of American and Chinese insolvency. Or incremental forfeiture of domestic ownership.
I wouldn’t agree. Even at national scale, these projects cost resources. And the resources of all agents (org, countries) are constrained.
While we could reason in "performance / watt" and "performance / people", "performance / whatever other resource involved", and "performance / opportunity cost of allocating these resources to this use case and not another", "performance / whatever unit of stable-ish currency" is a convenient and often "good enough" approximation that somewhat encapsulates them all.
A simplification, like any model, but still useful.
US energy is constrained by the utility monopolies/oligopolies which have to extract more rents, specifically by increasing costs. Their profit is a percentage of cost, these perverse incentives + oligopolies will make it increasingly expensive to make anything (including AI) in US.
The dynamics vastly favor China, part of the reason the US sprinting towards "ASI" isn't totally boneheaded is that the US and its industry needs a hail mary play to "win" the game, if they play it safe they lose for sure.
I'd be fine with a world without AI, honestly. Nobody really wins this race except the very wealthy. And I don't think it's really going to play out the way the wealthy think it will. It's more like a dog catching a car than it is a race.
"The dog that caught the car" refers to how dogs sometimes chase cars. Suppose the car stops and the dog catches up - what is it going to do? It has no plan, it has no purpose, it isn't going to bite the car, it isn't going to get anything out of catching the car. The car may even run it over. I intended it basically as "play stupid games, win stupid prizes", or "be careful what you wish for".
My observation is that the dog sniffs all the tires, picks one tire, lifts one leg and does the deed. I don't know if its a way of marking territory or domination. We need a dogatologist to explain what it means.
It seems very possible that we have at least five years of real limitations on compute coming up. Maybe ten, depending on ASML. I wonder what an overshoot looks like. I also wonder if there might be room for new entrants in a compute-scarce environment.
For instance, at some point, could Coreweave field a frontier team as it holds back 10% of its allocations over time? Pretty unusual situation.
AKA, the beginning of big companies being able to roll over small companies with moar money
(note: I don't expect this to actually happen until the AI gets good enough to either nearly entirely replace humans or solve cooperation, but the long term trend of scarce AI will go towards that direction)
The scarcity framing assumes compute is the bottleneck. For most production deployment's Ive seen, the actual bottleneck is evaluation and knowing what to trust.
You can throw cheaper models at a problem all day but, if you can't measure where the model fails on your data, You're just making mistakes faster at a lower cost.
Open Weight models are 6 months to a year behind SOTA. If you were building a company a year ago based on what AI could do then, you can build a company today with models that run locally on a user's computer. Yes that may mean requiring your customers to buy Macbooks or desktops with Nvidia GPUs, but if your product actually improves productivity by any reasonable amount, that purchase cost is quickly made up for.
I'll argue that for anything short of full computer control or writing code, the latest Qwen model will do fine. Heck you can get a customer service voice chat bot running in 8GB of VRAM + a couple gigs more for the ASR and TTS engine, and it'll be more powerful than the hundreds of millions spent on chat bots that were powered by GPT 4.x.
This is like arguing the age of personal computing was over because there weren't enough mainframes for people to telnet into.
It misses the point. Yes deployment and management of personal PCs was a lot harder than dumb terminal + mainframe, but the future was obvious.
The mainframe/PC analogy is spot on. And the hardware floor keeps dropping - you can grab a mini PC with 32-64GB RAM for a few hundred bucks and run surprisingly capable quantized models locally. Something like https://terminalbytes.com/best-mini-pcs-for-home-lab-2025/ shows the kind of hardware that's now available at consumer prices. The "scarcity" framing only makes sense if you assume everyone needs frontier-tier models for everything.
I've seen this claimed, but I'm not sure it's been true for my use cases? I should try a more involved analysis but so far open models seem much less even in their skills. I think this makes sense if a lot of them are built based on distillations of larger models. It seems likely that with task specific fine tuning this is true?
> I've seen this claimed, but I'm not sure it's been true for my use cases?
I'd be surprised if it isn't true for your use cases. If you give GLM-5.1 and Optus 4.6 the same coding task, they will both produce code that passes all the tests. In both cases the code will be crap, as no model I've seen produces good code. GLM-5.1 is actually slightly better at following instructions exactly than Optus 4.6 (but maybe not 4.7 - as that's an area they addressed).
I've asked GLM-5.1 and Opus 4.6 to find a bug caused by a subtle race condition (the race condition leads to a number being 15172580 instead of 15172579 after about 3 months of CPU time). Both found it, in a similar amount of time. Several senior engineers had stared at the code for literally days and didn't find it.
There is no doubt the models do vary in performance at various tasks, but we are talking the difference between Ferrari vs Mercedes in F1. While the differences are undeniable, this isn't the F1. Things take a year to change there. The performance of the models from Anthropic and OpenAI literally change day by day, often not due to the model itself but because of the horsepower those companies choose to give them on the day, or them tweaking their own system prompts. You can find no end of posts here from people screaming in frustration the thing that worked yesterday doesn't work today, or suddenly they find themselves running out of tokens, or their favoured tool is blocked. It's not at all obvious the differences between the open-source models and the proprietary ones are worse than those day to day ones the proprietary companies inflict on us.
If you don't know C, in older versions that can be a catastrophic failure. (The issue is so serious in modern C `free(NULL)` is a no-op.) If it's difficult to get a `FOO == NULL` without extensive mocking (this is often the case) most programmers won't do it, so it won't be caught by unit tests. The LLMs almost never get unit test coverage up high enough to catch issues like this without heavy prompting.
But that's the least of it. The models (all of them) are absolutely hopeless at DRY'ing out the code, and when they do turn it into spaghetti because they seem almost oblivious to isolation boundaries, even when they are spelt out to them.
None of this is a problem if you are vibe coding, but you can only do that when you're targeting a pretty low quality level. That's entirely appropriate in some cases of course, but when it isn't you need heavy reviews from skilled programmers. No senior engineer is going to stomach the repeated stretches of almost the "same but not quite" code they churn out.
You don't have to take my word for it. Try asking Google "do llm's produce verbose code".
C is fundamentally a bad target for LLMs. Humans get C wrong all the time, so we can not hope the nascent LLM, which has been trained on 95% code that does automatic memory management, to excel here.
I always found myself writing verbose copypasta code first, then compress it down based on the emerging commonalities. I think doing it the other way around is likely to lead to a worse design. Can you not tell the LLM to do the same? Honest question.
> I always found myself writing verbose copypasta code first, then compress it down based on the emerging commonalities. I think doing it the other way around is likely to lead to a worse design.
I do pretty much the same thing, which is to say I "write code using a brain dump", "look for commonalities that tickle the neurons", then "refactor". Lather, rinse, and repeat until I'm happy.
> Can you not tell the LLM to do the same?
You can tell them until you're blue in the face. They ignore you.
I'm sure this is a temporary phase. Once they solve the problem, coding will suffer the same fate as blacksmiths making nails. [0] To solve it they need to satisfy two conflicting goals - DRY the code out, while keeping interconnections between modules to a minimum. That isn't easy. In fact it's so hard people who do it well and can do it across scales are called senior software engineers. Once models master that trick, they won't be needed any more.
By "they" I mean "me".
[0] Blacksmiths could produce 1,000 or so a day, but it must have been a mind-numbing day even if it paid the bills. Then automation came along, and produced them at over a nail per second.
a) The agent doesn't need to read the implementation of anything - you can stuff the entire projects headers into the context and the LLM can have a better birds-eye view of what is there and what is not, and what goes where, etc.
and
b) Enforcing Parse, don't Validate using opaque types - the LLM writing a function that uses a user-defined composite datatype has no knowledge of the implementation, because it read only headers.
`free(NULL)` is harmless in C89 onwards. As I said, programmers freeing NULL caused so many issues they changed the API. It doesn't help that `malloc(0)` returns NULL on some platforms.
If you are writing code for an embedded platform with some random C compiler, all bets on what `free(NULL)` does are off. That means a cautious C programmer who doesn't know who will be using their code never allows NULL to be passed to `free()`.
In general, most good C programmers are good because they suffer a sort of PTSD from the injuries the language has inflicted on them in the past. If they aren't avoiding passing NULL to `free()`, they haven't suffered long enough to be good.
> That means a cautious C programmer who doesn't know who will be using their code never allows NULL to be passed to `free()`.
If your compiler chokes on `free(NULL)` you have bigger problems that no LLM (or human) can solve for you: you are using a compiler that was last maintained in the 80s!
If your C compiler doesn't adhere to the very first C standard published, the problem is not the quality of the code that is written.
> If they aren't avoiding passing NULL to `free()`, they haven't suffered long enough to be good.
I dunno; I've "suffered" since the mid-90s, and I will free NULL, because it is legal in the standard, and because I have not come across a compiler that does the wrong thing on `free(NULL)`.
So what would be the best practice in a situation like that? I would (naively?) imagine that a null pointer would mostly result from a malloc() or some other parts of the program failing, in which case would you not expect to see errors elsewhere?
> imagine that a null pointer would mostly result from a malloc() or some other parts of the program failing, in which case would you not expect to see errors elsewhere?
Oh yes, you probably will see errors elsewhere. If you are lucky it will happen immediately. But often enough millions of executed instructions later, in some unrelated routine that had its memory smashed. It's not "fun" figuring out what happened. It could be nothing - bit flips are a thing, and once you get the error rate low enough the frequency of bit flips and bugs starts to converge. You could waste days of your time chasing an alpha particle.
I saw the author of curl post some of this code here a while back. I immediately recognised the symptoms. Things like:
if (NULL == foo) { ... }
Every 2nd line was code like that. If you are wondering, he wrote `(NULL == foo)` in case he dropped an `=`, so it became `(NULL = foo)`. The second version is a syntax error, whereas `(foo = NULL)` is a runtime disaster. Most of it was unjustified, but he could not help himself. After years of dealing with C, he wrote code defensively - even if it wasn't needed. C is so fast and the compilers so good the coding style imposes little overhead.
