That is a very astute and concise way to explain everything about how the frontier labs are behaving and how they're trying to push more people to pay token rates for the best models. At the current subscription prices ($100 or $200 a month for a generous, though bounded, amount of tokens), frontier models are a no-brainer, most folks and companies will use them. But, at token rates, 10x or 100x the cost of open models or what I was spending on the frontier models a month ago? That is a harder question to answer "yes" to. I certainly wouldn't spend $1000 a month for the best model, much less $10,000; my employer might pay $1000/month, but definitely not $10,000. The frontier labs need everyone to answer "yes" to spending 100x what they currently spend to justify the valuations, and it's just not going to happen as long as everyone knows how to make these models.
Both OpenAI and Anthropic are trying to figure that out now. Anthropic, in particular, has their finger on the trigger...they want to push people to usage-based billing for Fable. But, OpenAI released 5.6 Sol, competitive with Fable (or close enough), and it's available via subscription (even the $20 subscription!), and there's no moat keeping someone from switching. If Anthropic really does end Fable access on the subscription plans in a few days, I predict a large market move back toward OpenAI.
The market isn't going to bear the cost of making the frontiers investment make sense.
It would have been an interesting experiment to charge more for it right away and see what the market would bear, rather than tease it for long enough for it to be presumably superseded any time now by whatever is next.
I think an interesting question is going to be, if models are a commodity, who is going to want to foot the very expensive bill to train them? I'm sure training cost will drop.. eventually, but I doubt it will happen fast enough for any of these companies.
And we can't ignore the power of "good enough". GLM5.2 may not be as good as the SOTA models, but it can be good enough for most, of not all, of our needs.
Think airlines - both passenger and freight. They have never come close to capturing all the economic value they enable.
That's not sign of commodity actor, just the opposite.
It's running, privately, in my homelab.
I think we are entering what I call the "have it your way" era. If an open source project doesn't do exactly what you want it to do, fork it, or create a new version. It's too easy.
This makes me a bit concerned about the future of open source. Upstreaming used to be worth it, since maintaining a fork is effort too. But now the balance has shifted significantly. Especially with many projects becoming a lot stricter about contributing, and some becoming outright hostile to AI. I can't blame them. But I think the effect will be that improvements are less likely to make it back to the community as AI adoption increases.
There is a lot of OSS software out there (e.g. in scientific communities) that I would say would barely qualify for each of those three attributes. The main reason it's valuable for the respective communities, is because it's the only thing that's available.
There's two reasons for that. The math is generally very unorthodox and alien for a seasoned developer, and software development practices are equally alien for the scientist who can understand and evolve the math behind it.
I have written a boundary element method evaluator for my Ph.D. not only math was alien, the required coding techniques for making it fast is very different for a standard developer. You have to have the perseverance and interest to do that. I chose that path intently and I do not regret a millisecond of it.
The problem is, if you don't have a dedicated team to continue that codebase (e.g.: like the Eigen team), your code is basically done and done. If somebody doesn't share the same passion, it's almost impossible for someone to take and carry it forward.
Oh, due to the math and optimizations, the code's structure need to be both documented and the next batch of developer(s) have to be tutored by the person who's giving the code to them.
There is, of course, the question of if that's making me dumber. It might be, but there are other brain training things I'm doing outside of that to force my brain to do the thing.
LLMs are poison for the brain, I'm almost certain of it, at least when used in the way most people are using them. If you drive everywhere because you don't want to walk (but you could), you're obviously going to be physically worse off than if you walked. This is the case with llms, if you have them do all the thinking, planning and action you're going to be cognitively worse off than if you didn't use them.
The average Joe can easily vibe code apps that took a small startup just a few years ago. If developers are also using AI to build the same simple apps - then yeah. They're not pushing themselves hard enough, and probably not using their brains as much anymore.
1966 saw the peak of calculator protests, where math teachers claimed similar things of calculators.
I'd bet there's far more 'good enoughs' than anything else out there. One of the reasons microsoft office is constantly churning subscription, etc is because they solved good enough decades ago and need to justify valuations that just don't matter for most of their user's use cases.
Not everyone is a software developer having to churn out the 101th SaaS that's just because some MBA refuses to hire a dev.
I have seen so many unnecessary forks of popular projects that I think it's better to stick with the original, even if that means it won't be perfect.
As cost to software goes to zero, these things become easily possible. In the past, I'd only fork top-quality software (things like `xsv` etc. which is easy to edit. These days even complex PHP software I fork with little trouble.
With lots of software, the value is in the data model and algorithm choices. Sometimes I even just point Claude Code / Codex at an open-source thing I want to vendor some functionality into my personal setup with and it gives me what I want. The hard part for me is modeling the data well. That takes experience with encountering things and it's hard to replicate the edges. LLMs often don't get the rough bits right. But someone else's hard work usually has accounted for this.
