The cost argument is a fallacy, because right now, either you have a trained human in the loop, or the model inevitably creates a mess.
But regardless, services are extremely cheap right now, to the point where every single company involved in generative AI are losing billions. Let’s see what happens when prices go up 10x.
After re-reading the post once again, because I honestly thought I was missing something obvious that would make the whole thing make sense, I started to wonder if the author actually understands the scope of a computer language. When he says:
> LLMs are far more nondeterministic than previous higher level languages. They also can help you figure out things at the high level (descriptions) in a way that no previous layer could help you dealing with itself. […] What about quality and understandability? If instead of a big stack, we use a good substrate, the line count of the LLM output will be much less, and more understandable. If this is the case, we can vastly increase the quality and performance of the systems we build.
How does this even work? There is no universe I can imagine where a natural language can be universal, self descriptive, non ambiguous, and have a smaller footprint than any purpose specific language that came before it.
You're going to pretty hard pressed to do Rust better than Rust.
There's minimal opportunity with lifetime annotations. I'm sure very small options elsewhere, too.
The idea of replacing Rust with natural language seems insane. Maybe I'm being naive, but I can't see why or how it could possibly be useful.
Rust is simply Chinese unless you understand what it's doing. If you translate it to natural language, it's still gibberish, unless you understand what it does and why first. In which case, the syntax is nearly infinitely more expressive than natural language.
That's literally the point of the language, and it wasn't built by morons!
I believe the author thinks of this problem in terms of “the LLM will figure it out”, i.e. it will be trained on enough code that compiles, that the LLM just needs to put the functional blocks together.
Which might work to a degree with languages like JavaScript.
@manuelabeledo: during 2025 I've been building a programming substrate called cell (think language + environment) that attempts to be both very compact and very expressive. Its goal is to massively reduce complexity to turn general purpose code more understandable (I know this is laughably ambitious and I'm desperately limited in my capabilities of pulling through something like that). But because of the LLM tsunami, I'm reconsidering the role of cell (or any other successful substrate): even if we achieve the goal, how will this interact with a world where people mostly write and validate code through natural language prompts? I never meant to say that natural language would itself be this substrate, or that the combination of LLMs and natural languages could do that: I still see that there will be a programming language behind all of this. Apologies for the confusion.
To be generous and steelman the author, perhaps what they're saying is that at each layer of abstraction, there may be some new low-hanging fruit.
Whether this is doable through orchestration or through carefully guided HITL by various specialists in their fields - or maybe not at all! - I suspect will depend on which domain you're operating in.
>After re-reading the post once again, because I honestly thought I was missing something obvious that would make the whole thing make sense, I started to wonder if the author actually understands the scope of a computer language.
The problem is you restrict the scope of a computer language to the familiar mechanisms and artifacts (parsers, compilers, formalized syntax, etc), instead of taking to be "something we instruct the computer with, so that it does what we want".
>How does this even work? There is no universe I can imagine where a natural language can be universal, self descriptive, non ambiguous, and have a smaller footprint than any purpose specific language that came before it.
Doesnt matter. Who said it needs to be "universal, self descriptive, non ambiguous, and have a smaller footprint than any purpose specific language that came before it"?
It's enough that is can be used to instruct computers more succintly and at a higher level of abstraction, and that a program will come out at the end, which is more or less (doesn't have to be exact), what we wanted.
Doesn't have to be "a clear definition", a rough defition within some quite lax boundaries is fine.
You can just say to Claude for example "Make me an app that accepts daily weight measurements and plots them in a graph" and it will make one. Tell it to use tha framework or this pattern, and it will do so too. Ask for more features as you go, in similar vague language. At some point your project is done.
Even before AI the vast majority of software is not written with any "clear definition" to begin with, there's some rought architecture and idea, and people code as they go, and often have to clarify or rebuilt things to get them as they want, or discover they want something slightly different or the initial design had some issues and needs changing.
This is the most handwaving per paragraph I've ever seen.
I think a fair summarization of your point is "LLM generated programs work well enough often enough to not need more constraints or validation than natural language", whatever that means.
If you take that as a true thing then sure why would you go deeper (eg, I never look at the compiled bytecode my high level languages produce for this exact reason - I'm extremely confident that translation is right to the point of not thinking about it anymore).
Most people who have built, maintained, and debugged software aren't ready to accept the premise that all of this is just handled well by LLMs at this point. Many many folks have lots of first hand experience watching it not be true, even when people are confidently claiming otherwise.
I think if you want to be convincing in this thread you need to go back one step and explain why the LLM code is "good enough" and how you determined that. Otherwise it's just two sides talking totally past each other.
>This is the most handwaving per paragraph I've ever seen.
Yes: "LLM generated programs work well enough often enough to not need more constraints or validation than natural language" if a fair summarization of my point.
Not sure the purpose of "whatever that means" that you added. It's clear what it means. Thought, casual language seems to be a problem for you. Do you only always discuss in formally verified proofs? If so, that's a you problem, not an us or LLM problem :)
>Most people who have built, maintained, and debugged software aren't ready to accept the premise that all of this is just handled well by LLMs at this point.
I don't know who those "most people are". Most developers already hand those tasks to LLMs, and more will in the future, as it's a market/job pressure.