Rust is popular because it gives you a similar result to C, but you don't need to have been beaten by 10 years of pain in order to produce safe Rust code. Sadly, it has other issues. Despite them, it's still the best C we have right now.
Write code? No. Use frontier models. They are subsidized and amazing and they get noticably better ever few months.
Literally anything else? Smaller models are fine. Classifiers, sentiment analysis, editing blog posts, tool calling, whatever. They go can through documents and extract information, summarize, etc. When making a voice chat system awhile back I used a cheap open weight model and just asked it "is the user done speaking yet" by passing transcripts of what had been spoken so far, and this was 2 years ago and a crappy cheap low weight model. Be creative.
I wouldn't trust them to do math, but you can tool call out to a calculator for that.
They are perfectly fine at holding conversations. Their weights aren't large enough to have every book ever written contained in them, or the details of every movie ever made, but unless you need that depth and breadth of knowledge, you'll be fine.
I just mean is the claim that the open source models where the closed models were 12 to 6 months ago true? They do seem to be for some specific tasks which is cool, but they seem even more uneven in skills than the frontier model. They're definitely useful tools, but I'm not sure if they're a match for frontier models from a year ago?
Frontier models from a year ago had issues with consistent tool calling, instruction following was pretty good but could still go off the rails from time to time.
Open weight models have those same issues. They are otherwise fine.
You can hook them up to a vector DB and build a RAG system. They can answer simple questions and converse back and forth. They have thinking modes that solve more complex problems.
They aren't going to discover new math theorems but they'll control a smart home and manage your calendar.
This isn't the first time they've dealt with scarcity, there's been supply chain scarcity four times since 2000. Post-dotcom boom, CDMA scarcity, HDD/flash scarcity, Pandemic scarcity.
The scarcity isn't long-term. Like all manufactured products, they'll ramp up production and flood the market with hardware, people will buy too much, market will drop. Boom and bust.
We're also still in the bubble. Eventually markets will no longer bear the lack of productivity/profit (as AI isn't really that useful) and there will be divestment and more hardware on the market as companies implode. Nobody is making 10x more from AI, they are just investing in it hoping for those profits which so far I don't think anyone has seen, other than in the companies selling the AI to other companies.
But more importantly, the models and inference keeps getting more efficient, so less hardware will do more in the future. We already have multiple models good enough for on-device small-scale work. In 5 years consumer chips and model inference will be so good you won't need a server for SOTA. When that happens, most of the billions invested in SOTA companies will disappear overnight, which'll leave a sizeable hole in the market.
> In 5 years consumer chips and model inference will be so good you won't need a server for SOTA.
Naw man, you crazy. If you tell me that in 5 years, consumer chips will be so good that I can run GPT-5.4-level AI on my phone, I'd find that plausible (I buy cheap phones). If you're telling me that in 5 years we won't need _servers_ because our _phones and/or desktops_ will be powerful enough to run the biggest newest LLMs in existence, I question your judgment, I think that prediction shows a deep uncreativity about how massively compute-hungry SOTA models will get.
The valuable things to do with inference will keep being a server niche because they'll keep being 1-2 OOM more compute-hungry than whatever consumer hardware can handle. Like gaming: my laptop can run games from 2015 at max settings no problem but the games actually worth getting excited about in 2026 still melt a $2k GPU, because whatever headroom the hardware gains, developers immediately spend on ray tracing and Nanite and modelling individual skin cells or whatever. I don't see any plausible reason to expect that the ceiling on "valuable server-side compute" or "inference capacity" will rise any more slowly than the on-device capability is rising.
My assumption is that in 2031, SOTA top-intelligence AI will be hosted on cloud servers like it is today, offering dirt-cheap access to capabilities we can't even dream of today, while your Android will be running some open-source GPT-5+ equivalent.
The thing is SOTA has a plateau. All LLMs work on the same principle: input goes in for training, reinforced by humans. There is only so much input (all recorded human knowledge), only so many human tweaks, that can produce only so much increased signal-to-noise in output. The machine can't read your mind, and there is no one truthful answer to most questions, so there will always be a limit on how accurate or correct or whatever any response will get. So at some point, you just can't make a better response. The agent harness, prompts, etc, are the only way to get better, and that's gonna be open source.
Add to that the algorithmic improvements on inference that's making inference faster with more context and higher quality. TurboQuant is just one example, more methods are coming out all the time. So the inference is getting more efficient.
At the same time, hardware can kind of keep getting infinitely better. Even if you can't make it smaller, you can make it more energy efficient, improve multitasking, more GPU cores/RAM or iGPUs, pack in more chips, improve cooling, use new materials... the sky's the limit.
Add all 3 together and at some point you will get Opus 4.7 on a phone with 40 t/s. At that point there's no way I'm paying for inference on a server. You can do RAG on-device, and image/video/voice is done by multi-modals. I want my agent chats replicated, but that's Google Drive. I want the agent to search the web, but that's Google Search. So eventually we're back to just doing what we do today (pre-AI) only with more automation.
The really advanced shit will come in 10 years, when we finally crack real memory and learning. That will absolutely be locked up in the cloud. But that's not an LLM, it's something else entirely. (slight caveat that WW3 will delay progress by 10-20 years)
This notion that "we don't have enough compute" does not cleanly reconcile with the fact that labs are burning cash faster than any cohort of companies in history.
If I am a grocery store that pays $1 for oranges and sells them for $0.50, I can't say, "I don't have enough oranges."
There is a major logic flaw in what you're saying.
'If I am a grocery store that pays $1 for oranges and sells them for $0.50, I can't say, "I don't have enough oranges."'
How about 'if I'm a grocery store and I see no limit on demand for oranges at $.50 but they are currently $1, I can say 'if oranges were cheaper I could sell orders of magnitude more of them'.
Buying oranges for $1 and selling for $0.5 is an investment into acquiring market share and customer relationships and a gamble on the price of oranges falling in the future.
Selling below cost is also called "predatory pricing". Sadly it's legal in US but it's something wealthy companies do to kill competitors and end up with captive customers.
The grocery store analogy works if compute is the orange.
But labs arent buying oranges — theyre buying the only orchard on the island, hoping it yields a fruit no ones grown yet. Burning $1B to net $500M isnt "I have too few oranges." Its "Im betting the farm Ill find a new one."
Both can be irrational. Theyre irrational in different ways.
"I built a ship to go to the Indies and bring back tea."
"Bro, the ship cost 100,000 pounds sterling and only brought back 50,000 pounds of tea. I don't care if you paid 12,500 pounds for the tea itself, you're losing money."
There is a very rational reason labs are spending everything they can get for more compute right now. The tea (inference) pays 60%+ margins. And that is rising. And that number is AFTER hyper scalars make their margins. There is an immense amount of profit floating around this system, and strategics at the edge believing they can build and control the demand through combined spend on training and inference in the proper ratios.
60%+ margins according to numbers which are not published publicly and have not AFAICT been audited.
Could they be accurate? Sure, I think people who claim this is impossible are overconfident. But I would encourage anyone who assumes they must be right to read a history of the Worldcom scandal. It's really quite easy for a person who wants to be making money (or an LLM who's been instructed to "run the accounts make no mistakes"!) to incorrectly categorize costs as capital investments when nobody's watching carefully.
Any materially false public statement by one of the foundation lab CEOs is a huge foot fault. I'm not saying they would never lie, but it would be a very, very dumb thing to do. That public information can be relied on by their private (very powerful) investors. I think if you're hearing these numbers ballparked in public settings, they are, as a prior, directionally accurate.
I agree, although I would emphasize that Worldcom is a great example of a CEO doing that very dumb thing. But I am not hearing these numbers ballparked in public settings. As far as I can tell, all the numbers people discuss for OpenAI or Anthropic margins come from anonymous leaks of internal documents.
ASML only makes a certain number of machines a year that can do extreme ultra-violet lithography.
Also - turbine blades limit power, according to Elon.
Between them - we cannot chip fabs past a certain rate, and we cannot stand up the datacenter to run these desired chips past a certain rate. Different people believe one or the other is the 'true' current bottleneck. The turbine supply chain scaling looks much more tractable -- EUV is essentially the most complicated production process humans have ever devised.
Is ASML really the bottleneck? Do you believe anybody but TSMC and few fabs could really use and acquire those machines? I don't know the throughput of a EUV device from ASML but I imagine you need :
- clean room, itself needing the infrastructure for it (size, airCo, filtering, electricity) and the staff to run and maintain that basically empty space
- wafers to "print" on, so that's a lot of water and logistic to manipulate them (so infrastructure for clean water and all chemicals) also with dedicated staff
- finally staff who would be able to design something significantly better than NVIDIA, Intel, Broadcom, IBM, etc while (and arguably that's the trickiest part IMHO) being able to get it good enough as at a scale that can be manufactured from their own fab.
so I'm wondering who can afford this kind of setup that can only then make use of ASML machines.
> (so infrastructure for clean water and all chemicals)
Fabs are some of the most complex chemical engineering sites (dealing with some of the most dangerous substances) in the world. So don't underestimate the complexity of this part.
Presumably ASML can increase production if demand is high enough the question is over what time frame. 5 years seems plausible to me but I honestly don't know what that number is.
Yes. And the fab companies and their suppliers are deliberately and wisely slow to scale up production to meet short term changes in demand. They've seen the history of the semiconductor industry, it's constant boom and bust cycles. But they have the highest op-ex costs of anyone. So when the party's over they are the ones who pay for it the most.
Yes. At least, the manufacturing of compute is. And a lot of the chain has been bitten hard by increasing capacity prematurely in the past so they're reticent to increase bandwidth at vast cost.
If only there were some form of cheap, widely manufactured power generation technology that didn't use turbines... Are they really going to wait until 2030 to get more turbines rather than invest in solar?
I am clueless in this field, but solar seems to be unreliable and yield fraction of power required. Do you have a suggestion on something to read and learn more?
Reliability: complete solar deployment includes some form of power storage. There are many variations, but chemical battery technology is improving the fastest, so it's gaining the most ground.