Since it's just a duplicate folder, I can always fall back if it fubars.
The latter as always been more durable. Linux doesn't have the mindshare it does because it's "free" as in beer - it's because it's "free" as in freedom.
The price of freedom, of brewing your own beer, is sometimes higher than buying it from the store. But for many folks, the control over the supply chain is what makes it worth it. In LLM-land, it might take a little bit of time for folks to catch up -- or maybe a lot of that is already in motion as companies get paranoid (and rightfully so) to frontier labs getting a little grabby about data. If you need a ZDR environment, "free" as in freedom has a very high premium that you will pay and rightfully so.
Seems reasonable
We could try steelmaning this argument instead: it's enough that most big companies who would otherwise have incentives to contribute.
Before FOSS got in fashion, around the early 2000s, most commercial companies wouldn't touch it as contributors and were openly avert to it, and to open sourcing their stuff. This can be the case again.
You can use an LLM to create anything but you still need to know what it is that you're building, and you need to think through how everything should work or the LLM will just fill it with sausage. You can tell that the models are still quite jagged and limited by the mixed quality from a lot of the software that these presumed trillion dollar companies are putting out. The future is sausage.
Makes perfect sense to anyone good at using these models. What doesn't make sense is that analogy. Typing prompts isn't even close to as difficult to baking bread.
It doesn't really, because whenever I ask them what did they actually create, its always a shitty dashboard or a finance tracker or something derivative and worse than what is out there
…but consider: the Q-tip. “Don’t use it to clean your ears”, but for most people that’s all they want to do with it, and empirical observation indicates that this dynamic results in either “using Q-tips irresponsibly” or “not using Q-tips”, with “uses Q-tips properly” being a small-to-vanishing proportion of the whole.
I think the opposite: I think the frontier labs have good margins on their inference unit costs.
We can already see what it costs to run near frontier-size models. There are independent business pivoting to serving these models at reasonable prices and they're competing on OpenRouter for costs much lower than frontier labs.
> Is there any guarantee that I'll be able to run a Opus 4.8-level model on my personal computer before the big AI labs decide to hike up the prices?
I would bet good money on prices going down significantly, not up.
If we get to the point where you can run an Opus 4.8 model on your local computer, it's going to be even cheaper for a datacenter to serve it on their hardware. That means prices crash, not that they're going to rise.
1. Much of those profits have to be immediately reinvested into model training runs to avoid being lapped by competitions.
2. Unit costs are irrelevant when the labs don't price per unit, and instead charge, for instance, $200 / month for $10k worth of tokens.
This isn't a steady state. Whatever the current situation is, I doubt it's sustainable.
Cost to generate all of the tokens divided by revenue generated by selling those tokens is what matters.
The subscription plans confuse a lot of people because that's what they see. They're not seeing the gigantic API bills from all of the tokens going into enterprise use cases.
The subscription plans are a small part of their income. Most users aren't maxing out 100% of their plan usage every week. I wouldn't be surprised if their average plan user was using less than 50% of their monthly quota each month.
Plans like that can produce a net increase in profit if they get consumers interested in the brand and pitching it at work. Giving them some extra token headroom on their $20/month or $100/month home plan is money well spent if it gets all of a company's developers advocating for enterprise plans with budgets exceeding $1000 per person.
Token prices are going down. Competition is global. A company could choose to keep their API prices high, but if another company comes in at 1/10th the price for 95% of the performance then they won't have many customers.
Using a full Claude Max 20x plan to 100% of weekly usage would easily cost you 2k through the API. While the Claude Max 20x plan is 200 a month.
I doubt many of their customers are on the 20X plan. Of those, I doubt many of them are using 100% of their weekly usage regularly.
Comparing the 100% maximum usage scenario of their most discounted plan against the API cost has been a trap in this conversation since it came out. I bet if we saw their financials it would be a tiny sliver in a pie chart somewhere.
Also, DeepSeek V4 Pro is cheap via any commodity API, and DeepSeek V4 Flash is essentially free at API prices like $0.09/M, $0.18/M out. This is generally not subsidized.
For a more practical local setup, Qwen3.6 27B on a used Nvidia 3090 (US$1300) or two is surprisingly nice. It needs clear instructions and you can't use it for hands-off vibecoding, but it's actually quite reasonable for hands-on programmers.
You're always guaranteed that you can stash away the open models!
If subsidies do end, demand for price efficiency per unit of intelligence will go way up.And because there's many players in the market, this demand should be met by at least some of them.
I wonder what he thinks was too harsh, still seems pretty bang on, I think it’s going to age well.
I think he now thinks agents can maybe program a little bit.
Having a thinking trace that is legible, coherent, and immediately implies the explicit turn output and/or tool use seems difficult if not impossible to reliably get from mixture models.
I predict MoE is a transitional technology, it's got too many problems and the benefits are...kinda grandfathered into the dogma at this point.