Amount of power: World solar power generation capacity is in the terawatts and rapidly increasing, there's no issue with its potential ceiling. As a bonus, it tends to work best on land that's useless for other purposes.
1. Supply can scale. You can point to COVID/supply-chain shocks, but the problem there is temporary changes. No one spins up a whole fab to address a 3 month spike. Whereas AI is not a temporary demand change.
2. Models are getting more efficient. DeepSeek V3 was 1/10th the cost of contemporary ChatGPT. Open weight models get more runnable or smarter every month. Cutting edge is always cutting edge, but if scarcity is real, model selection will adjust to fit it.
Initially I thought "Well... good for AI companies because they can then charge more" but IMHO that's a very tricky position because it means the cheap wave is behind us.
It's one thing to "sell" free or symbolically cheap stuff, it's another to have an actual client who will do the math and compare expenditure vs actually delivered value.
> and compare expenditure vs actually delivered value
Which means that the hype production will be driven up another few notches to make people doubt their rational findings and keep them in irrational territory just a tad longer. Every minute converts to dollars spent on tokens.
This isn’t really looking like AI scarcity it’s more like compute becoming the bottleneck : when the access depends on chips energy and capital it stops being a pure software game and the winners are often whoever can secure capacity first
Why is written with an assumption that we have finite hardware production capacity? Industrial processes can scale up, new factories can come online… it will take a while but the whole point of economics is that supply will scale to meet demand. The shortage is a temporary, point-in-time metric.
And that’s not considering the software innovation that can happen in the meantime.
The economic hypothesis that has dominated the past hundred years is that economic growth is infinite because resources are infinite and (almost) free. We all know this is unrealistic and disconnected from our human condition.
Regarding "innovation", I agree with your idea. I even think that the major innovation will be to transpose models locally, using reduced infrastructures that will still be sufficient for the majority of use cases.
... and I have this little idea in the back of my mind: when companies can no longer keep up with demand and people have (albeit more limited and reduced) local capacity, minds will start focusing on techniques (more humble and modest ones) to keep part of the system running locally, without dependency.
I know it may sound ridiculous, but it could actually become a way to break away from the business models that have been developed over the past few decades. Broadly speaking, this even amounts to saying that the biggest victims of AI could be the companies that bet on AI as a service.
Yet I know my vision is way too idealistic but I'm coming to imagine that a human brain, although less efficient in the long run, remains a reliable way to control the resulting costs and could even turn out to be more advantageous and more readily available than its silicon-based counterpart.
The human brain is incredibly efficient (Approximately 20W of energy consumption¹). These AI systems use many orders of magnitude more energy than human equivalents.
The problem with this idea is that someone can, and likely will, come up with the next best architecture that leapfrogs the current frontier models at least once a year, likely faster, for the foreseeable future. This means by the time you've manufactured your LLM on an ASIC, it's 4-5 generations behind, and probably much less efficient than current SOTA model at scale.
It won't make sense for ASIC LLMs to manifest until things start to plateau, otherwise it'll be cheaper to get smarter tokens on the cloud for almost all use cases.
That said, a 10 trillion parameter model on a bespoke compute platform overcomes a lot of efficiency and FOOM aspects of the market fit, so the angle is "when will models that can be run on an asic be good enough that people will still want them for various things even if the frontier models are 10x smarter and more efficient"
I think we're probably a decade of iteration on LLMs out, at least, and the entire market could pivot if the right breakthrough happens - some GPT-2 moment demonstrating some novel architecture that convinces the industry to make the move could happen any time now.
I don't think so. GB200 prices are GOING UP. A100s are still expensive. This implies massive utilization and demand, no? These machines are not sitting idle, or prices would drop in the very competitive hyperscaler environment.
Hard to say at this point. I'm sure you can run your LLM chips 24/7 for training and for the public to make weird thirst-trap videos about Judy Hopps but how real is the utilization and demand, really? Maybe very real, maybe not, I don't think we can know yet.
Its like being back in 1850 and you build the world's first amusement park where the rides are free or very cheap. People are like Amusement parks are the next big thing since Steam Boats! And tons of other rich people start to build huge amusement parks everywhere. The people who are skilled at making amusement park rides will increase their prices, and since the first amusement parks are free so they can get the public going to them demand will be huge.
But how sustainable is that? - well obviously we know from history that amusement parks did, in fact, take over the world and most people spent virtually all their time and money at amusement parks - I think the Crimean War was even fought over some religious-based theme park in Israel - until moving pictures came out, so it worked out for them, but for AI?
Companies who could see it clearly and ignored the "AI is a bubble duh" crowd will ultimately get benefited by the GPUs they already acquired. The companies who acted cautiously will get burned.
Companies who become dependent on AI to "optimize" their offering/processes are going to be faced with some serious vendor lock-in unless they do it in such a way that they can swap out the foundation model.
Trying to up-tier fractional improvements in something that can't be quantified easily, and !BONUS! with gated access it can't be as easily analyzed by the (low profit) AI analyzers/benchmarkers.
Foster paranoia among top executives that the fractional/debatable improvement is a MUST HAVE to STAY COMPETITIVE in your industry.
Meanwhile, I have not seen any improvement in software in the now almost ?three to four? years than mainstream LLM and AI coding assistance has arrived on the scene.
Although I will hold out the possibility that software has actually gotten far worse for the end user, because AI code is being dedicated to revenue enhancement and dark data collection.
We just had a realization during a demo call the other day:
The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up. Not being dependent on LLMs for your fundamental product’s value will be a major advantage, at least in pricing.
Yup. Also regardless of price they need to spend more and more as the project collapses under the inevitable incidental complexity of 30k lines of code a day.
It's similar to how if you know what you're doing you can manage a simple VPS and scale a lot more cost effectively than something like vercel.
In a saturated market margins are everything. You can't necessarily afford to be giving all your margins to anthropic and vercel.
I also can’t wait for the time when few know how to code. Just like how many folks don’t know html from css when the homebrew website went away.
Their might always be llms, but the dependence is an interesting topic.
Look no further to be honest; look at older generation programming languages like COBOL and how sought-after good developers for that language are.
But I'm also afraid / certain that LLMs are able to figure out legacy code (as long as enough fits in their context window), so it's tenuous at best.
Also, funny you mentioned HTML / CSS because for a while (...in the 90's / 2000's) it looked like nobody needed to actually learn those because of tools like Dreamweaver / Frontpage.
The issue with COBOL code is that it’s hidden. It’s mostly internal systems so little code available for training. HTML, TypeScript, JavaScript, C, etc, are readily available, billions of code lines.
Well, on the 2nd paragraph, I have no illusion they’ll figure out more as they are being trained. I am more thinking of the custodians (as coders turn into that)
Say you are a good coder now, but you are becoming a custodian, checking the llm work will slowly erode your skills. Maybe if you got a good memory or an amazing skillset it might be some time, but if you don’t use it, you lose it.
COBOL developers are sought after but still paid less than a grad doing crud. Is that the future?
The whole “you can make a lot of money programming in COBOL” is one of those myths that needs to die.
Even the briefest of Google searches show they make around the same as any other enterprise dev if not slightly less.
How are COBOL developers "sought after"? That's an oft-repeated but woefully incorrect meme.
FAANG new grads make more. If the COBOL devs had upskilled throughout their career they'd be Senior Staff/Principal+ and making 5-10x more than they do today.
A time when few know how to code?
I think that was about 10 years ago…
> The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up.
It's not that clear. Sure, hardware prices are going up due to the extremely tight supply, but AI models are also improving quickly to the point where a cheap mid-level model today does what the frontier model did a year ago. For the very largest models, I think the latter effect dominates quite easily.
>> The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up.
> It's not that clear. Sure, hardware prices are going up due to the extremely tight supply, but AI models are also improving quickly to the point where a cheap mid-level model today does what the frontier model did a year ago.
I agree; I got some coding value out of Qwen for $10/m (unlimited tokens); a nice harness (and some tight coding practices) lowers the distance between SOTA and 6mo second-tier models.
If I can get 80% of the way to Anthropic's or OpenAI's SOTA models using 10$/m with unlimited tokens, guess what I am going to do...
GitHub Copilot is already $10 and I don't even use up the requests every month, it's the most bang for buck LLM service I've used.
Until May
What’s happening in May?
Github Copilot switches all users from per prompt to per token billing
There's only so far engineers can optimise the underlying transformer technique, which is and always has been doing all the heavy lifting in the recent ai boom. It's going to take another genius to move this forward. We might see improvements here and there but the magnitudes of the data and vram requirements I don't think will change significantly
State space models are already being combined with transformers to form new hybrid models. The state-space part of the architecture is weaker in retrieving information from context (can't find a needle in the haystack as context gets longer, the details effectively get compressed away as everything has to fit in a fixed size) but computationally it's quite strong, O(N) not O(N^2).
I’ve read and heard from Semi Analysis and other best-in-class analysts that the amount of software optimizations possible up and down the stack is staggering…
How do you explain that capabilities being equal, the cost per token is going down dramatically?
Optimizations, like I said. They'll never hack away the massive memory requirements however, or the pre training... Imagine the memory requirements without the pre training step....this is just part and parcel of the transformer architecture.
And a lot of these improvements are really just classic automation or chaining together yet more transformer architectures, to fix issues the transformer architecture creates in the first place (hallucinations, limited context)
We are processing same data for the last 2 years.
Inference prices droped like 90 percent in that time (a combination of cheaper models, implicit caching, service levels, different providers and other optimizations).
Quality went up. Quantity of results went up. Speed went up.
Service level that we provide to our clients went up massively and justfied better deals. Headcount went down.
What's not to like?
The decline of independent thoughts for one. As people become reliant on LLMs to do their thinking for them and solve all problems that they stumble upon, they become a shell of their previous self.
Sadly, this is already happening.
We'll need to do faux mental work like how we do faux labor work.
There is no decline. Human assets were always too expensive to process some additional information. We are simply processing lot more of low signal data.
Actually some of our analysts are empowered by the tools at their disposal. Their jobs are safe and necessary. Others were let go.
Clients are happy to get fuller picture of their universe, which drives more informed decissions . Everybody wins.
You are free to believe what you want, but what you describe does not match what I’ve seen from society as a whole. I’m just going to leave this here: https://www.media.mit.edu/projects/your-brain-on-chatgpt/ove...
Are you being satirical?
The headcount that went down probably isn’t too thrilled about it.
Yes, probably. But the others gained skills and tools that made their jobs secure.
Right but the question wasn’t were some people better off. It is what’s not to like?
You also have to look at how exposed your vendors are to cost increases as well.
Your company may have the resources to effectively shift to cheaper models without service degradation, but your AI tooling vendors might not. If you pay for 5 different AI-driven tools, that's 5 different ways your upstream costs may increase that you'll need to pass on to customers as well.
> Not being dependent on LLMs for your fundamental product’s value
I think more specifically not being dependent on someone else's LLM hardware. IMO having OSS models on dedicated hardware could still be plenty viable for many businesses, granted it'll be some time before future OSS reaches today's SOTA models in performance.
What's weird though is the bifurcation in pricing in the market: aka if your app can function on a non-frontier level AI you can use last years model at a fraction of the cost.
That'll be (part of) the big market correction, but also speaking broadly; as investor money dries up and said investors want to see results, many new businesses or products will realise they're not financially viable.
On a small scale that's a tragedy, but there's plenty of analysts that predict an economic crash and recession because there's trillions invested in this technology.
No shit. People are just figuring this out now?
This is the “Building my entire livelihood on Facebook, oh no what?” all over again.
Oh no sorry I forgot, your laptops LLM can draw a potato, let me invest in you.
Indeed, it was clear from the beginning, "AI" companies want to become infrastructure and a critical dependency for businesses, so they can capture the market and charge whatever they want. They will have all the capital and data needed to eventually swallow those businesses too, or more likely sell it to anyone who wants the competitive advantage.
Seriously.
> We just had a realization during a demo call the other day
These tools have been around for years now. As they've improved, dependency on them has grown. How is any organization only just realizing this?
That's like only noticing the rising water level once it starts flooding the second floor of the house.
https://news.ycombinator.com/item?id=47804804
in fact I am betting opposite. frontier models are getting not THAT much better anymore at all, for common business needs at least. but the OSS models keep closing the gap. which means if trajectories hold there will be a near future moment probably where the big provider costs suddenly drop shaerply once the first viable local models consistently can take over tasks normally on reasonable hardware. Right now probably frontier providers rush for as much money as they possible can before LLMs become a true commodity for the 80% usecases outside of deep expert areas they will have an edge over as specialist juggernauts (iE a cybersecurity premium model).
So its all a house of cards now, and the moment the bubble bursts is when local open inference has closed the gap. looks like chinese and smaller players already go hard into this direction.
Local open inference can address hardware scarcity by repurposing the existing hardware that users need anyway for their other purposes. But since that hardware is a lot weaker than a proper datacenter setup, it will mostly be useful for running non-time-critical inference as a batch task.
Many users will also seek to go local as insurance against rug pulls from the proprietary models side (We're not quite sure if the third-party inference market will grow enough to provide robust competition), but ultimately if you want to make good utilization of your hardware as a single user you'll also be pushed towards mostly running long batch tasks, not realtime chat (except tiny models) or human-assisted coding.
Absolutely. Pricing exposure is the quiet story under all the waves of AI hype. Build for convenience → subsidise for dependence → meter for margin is a well-worn playbook, and AI-dependent companies are about to find out what phase three feels like.
Hyperscalers are spending a fortune so we think AI = API, but renting intelligence is a business model, not a technical inevitability.
Shameless link to my post on this: https://mjeggleton.com/blog/AIs-mainframe-moment
How is that surprising? We've been taking that into account for any LLM related tooling for over a year now that we either can drop it, or have it designed in a way that we can switch to a selfhosted model when throwing money at hardware would pay for itself quickly.
It's just another instance of cloud dependency, and people should've learned something from that over the last two decades.
Not so much that it was surprising, rather that we looked at a competitor’s site and noticed that a) their prices went way up and b) their branding changed to be heavily AI-first.
So we thought, hmm, “wonder if they are increasing prices to deal with AI costs,” and then projected that into a future where costs go up.
We don’t have this dependence ourselves, so this seems to be a competitive advantage for us on pricing.
Marginal costs matter in this world.
The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up
Or they'll price the true cost in from the start, and make massive profits until the VC subsidies end... I know which one I'd do.
We don't know what anthropic's true costs are. So pricing that in is at best a guess.
I wonder if it could be that they won't because the real mechanism is that AI wrapper pricing power is weak (switching costs near zero) but state of the art models makes it difficult to lower prices due to higher cost.
Also: AI dependance could be explicit AI API usage by the product itself, but also anything else, like: AI assisted coding, AI used by humans in other surrounding workflows, etc.
Yeah that's actually what I initially meant: just dependence on AI as a technology, not purely the usage costs. I didn't spell that out well enough.
This was as clear as the sky when the first llm based businesses popped up. How did you realize this only now?
Replied here: https://news.ycombinator.com/item?id=47804804
And I don't really mean new businesses that are entirely built around LLMs, rather existing ones that pivoted to be LLM-dependent – yet still have non-LLM-dependent competitors.
Yea that would've been extremely short sighted from your competitors. Thanks for linking the response!
same as Uber… in the beginning everyone pretty much new that the cost of rides cannot possibly be that cheap and that it is subsudized. once you corner the market etc people just got used to “real” prices to the poibt that now there are often cheaper alternatives than Uber but people still Uber…
Its also quite interesting to read about Uber exploits their drivers and discriminating algorithms. Cory Doctorow mentioned it in his latest book, sadly cant link the direct sources.
Not really, the next move is to establish standards groups requiring the use of AI in product development. A mix of industry and governmental mandates. What you view are viewing as COGS instead becomes instead a barrier to entry.
Constraints can lead to innovation. Just two things that I think will get dramatically better now that companies have incentive to focus on them:
* harness design
* small models (both local and not)
I think there is tremendous low hanging fruit in both areas still.
China already operates like this. Low cost specialized models are the name of the game. Cheaper to train, easy to deploy.
The US has a problem of too much money leading to wasteful spending.
If we go back to the 80s/90s, remember OS/2 vs Windows. OS/2 had more resources, more money behind it, more developers, and they built a bigger system that took more resources to run.
Mac vs Lisa. Mac team had constraints, Lisa team didn't.
Unlimited budgets are dangerous.
Though I do agree with you, I just came back from a trip to China (Shanghai more specifically) and while attending a couple AI events, the overwhelming majority of people there were using VPNs to access Claude code and codex :-/
Parent's point was about deployment, not agentic coding.
On the Mac vs Lisa team, I generally agree but wasn't there a strong tension on budget vs revenue on Mac vs Apple II? And that Apple II had even more constrained budget per machine sold which led to the conflict between Mac and Apple II teams. (Apple II team: "We bring in all the revenue+profit, we offer color monitors, we serve businesses and schools at scale. Meanwhile, Steve's Mac pirate ship is a money pit that also mocks us as the boring Navy establishment when we are all one company!")
By the logic of constraints (on a unit basis), Apple II should have continued to dominate Mac sales through the early 90s but the opposite happened.
Perhaps its because american hyperscalers want unlimited upside for their capital?
> Low cost specialized models
Can you elaborate on this? Is this something that companies would train themselves?
You can fine-tune a model, but there are also smaller models fine-tuned for specific work like structured output and tool calling. You can build automated workflows that are largely deterministic and only slot in these models where you specifically need an LLM to do a bit of inference. If frontier models are a sledgehammer, this approach is the scalpel.
A common example would be that people are moving tasks from their OpenClaw setup off of expensive Anthropic APIs onto cheaper models for simple tasks like tagging emails, summarizing articles, etc.
Combined with memory systems, internal APIs, or just good documentation, a lot of tasks don't actually require much compute.
It has been a very bad bet that hardware will not evolve to exceed the performance requirements of today's software tomorrow, just as it is a bad bet that tomorrow someone will rewrite today's software to be slower.
Eh, but then as hardware evolves, the software will also follow suit. We’ve had an explosion of compute performance and yet software is crawling for the same tasks we did a decade ago.
Better hardware ensures that software that is “finished” today will run at acceptable levels of performance in the future, and nothing more.
I think we won’t see software performance improve until real constraints are put on the teams writing it and leaders who prioritize performance as a North Star for their product roadmap. Good luck selling that to VCs though.
Harness is a big one, Claude Code still has trouble editing files with tabs. I wonder how many tokens per day are wasted on Claude attempting multiple times to edit a file.
The future is now, I guess
Yep.
As a recent example in AI space itself. China had scarce GPU resources, quite obvious why => DeepSeek training team had to invent some wheels and jump through some hoops => some of those methods have since become 'industry standard' and adopted by western labs who are now jumping through the same hoops despite enjoying massive computeresources, for the sake of added efficiency.
Absolutely. Anyone working on inference token level knows how wasteful it all is especially in multimodal tokens.
Could not agree more, this will spur innovation in all aspects of local models is my hunch.
What do you mean by harness here?
When you go to the command line and type “Claude”, there is an LLM, and everything else is the harness
I'm having an hard time getting my mind to see this.
> Users should re-tune their prompts and harnesses accordingly.
I read this in the press release and my mind thought it meant test harness. Then there was a blog post about long running harnesses with a section about testing which lead me to a little more confusion.
Yes, the word 'harness' is consistently used in the context as a wrapper around the LLM model not as 'test harness'.
This field is chock full of people using terms incorrectly, defining new words for things that already had well known names, overloading terms already in use. E.g. shard vs partition. TUI which already meant "telephony user interface ". "Client" to mean "server" in blockchain.
Some people also call evaluations "tests". There are unexpected things that come along with new models, like the model in a workflow you'd set up suddenly starts calling a tool and never stops or decides to no longer call a particular tool, so running your existing evaluations to catch regressions like this and potentially updating the prompts is considered "testing" your prompts and harnesses.
I understood this concept with this simple equation: Agent = LLM + harness
It’s the tool that calls the model, give it access to the local file system, calls the actual tools and commands for the model, etc, and provide the initial system prompt.
Basically a clever wrapper around the Anthropic / OpenAI / whatever provider api or local inference calls.
pi vs. claude code vs. codex These are all agent harnesses which run a model (in pi's case, any model) with a system prompt and their own default set of tools.
[dead]
There's other side to it too.
Whoever running and selling their own models with inference is invested into the last dime available in the market.
Those valuations are already ridiculously high be it Anthropic or OpenAI to the tune of couple of trillion dollars easily if combind.
All that investment is seeking return. Correct me if I'm wrong.
Developers and software companies are the only serious users because they (mostly) review output of these models out of both culture and necessity.
Anywhere else? Other fields? There these models aren't any useful or as useful while revenue from software companies by no means going to bring returns to the trillion dollar valuations. Correct me if I'm wrong.
To make the matter worst, there's a hole in the bucket in form of open weight models. When squeezed further, software companies would either deploy open weight models or would resort to writing code by hand because that's a very skilled and hardworking tribe they've been doing this all their lives, whole careers are built on that. Correct me if I'm wrong.
Eventually - ROI might not be what VCs expect and constant losses might lead to bankruptcies and all that build out of data centers all of sudden would be looking for someone to rent that compute capacity result of which would be dime a dozen open weight model providers with generous usage tiers to capitalize on that available compute capacity owners of which have gone bankrupt and can't use it any more wanting to liquidate it as much as possible to recoup as much investment as possible.
EDIT: Typos
It feels like a repeat of the dot com infrastructure buildup that spurred the whole 2005 explosion in affordable hosting and new companies. This will probably leave us massive access to affordable compute in a couple of years.
For power maybe. But the expected lifetime of GPU hardware is 3 years before they fail completely
OpenAI has an absurdly high valuation given their cash burn vs RRR.
Anthropic's is far more reasonable.
It makes no sense to lump these two companies together when talking about valuation. They have completely different financial dynamics
No matter how low and reasonably Anthropic is valued, don't think $200 Max plans are going to recoup the investment + some return on top because size of the software industry is not that huge and profit margins for AI inference aren't very high either.
> because size of the software industry is not that huge
I onboarded marketing on a premium team Claude seat yesterday. And one of our sales vibecoded an internal tool in the last three weeks using Claude Code that they now use every day. I wouldn’t have imagined it a month ago. We still had to take care of deployment for him, but things are moving fast.
Pro and Max plans are probably a drop in the bucket for them.
Seems like everybody an their mothers are using max plans these days. I wouldn't be surprised if LTV of each customer was big enough to justify spending.
Assuming there are 10 million developers and everyone is at $200 max plan, that would be $2 billion/month or $24 billion/year maximum.
Note - this is just the revenue not the profit. No salaries, no compute paid for. Just plain revenue. Profit would be way less.
But even that - if we take it to $24 billion/year and we take a 10x multiple, the company is barely valued at $240 billon dollar, lets be generous and make it double at $480 billion and then round it up to $500 billion for a nice round number.
Far far from the $800 billion valuation Anthropic is looking at.
Only a matter of time.
EDIT: Fixed math
While I agree with you that AI companies are overvalued, I think 10 million developers at $200 per month makes 2 billion.
Thanks for pointing out. I updated the comment.
Companies are spending far more than $200/month/developer. The $200 Max plan is a great value but you hit limits far too soon, and it also doesn't cover any of the other styles of integrations and tools that you can build and use to help your developers, like code review suggestions, which at the very least would come from additional Max plans, and not from the individual developers' plans.
Pro and Max plans are a tiny fraction of their revenue. Many businesses are spending thousands of dollars per head per month.
Really? Not challenging you, genuinely asking for more details.
If that is true, I think AI is counterproductive from the bean counter's standpoint.
Not counterproductive because people aren't just sitting back in the rest of the time while AI does work. They do more work. $3k per head on Claude is nothing if your devs get 2x more work done.
Do they review the code? Because in my experience using Claude Opus 4.6 generates code that would be buggy and the tests would be written agains that buggy code with wrong assumptions that certainly would pass with flying colors.
It is only when you look closed you get to know what the hell has happened!
And nor are they trying to. If you are spending $200 a month, you are a mere tolerated nuisance. The company I work for gives every developer a $5000 a month allowance to Claude and I think there are around 500-600 people eligible for it.
comments like these is probably while significant contingent of HN talks heavily about AI bubble. burning that kind of cash always ends up like the right move, godspeed :)
I work for a consulting company, we know exactly how much each billable person makes the company and whether the ROI is worth it.
We don’t have to measure “productivity gains”.
I personally don’t come close to that and neither do I suspect most of us. Between using my $20 a month ChatGPT subscription with Codex and the amount of time I spend on Zoom calls “adding on to what Becky said” and “looking at things from the 1000 foot view”
On another note, the grunt work I use to delegate to a junior consultant, I now can get Claude to do in a fraction of the the time. They were making a lot more than the worse case of $72K a year fully allocated. But honestly most of that work is done with my $240 a year ChatGPT subscription + maybe $600 in Claude at the current prices/limits
> would resort to writing code by hand because that's a very skilled and hardworking tribe they've been doing this all their lives
Shush, don't tell that to the AI coding acolytes.
There is a lot of demand still coming for sure but I think I'm more optimistic. Ready to eat my hat on this but
- higher prices will result in huge demand destruction too. Currently we're burning a lot of tokens just because they're cheap, but a lot of heavy users are going to spend the time moving flows over to Haiku or onprem micro models the moment pricing becomes a topic.
- data centers do not take that long to build, probably there are bottlenecks in weird places like transformers that will cause some hicups, but nvidia's new stuff is waay more efficient and the overall pipeline of stuff coming online is massive.
- probably we will see some more optimization at the harness level still for better caching, better mix of smaller models for some use, etc etc.
These companies have so much money and they at least anthropic and openai are playing winner takes it all stakes, with competition from the smaller players too. I think they're going to be feeding us for free to win favour for quite a while still.
Let's see though.
I agree and I am amazed at how much money some individuals and also a friend's company burn on token costs. I get huge benefits from this tech just using gemini-cli and Antigravity a few times a week, briefly. I also currently invest about $15/month in GLM-5.1 running Hermes Agent on a small dedicated VPS - fantastically good value for getting stuff done and this requires little of my time besides planning what I need done.
I think the token burners are doing it wrong. I think that long term it is better to move a little slower, do most analysis and thinking myself, and just use AI when the benefits are large while taking little of my time and money to use the tools.
This is probably even the "fun" part of the whole picture. The purely dystopia starts when investment firms just silently grow bigger and bigger data centers like cancer. There will be no press releases, no papers, no chance anyone without billions will even know the details yet alone get access. One day we realise the worlds resources (maybe not as in the paperclip maximiser, but as in memory, energy, GPUs, water, locations) are consumed by trading models and the data centres are already guarded by robot armies. While we were distracted frighting with anthropic and openAI the real war was already over. Mythos is one sign in this direction but i also met a few people who were claiming to fund fairly large research and training operations just by internal models working on financial markets. I have no way to verify those claims but this happened 3 times now and the papers/research they were working on looked pretty solid and did not seem like they were running kimi openclaw on polymarket but actual models on some significant funds. Would be really interested if anyone here has some details on this reality. I would also not be surprised if this is a thing that people in SF just claim to sound dangerous and powerful.
The fun times will also end for consumers soon enough, when the oligopolies are established and investors start asking for returns.
one graph, One graph and the author is pinning an entire theory on it?
Infra is always limited, even at hyper scalers. This leads to a bunch of tools dfofr caching, profiling and generally getting performance up, not to mention binpacking and all sorts of other "obvious" things.
On X I had seen him mostly posting memes so this post seems par for the course
> Infra is always limited, even at hyper scalers
I think maybe infra is limited only at hyperscalers. For the rest of us it's just how much capacity to we want to rent from the hyperscalars.
It's kind of a recent cloud-native mindset, since back in the day when you ran your own hardware scaling and capacity was always top of mind. Looks like AI compute might be like that again, for the time being.
One graph, about 100 words, AI in the title: Hacker News front page.
Not bad for a coffee break of effort.
Well it's in the books. O(n^2) algorithms are bad in the long run, transformers algorithm has such complexity, so not a big surprise we hit the limits.
It’s time to be AI-independent. It’s like AWS, for most of us, it’s not worth it.
The US is bound by energy and China is bound by compute power. The one who solves its limitation first will end this “Scarcity Era”.
China is installing something like 500 GW of wind and solar per year now. Even if they're only able to build and otherwise access chips that have half the SoTA performance per watt, they will win.
Performance per dollar may be more important than performance per watt here, though
A dollar is an entirely fictional unit and trillions of it can be manufactured at no cost, while watts are constrained by the laws of physics, photons/electrons, supply chain of electricity and all that fun stuff in the real world.
A dollar is still a useful unit as "the fraction of the economy that can be controlled by currency". It's true that printing a huge pile of it and throwing it at GPUs wouldn't instantly convert into more GPUs, but it would meaningfully represent that other things are being squeezed out to allocate more resources to GPU production even so. That such reallocation is inefficient, arguably immoral, and highly questionable in the long term versus other options wouldn't stop that from being ture.
>dollar is an entirely fictional unit and trillions of it can be manufactured at no cost
If the abstraction works better for you this way, call them interchangeable units of American and Chinese insolvency. Or incremental forfeiture of domestic ownership.
> A dollar is an entirely fictional unit and trillions of it can be manufactured at no cost
It’s still a useful proxy for resources allocation and viability.
..unless you're actually reasoning at nation-scale where OP's points apply
I wouldn’t agree. Even at national scale, these projects cost resources. And the resources of all agents (org, countries) are constrained.
While we could reason in "performance / watt" and "performance / people", "performance / whatever other resource involved", and "performance / opportunity cost of allocating these resources to this use case and not another", "performance / whatever unit of stable-ish currency" is a convenient and often "good enough" approximation that somewhat encapsulates them all.
A simplification, like any model, but still useful.
Win what exactly?
China's domestic chips are increasingly close to state-of-the-art. The US electrical grid is... not.
US energy is constrained by the utility monopolies/oligopolies which have to extract more rents, specifically by increasing costs. Their profit is a percentage of cost, these perverse incentives + oligopolies will make it increasingly expensive to make anything (including AI) in US.
Or simply by the fact that increasing production takes time? Any power plant takes years to build?
Years, is like a lifetime for AI at this point...
> increasing production takes time?
This is true of nearly everything (except money). I'm not sure of the point you are trying to make.
Not solar. China and to a lesser extent India are pumping out huge solar farms in months.
The dynamics vastly favor China, part of the reason the US sprinting towards "ASI" isn't totally boneheaded is that the US and its industry needs a hail mary play to "win" the game, if they play it safe they lose for sure.
I'd be fine with a world without AI, honestly. Nobody really wins this race except the very wealthy. And I don't think it's really going to play out the way the wealthy think it will. It's more like a dog catching a car than it is a race.
> It's more like a dog catching a car than it is a race.
What does this mean? I didn't understand the analogy.
A car caught by a dog has no purpose. The activity concludes with no output.
"The dog that caught the car" refers to how dogs sometimes chase cars. Suppose the car stops and the dog catches up - what is it going to do? It has no plan, it has no purpose, it isn't going to bite the car, it isn't going to get anything out of catching the car. The car may even run it over. I intended it basically as "play stupid games, win stupid prizes", or "be careful what you wish for".
My observation is that the dog sniffs all the tires, picks one tire, lifts one leg and does the deed. I don't know if its a way of marking territory or domination. We need a dogatologist to explain what it means.
That was quite the unexpected anticlimactic ending. I’m sure Terry Pratchett would be proud.
We did it reddit!
It seems very possible that we have at least five years of real limitations on compute coming up. Maybe ten, depending on ASML. I wonder what an overshoot looks like. I also wonder if there might be room for new entrants in a compute-scarce environment.
For instance, at some point, could Coreweave field a frontier team as it holds back 10% of its allocations over time? Pretty unusual situation.
Jensen just said that if the signal/commitments are there, ASML can scale in 2-3 years.
With Anthropic buying compute in dark alleys I’d assume that day is coming..
AKA, the beginning of big companies being able to roll over small companies with moar money
(note: I don't expect this to actually happen until the AI gets good enough to either nearly entirely replace humans or solve cooperation, but the long term trend of scarce AI will go towards that direction)
The scarcity framing assumes compute is the bottleneck. For most production deployment's Ive seen, the actual bottleneck is evaluation and knowing what to trust.
You can throw cheaper models at a problem all day but, if you can't measure where the model fails on your data, You're just making mistakes faster at a lower cost.
Compute gets cheaper. Reliable evaluation doesn't.
That’s a lot of speculation based on one graph. It doesn’t cover whatever Google is doing, for example.
To bang on the same damn drum:
Open Weight models are 6 months to a year behind SOTA. If you were building a company a year ago based on what AI could do then, you can build a company today with models that run locally on a user's computer. Yes that may mean requiring your customers to buy Macbooks or desktops with Nvidia GPUs, but if your product actually improves productivity by any reasonable amount, that purchase cost is quickly made up for.
I'll argue that for anything short of full computer control or writing code, the latest Qwen model will do fine. Heck you can get a customer service voice chat bot running in 8GB of VRAM + a couple gigs more for the ASR and TTS engine, and it'll be more powerful than the hundreds of millions spent on chat bots that were powered by GPT 4.x.
This is like arguing the age of personal computing was over because there weren't enough mainframes for people to telnet into.
It misses the point. Yes deployment and management of personal PCs was a lot harder than dumb terminal + mainframe, but the future was obvious.
The mainframe/PC analogy is spot on. And the hardware floor keeps dropping - you can grab a mini PC with 32-64GB RAM for a few hundred bucks and run surprisingly capable quantized models locally. Something like https://terminalbytes.com/best-mini-pcs-for-home-lab-2025/ shows the kind of hardware that's now available at consumer prices. The "scarcity" framing only makes sense if you assume everyone needs frontier-tier models for everything.
I've seen this claimed, but I'm not sure it's been true for my use cases? I should try a more involved analysis but so far open models seem much less even in their skills. I think this makes sense if a lot of them are built based on distillations of larger models. It seems likely that with task specific fine tuning this is true?
> I've seen this claimed, but I'm not sure it's been true for my use cases?
I'd be surprised if it isn't true for your use cases. If you give GLM-5.1 and Optus 4.6 the same coding task, they will both produce code that passes all the tests. In both cases the code will be crap, as no model I've seen produces good code. GLM-5.1 is actually slightly better at following instructions exactly than Optus 4.6 (but maybe not 4.7 - as that's an area they addressed).
I've asked GLM-5.1 and Opus 4.6 to find a bug caused by a subtle race condition (the race condition leads to a number being 15172580 instead of 15172579 after about 3 months of CPU time). Both found it, in a similar amount of time. Several senior engineers had stared at the code for literally days and didn't find it.
There is no doubt the models do vary in performance at various tasks, but we are talking the difference between Ferrari vs Mercedes in F1. While the differences are undeniable, this isn't the F1. Things take a year to change there. The performance of the models from Anthropic and OpenAI literally change day by day, often not due to the model itself but because of the horsepower those companies choose to give them on the day, or them tweaking their own system prompts. You can find no end of posts here from people screaming in frustration the thing that worked yesterday doesn't work today, or suddenly they find themselves running out of tokens, or their favoured tool is blocked. It's not at all obvious the differences between the open-source models and the proprietary ones are worse than those day to day ones the proprietary companies inflict on us.
> In both cases the code will be crap, as no model I've seen produces good code.
I'm wondering if you have actually used claude code because results are not so catastrophic as you describe them.
I used LLMs to write what seems like far too many lines of code now. This is an example Opus 4.6 running at maximum wrote in C:
If you don't know C, in older versions that can be a catastrophic failure. (The issue is so serious in modern C `free(NULL)` is a no-op.) If it's difficult to get a `FOO == NULL` without extensive mocking (this is often the case) most programmers won't do it, so it won't be caught by unit tests. The LLMs almost never get unit test coverage up high enough to catch issues like this without heavy prompting.But that's the least of it. The models (all of them) are absolutely hopeless at DRY'ing out the code, and when they do turn it into spaghetti because they seem almost oblivious to isolation boundaries, even when they are spelt out to them.
None of this is a problem if you are vibe coding, but you can only do that when you're targeting a pretty low quality level. That's entirely appropriate in some cases of course, but when it isn't you need heavy reviews from skilled programmers. No senior engineer is going to stomach the repeated stretches of almost the "same but not quite" code they churn out.
You don't have to take my word for it. Try asking Google "do llm's produce verbose code".
C is fundamentally a bad target for LLMs. Humans get C wrong all the time, so we can not hope the nascent LLM, which has been trained on 95% code that does automatic memory management, to excel here.
I always found myself writing verbose copypasta code first, then compress it down based on the emerging commonalities. I think doing it the other way around is likely to lead to a worse design. Can you not tell the LLM to do the same? Honest question.
> I always found myself writing verbose copypasta code first, then compress it down based on the emerging commonalities. I think doing it the other way around is likely to lead to a worse design.
I do pretty much the same thing, which is to say I "write code using a brain dump", "look for commonalities that tickle the neurons", then "refactor". Lather, rinse, and repeat until I'm happy.
> Can you not tell the LLM to do the same?
You can tell them until you're blue in the face. They ignore you.
I'm sure this is a temporary phase. Once they solve the problem, coding will suffer the same fate as blacksmiths making nails. [0] To solve it they need to satisfy two conflicting goals - DRY the code out, while keeping interconnections between modules to a minimum. That isn't easy. In fact it's so hard people who do it well and can do it across scales are called senior software engineers. Once models master that trick, they won't be needed any more.
By "they" I mean "me".
[0] Blacksmiths could produce 1,000 or so a day, but it must have been a mind-numbing day even if it paid the bills. Then automation came along, and produced them at over a nail per second.
> C is fundamentally a bad target for LLMs.
I found it exceptionally good, because:
a) The agent doesn't need to read the implementation of anything - you can stuff the entire projects headers into the context and the LLM can have a better birds-eye view of what is there and what is not, and what goes where, etc.
and
b) Enforcing Parse, don't Validate using opaque types - the LLM writing a function that uses a user-defined composite datatype has no knowledge of the implementation, because it read only headers.
Is foo a pointer in your example? Is free(NULL) not a valid operation?
Yes `foo` is a pointer.
`free(NULL)` is harmless in C89 onwards. As I said, programmers freeing NULL caused so many issues they changed the API. It doesn't help that `malloc(0)` returns NULL on some platforms.
If you are writing code for an embedded platform with some random C compiler, all bets on what `free(NULL)` does are off. That means a cautious C programmer who doesn't know who will be using their code never allows NULL to be passed to `free()`.
In general, most good C programmers are good because they suffer a sort of PTSD from the injuries the language has inflicted on them in the past. If they aren't avoiding passing NULL to `free()`, they haven't suffered long enough to be good.
> That means a cautious C programmer who doesn't know who will be using their code never allows NULL to be passed to `free()`.
If your compiler chokes on `free(NULL)` you have bigger problems that no LLM (or human) can solve for you: you are using a compiler that was last maintained in the 80s!
If your C compiler doesn't adhere to the very first C standard published, the problem is not the quality of the code that is written.
> If they aren't avoiding passing NULL to `free()`, they haven't suffered long enough to be good.
I dunno; I've "suffered" since the mid-90s, and I will free NULL, because it is legal in the standard, and because I have not come across a compiler that does the wrong thing on `free(NULL)`.
So what would be the best practice in a situation like that? I would (naively?) imagine that a null pointer would mostly result from a malloc() or some other parts of the program failing, in which case would you not expect to see errors elsewhere?
> imagine that a null pointer would mostly result from a malloc() or some other parts of the program failing, in which case would you not expect to see errors elsewhere?
Oh yes, you probably will see errors elsewhere. If you are lucky it will happen immediately. But often enough millions of executed instructions later, in some unrelated routine that had its memory smashed. It's not "fun" figuring out what happened. It could be nothing - bit flips are a thing, and once you get the error rate low enough the frequency of bit flips and bugs starts to converge. You could waste days of your time chasing an alpha particle.
I saw the author of curl post some of this code here a while back. I immediately recognised the symptoms. Things like:
Every 2nd line was code like that. If you are wondering, he wrote `(NULL == foo)` in case he dropped an `=`, so it became `(NULL = foo)`. The second version is a syntax error, whereas `(foo = NULL)` is a runtime disaster. Most of it was unjustified, but he could not help himself. After years of dealing with C, he wrote code defensively - even if it wasn't needed. C is so fast and the compilers so good the coding style imposes little overhead.Rust is popular because it gives you a similar result to C, but you don't need to have been beaten by 10 years of pain in order to produce safe Rust code. Sadly, it has other issues. Despite them, it's still the best C we have right now.
Having used Claude Code extensively, catastrophic is a perfect word to describe it.
What are you trying to do?
Write code? No. Use frontier models. They are subsidized and amazing and they get noticably better ever few months.
Literally anything else? Smaller models are fine. Classifiers, sentiment analysis, editing blog posts, tool calling, whatever. They go can through documents and extract information, summarize, etc. When making a voice chat system awhile back I used a cheap open weight model and just asked it "is the user done speaking yet" by passing transcripts of what had been spoken so far, and this was 2 years ago and a crappy cheap low weight model. Be creative.
I wouldn't trust them to do math, but you can tool call out to a calculator for that.
They are perfectly fine at holding conversations. Their weights aren't large enough to have every book ever written contained in them, or the details of every movie ever made, but unless you need that depth and breadth of knowledge, you'll be fine.
I just mean is the claim that the open source models where the closed models were 12 to 6 months ago true? They do seem to be for some specific tasks which is cool, but they seem even more uneven in skills than the frontier model. They're definitely useful tools, but I'm not sure if they're a match for frontier models from a year ago?
Frontier models from a year ago had issues with consistent tool calling, instruction following was pretty good but could still go off the rails from time to time.
Open weight models have those same issues. They are otherwise fine.
You can hook them up to a vector DB and build a RAG system. They can answer simple questions and converse back and forth. They have thinking modes that solve more complex problems.
They aren't going to discover new math theorems but they'll control a smart home and manage your calendar.
Buy new Macs from where? There is a shortage of RAM, SSD, GPUs, and the CPU shortage just started.
That’s nonsense. Local models don’t have any of the nuance in text responses. I find them more akin to GPT 3.5 than even 4.x
Energy scarcity will drive more innovation in local silicon and local inference. Apple will be the unexpected beneficiary of this reality.
>For the first time since the 2000s, technology companies are confronting the limits of their supply chain.
I thought there'd been a shortage of cheap GPUs since ChatGPT took off and also before that in various crypto booms. I'm not sure it's a new thing.
But that concerned mostly only gamers and cryptominers. AI is supposed to be replacing traditional software development, which affects everything.
This isn't the first time they've dealt with scarcity, there's been supply chain scarcity four times since 2000. Post-dotcom boom, CDMA scarcity, HDD/flash scarcity, Pandemic scarcity.
The scarcity isn't long-term. Like all manufactured products, they'll ramp up production and flood the market with hardware, people will buy too much, market will drop. Boom and bust.
We're also still in the bubble. Eventually markets will no longer bear the lack of productivity/profit (as AI isn't really that useful) and there will be divestment and more hardware on the market as companies implode. Nobody is making 10x more from AI, they are just investing in it hoping for those profits which so far I don't think anyone has seen, other than in the companies selling the AI to other companies.
But more importantly, the models and inference keeps getting more efficient, so less hardware will do more in the future. We already have multiple models good enough for on-device small-scale work. In 5 years consumer chips and model inference will be so good you won't need a server for SOTA. When that happens, most of the billions invested in SOTA companies will disappear overnight, which'll leave a sizeable hole in the market.
> In 5 years consumer chips and model inference will be so good you won't need a server for SOTA.
Naw man, you crazy. If you tell me that in 5 years, consumer chips will be so good that I can run GPT-5.4-level AI on my phone, I'd find that plausible (I buy cheap phones). If you're telling me that in 5 years we won't need _servers_ because our _phones and/or desktops_ will be powerful enough to run the biggest newest LLMs in existence, I question your judgment, I think that prediction shows a deep uncreativity about how massively compute-hungry SOTA models will get.
The valuable things to do with inference will keep being a server niche because they'll keep being 1-2 OOM more compute-hungry than whatever consumer hardware can handle. Like gaming: my laptop can run games from 2015 at max settings no problem but the games actually worth getting excited about in 2026 still melt a $2k GPU, because whatever headroom the hardware gains, developers immediately spend on ray tracing and Nanite and modelling individual skin cells or whatever. I don't see any plausible reason to expect that the ceiling on "valuable server-side compute" or "inference capacity" will rise any more slowly than the on-device capability is rising.
My assumption is that in 2031, SOTA top-intelligence AI will be hosted on cloud servers like it is today, offering dirt-cheap access to capabilities we can't even dream of today, while your Android will be running some open-source GPT-5+ equivalent.
The thing is SOTA has a plateau. All LLMs work on the same principle: input goes in for training, reinforced by humans. There is only so much input (all recorded human knowledge), only so many human tweaks, that can produce only so much increased signal-to-noise in output. The machine can't read your mind, and there is no one truthful answer to most questions, so there will always be a limit on how accurate or correct or whatever any response will get. So at some point, you just can't make a better response. The agent harness, prompts, etc, are the only way to get better, and that's gonna be open source.
Add to that the algorithmic improvements on inference that's making inference faster with more context and higher quality. TurboQuant is just one example, more methods are coming out all the time. So the inference is getting more efficient.
At the same time, hardware can kind of keep getting infinitely better. Even if you can't make it smaller, you can make it more energy efficient, improve multitasking, more GPU cores/RAM or iGPUs, pack in more chips, improve cooling, use new materials... the sky's the limit.
Add all 3 together and at some point you will get Opus 4.7 on a phone with 40 t/s. At that point there's no way I'm paying for inference on a server. You can do RAG on-device, and image/video/voice is done by multi-modals. I want my agent chats replicated, but that's Google Drive. I want the agent to search the web, but that's Google Search. So eventually we're back to just doing what we do today (pre-AI) only with more automation.
The really advanced shit will come in 10 years, when we finally crack real memory and learning. That will absolutely be locked up in the cloud. But that's not an LLM, it's something else entirely. (slight caveat that WW3 will delay progress by 10-20 years)
This notion that "we don't have enough compute" does not cleanly reconcile with the fact that labs are burning cash faster than any cohort of companies in history.
If I am a grocery store that pays $1 for oranges and sells them for $0.50, I can't say, "I don't have enough oranges."
There is a major logic flaw in what you're saying.
'If I am a grocery store that pays $1 for oranges and sells them for $0.50, I can't say, "I don't have enough oranges."'
How about 'if I'm a grocery store and I see no limit on demand for oranges at $.50 but they are currently $1, I can say 'if oranges were cheaper I could sell orders of magnitude more of them'.
Buying oranges for $1 and selling for $0.5 is an investment into acquiring market share and customer relationships and a gamble on the price of oranges falling in the future.
> acquiring market share and customer relationships
The whole setup rests on this, and it seems mythical to me. These guys have basically equivalent products at this point.
Selling below cost is also called "predatory pricing". Sadly it's legal in US but it's something wealthy companies do to kill competitors and end up with captive customers.
> Buying oranges for $1 and selling for $0.5 is an investment into acquiring market share and customer relationships
It's a delusion that customers are going to remain with the behemoths when a Qwen model run by an independent is $10/m, unlimited usage.
This is not a market that can be locked-in with network effects, and the current highly-invested players have no moat.
[dead]
The grocery store analogy works if compute is the orange.
But labs arent buying oranges — theyre buying the only orchard on the island, hoping it yields a fruit no ones grown yet. Burning $1B to net $500M isnt "I have too few oranges." Its "Im betting the farm Ill find a new one."
Both can be irrational. Theyre irrational in different ways.
You can if you're exhausting the global production of oranges.
If there were more oranges you’d pay less to buy them and your economics would work out.
Not sure if this is a joke or not, but competitive pressure still exists. This only really holds if you're the only orange seller.
You misunderstand.
"I built a ship to go to the Indies and bring back tea."
"Bro, the ship cost 100,000 pounds sterling and only brought back 50,000 pounds of tea. I don't care if you paid 12,500 pounds for the tea itself, you're losing money."
There is a very rational reason labs are spending everything they can get for more compute right now. The tea (inference) pays 60%+ margins. And that is rising. And that number is AFTER hyper scalars make their margins. There is an immense amount of profit floating around this system, and strategics at the edge believing they can build and control the demand through combined spend on training and inference in the proper ratios.
60%+ margins according to numbers which are not published publicly and have not AFAICT been audited.
Could they be accurate? Sure, I think people who claim this is impossible are overconfident. But I would encourage anyone who assumes they must be right to read a history of the Worldcom scandal. It's really quite easy for a person who wants to be making money (or an LLM who's been instructed to "run the accounts make no mistakes"!) to incorrectly categorize costs as capital investments when nobody's watching carefully.
Any materially false public statement by one of the foundation lab CEOs is a huge foot fault. I'm not saying they would never lie, but it would be a very, very dumb thing to do. That public information can be relied on by their private (very powerful) investors. I think if you're hearing these numbers ballparked in public settings, they are, as a prior, directionally accurate.
I agree, although I would emphasize that Worldcom is a great example of a CEO doing that very dumb thing. But I am not hearing these numbers ballparked in public settings. As far as I can tell, all the numbers people discuss for OpenAI or Anthropic margins come from anonymous leaks of internal documents.
What limits LLM inference accelerators? I heard about Groq (https://groq.com/) not sure how much it pushes away the problem.
ASML only makes a certain number of machines a year that can do extreme ultra-violet lithography.
Also - turbine blades limit power, according to Elon.
Between them - we cannot chip fabs past a certain rate, and we cannot stand up the datacenter to run these desired chips past a certain rate. Different people believe one or the other is the 'true' current bottleneck. The turbine supply chain scaling looks much more tractable -- EUV is essentially the most complicated production process humans have ever devised.
Is ASML really the bottleneck? Do you believe anybody but TSMC and few fabs could really use and acquire those machines? I don't know the throughput of a EUV device from ASML but I imagine you need :
- clean room, itself needing the infrastructure for it (size, airCo, filtering, electricity) and the staff to run and maintain that basically empty space - wafers to "print" on, so that's a lot of water and logistic to manipulate them (so infrastructure for clean water and all chemicals) also with dedicated staff - finally staff who would be able to design something significantly better than NVIDIA, Intel, Broadcom, IBM, etc while (and arguably that's the trickiest part IMHO) being able to get it good enough as at a scale that can be manufactured from their own fab.
so I'm wondering who can afford this kind of setup that can only then make use of ASML machines.
> (so infrastructure for clean water and all chemicals)
Fabs are some of the most complex chemical engineering sites (dealing with some of the most dangerous substances) in the world. So don't underestimate the complexity of this part.
Well that was part of my point, not everybody is TSMC. It's not "just" getting an ASML machine and voila, you're good to go.
Presumably ASML can increase production if demand is high enough the question is over what time frame. 5 years seems plausible to me but I honestly don't know what that number is.
It's ... really long, according to Dylan Patel on the Dwarkesh Podcast. The supply chain is extremely deep and complex.
Yes. And the fab companies and their suppliers are deliberately and wisely slow to scale up production to meet short term changes in demand. They've seen the history of the semiconductor industry, it's constant boom and bust cycles. But they have the highest op-ex costs of anyone. So when the party's over they are the ones who pay for it the most.
Is global compute bottlenecked by one company?
Yes. At least, the manufacturing of compute is. And a lot of the chain has been bitten hard by increasing capacity prematurely in the past so they're reticent to increase bandwidth at vast cost.
If only there were some form of cheap, widely manufactured power generation technology that didn't use turbines... Are they really going to wait until 2030 to get more turbines rather than invest in solar?
I am clueless in this field, but solar seems to be unreliable and yield fraction of power required. Do you have a suggestion on something to read and learn more?
Reliability: complete solar deployment includes some form of power storage. There are many variations, but chemical battery technology is improving the fastest, so it's gaining the most ground.
Amount of power: World solar power generation capacity is in the terawatts and rapidly increasing, there's no issue with its potential ceiling. As a bonus, it tends to work best on land that's useless for other purposes.
Google china solar deployments to read about the logistics end of it
This is wrong along multiple axes.
1. Supply can scale. You can point to COVID/supply-chain shocks, but the problem there is temporary changes. No one spins up a whole fab to address a 3 month spike. Whereas AI is not a temporary demand change.
2. Models are getting more efficient. DeepSeek V3 was 1/10th the cost of contemporary ChatGPT. Open weight models get more runnable or smarter every month. Cutting edge is always cutting edge, but if scarcity is real, model selection will adjust to fit it.
Initially I thought "Well... good for AI companies because they can then charge more" but IMHO that's a very tricky position because it means the cheap wave is behind us.
It's one thing to "sell" free or symbolically cheap stuff, it's another to have an actual client who will do the math and compare expenditure vs actually delivered value.
> and compare expenditure vs actually delivered value
Which means that the hype production will be driven up another few notches to make people doubt their rational findings and keep them in irrational territory just a tad longer. Every minute converts to dollars spent on tokens.
Definitely feeling this - the subsidized subscription plans are already starting to buckle.
This isn’t really looking like AI scarcity it’s more like compute becoming the bottleneck : when the access depends on chips energy and capital it stops being a pure software game and the winners are often whoever can secure capacity first
Neither is this the first time nor are they really confronting it
A few years ago, I purchased a handful of 250GB SSDs from amazon for $17.00 each.
Last year, I purchased a few 8TB hard drives for $80.00 each.
Today, I am sad. ;-(
"adaptive" thinking
The current inference system is on a down slope.
It remains to be seen what new wave of AI system or systems will replace it, making the whole current architecture obsolete.
Meanwhile, they are milking it, in the name of scarcity.
distillation is an equalizing force
Distillation doesn't give you an equivalent model.
Why is written with an assumption that we have finite hardware production capacity? Industrial processes can scale up, new factories can come online… it will take a while but the whole point of economics is that supply will scale to meet demand. The shortage is a temporary, point-in-time metric.
And that’s not considering the software innovation that can happen in the meantime.
The economic hypothesis that has dominated the past hundred years is that economic growth is infinite because resources are infinite and (almost) free. We all know this is unrealistic and disconnected from our human condition.
Regarding "innovation", I agree with your idea. I even think that the major innovation will be to transpose models locally, using reduced infrastructures that will still be sufficient for the majority of use cases.
The models have already plateaued, you don't need latest and greatest.
... and I have this little idea in the back of my mind: when companies can no longer keep up with demand and people have (albeit more limited and reduced) local capacity, minds will start focusing on techniques (more humble and modest ones) to keep part of the system running locally, without dependency.
I know it may sound ridiculous, but it could actually become a way to break away from the business models that have been developed over the past few decades. Broadly speaking, this even amounts to saying that the biggest victims of AI could be the companies that bet on AI as a service.
Yet I know my vision is way too idealistic but I'm coming to imagine that a human brain, although less efficient in the long run, remains a reliable way to control the resulting costs and could even turn out to be more advantageous and more readily available than its silicon-based counterpart.
The human brain is incredibly efficient (Approximately 20W of energy consumption¹). These AI systems use many orders of magnitude more energy than human equivalents.
1. https://pmc.ncbi.nlm.nih.gov/articles/PMC8364152/
Does this also mean ram prices are not coming down anytime soon?
> Does this also mean ram prices are not coming down anytime soon?
One person replies "yes". Another replies "no".
This concludes our press conference.
<3 HN
yes, and it will keep increasing
they already are
It's artificial scarcity. LLM inference will soon be commodity as cloud.
There is a 2-3years still before ASIC LLM inferences will catch up.
The problem with this idea is that someone can, and likely will, come up with the next best architecture that leapfrogs the current frontier models at least once a year, likely faster, for the foreseeable future. This means by the time you've manufactured your LLM on an ASIC, it's 4-5 generations behind, and probably much less efficient than current SOTA model at scale.
It won't make sense for ASIC LLMs to manifest until things start to plateau, otherwise it'll be cheaper to get smarter tokens on the cloud for almost all use cases.
That said, a 10 trillion parameter model on a bespoke compute platform overcomes a lot of efficiency and FOOM aspects of the market fit, so the angle is "when will models that can be run on an asic be good enough that people will still want them for various things even if the frontier models are 10x smarter and more efficient"
I think we're probably a decade of iteration on LLMs out, at least, and the entire market could pivot if the right breakthrough happens - some GPT-2 moment demonstrating some novel architecture that convinces the industry to make the move could happen any time now.
I don't think so. GB200 prices are GOING UP. A100s are still expensive. This implies massive utilization and demand, no? These machines are not sitting idle, or prices would drop in the very competitive hyperscaler environment.
Hard to say at this point. I'm sure you can run your LLM chips 24/7 for training and for the public to make weird thirst-trap videos about Judy Hopps but how real is the utilization and demand, really? Maybe very real, maybe not, I don't think we can know yet.
Its like being back in 1850 and you build the world's first amusement park where the rides are free or very cheap. People are like Amusement parks are the next big thing since Steam Boats! And tons of other rich people start to build huge amusement parks everywhere. The people who are skilled at making amusement park rides will increase their prices, and since the first amusement parks are free so they can get the public going to them demand will be huge.
But how sustainable is that? - well obviously we know from history that amusement parks did, in fact, take over the world and most people spent virtually all their time and money at amusement parks - I think the Crimean War was even fought over some religious-based theme park in Israel - until moving pictures came out, so it worked out for them, but for AI?
so much for all that hardware that was going to be obsolete in 3 years...
Companies who could see it clearly and ignored the "AI is a bubble duh" crowd will ultimately get benefited by the GPUs they already acquired. The companies who acted cautiously will get burned.
wasnt ai supposed to get us post-scarcity?
That worked out, for the founders of frontier labs at least.
Companies who become dependent on AI to "optimize" their offering/processes are going to be faced with some serious vendor lock-in unless they do it in such a way that they can swap out the foundation model.
Pay for the latest AI for EXCLUSIVE POWERS ...
Trying to up-tier fractional improvements in something that can't be quantified easily, and !BONUS! with gated access it can't be as easily analyzed by the (low profit) AI analyzers/benchmarkers.
Foster paranoia among top executives that the fractional/debatable improvement is a MUST HAVE to STAY COMPETITIVE in your industry.
Meanwhile, I have not seen any improvement in software in the now almost ?three to four? years than mainstream LLM and AI coding assistance has arrived on the scene.
Although I will hold out the possibility that software has actually gotten far worse for the end user, because AI code is being dedicated to revenue enhancement and dark data collection.
"The first hit is free....."
And folks are just now realizing the SaaS token provider rug-pull?
How convenient, especially since everything has some LLM slop interaction.
But that rug isnt going to pull itself!
[dead]
[dead]
[flagged]
[dead]