(I'm not saying it's good or good enough as a quality assessment. In fact, I don't particularly like it. But I am saying it's "good enough" as in, people will deem it good enough to be shipped).
> I don't know who those "most people are". Most developers already hand those tasks to LLMs, and more will in the future, as it's a market/job pressure.
This is definitely not true. Outside of the US, very few devs can afford to pay for the computer and/or services. And in a couple years, I believe, devs in the US will be in for a rude awakening when the current prices skyrocket.
The "whatever that means" isn't a judgement jab at your point, merely acknowledging the hand waving of my own with "good enough".
I hope this comment thread helps with your cheeky jab that I might have a problem understanding or using casual language.
I'm not sure if it's moving the goalpost or not to back away from a strong claim that LLMs are at the "good enough" (whatever that means!) level now and instead fall back to "some devs will just ship it and therefore that's good enough, by definition".
Regardless, I think we agree that, if LLMs are "good enough" in this way then we can think a lot less about code and logic and instead focus on prompts and feature requests.
I just don't think we agree on what "good enough" is, if current LLMs produce it with less effort than alternatives, and if most devs already believe the LLM generated code is good enough for that.
I use LLMs for a lot of dev work but I haven't personally seen these things one- or even many- shot things to the level I'd feel comfortable being on call for.
>I just don't think we agree on what "good enough" is, if current LLMs produce it with less effort than alternatives, and if most devs already believe the LLM generated code is good enough for that.
Don't need to consider what they think, one can just see their "revealed preferences", what they actually do. Which for the most part is adopting agents.
>I use LLMs for a lot of dev work but I haven't personally seen these things one- or even many- shot things to the level I'd feel comfortable being on call for.
That's true for many devs one might have working for their team as well. Or even one's self. So we review, we add tests, and so on. So we do that when the programming language is a "real" programming language too, doesn't have to change when it is natural language to an agent. What I'm getting at, is, that this is not a show stopper to the point of TFA.
In the same way in Rust you can download a package with Cargo and use it without reimplementing it, an LLM can download and explore all written human knowledge to produce a solution.
Or how you can efficiently loop over all combinations of all inputs in a short computer program, it will just take awhile!
If you have a programming language where finding an efficient algorithm is a compiler optimization, then your programs can get a lot shorter.
Even those are way more predictable than LLMs, given the same input. But more importantly, LLMs aren’t stateless across executions, which is a huge no-no.
> But more importantly, LLMs aren’t stateless across executions, which is a huge no-no.
They are, actually. A "fresh chat" with an LLM is non-deterministic but also stateless. Of course agentic workflows add memory, possibly RAG etc. but that memory is stored somewhere in plain English; you can just go and look at it. It may not be stateless but the state is fully known.
Using the managed runtime analogy, what you are saying is that, if I wanted to benchmark LLMs like I would do with runtimes, I would need to take the delta between versions, plus that between whatever memory they may have. I don’t see how that helps with reproducibility.
Perhaps more importantly, how would I quantify such “memory”? In other words, how could I verify that two memory inputs are the same, and how could I formalize the entirety of such inputs with the same outputs?
But that’s not the point I’m trying to make here. JIT compilers are vastly more predictable than LLMs. I can take any two JVMs from any two vendors, and over several versions and years, I’m confident that they will produce the same outputs given the same inputs, to a certain degree, where the input is not only code but GC, libraries, etc.
I cannot do the same with two versions of the same LLM offering from a single vendor, that had been released one year apart.
Enough so that I've never had a runtime issue because the compiler did something odd once, and correct thr next time. At least in c#. If Java is doing that, then stop using it...
If the compiler had an issue like LLMs do, the half my builds would be broken, running the same source.
There are people out there who truly believe that they can outsource the building of highly complex systems by politely asking a machine, and ultimately will end up tasking the same machine to tell them how these systems should be built.
Now, if I were in business with any of these people, why would I be paying them hundreds of thousands, plus the hundreds of thousands in LLM subscriptions they need to barely function, when they cannot produce a single valuable thought?
I've been an Apple One subscriber for over three years now. For the past few months, as soon as you open the TV+ app, a Peacock ad starts playing really loud.
Then I don’t know what to tell you. I just opened the app again, and right there in the home section I’m seeing an ad for the Super Bowl in Peacock. If you don’t get that, great, but I’m far from the only one complaining about it.
It would be great if folks would stop assuming this is on me and not Apple. There are Peacock ads in the TV app Home Screen, and they are targeted to One Family and Premier subscribers.
Maybe it’s just me. I’ve been an Apple One subscriber for a long time now. The Peacock commercial I’m talking about plays right when I open the app, almost full screen and quite loud. It seems to be some sort of add-on offer for Apple One subscribers.
While I agree that third party advertising is not the same as playing trailers from other same platform shows, once you are in the app, these highly promoted shows are really hard to miss, regardless of how many trailers are placed at the beginning of another show.
For Shrinking, for instance, they placed an almost full screen, auto play trailer in the main carousel. It is also first in the top ten shows, and it appears in a number of other lists.
Regardless of all this, they do play unrelated promotions for their add ons like some sports stuff or the Peacock deal.
It's an ironic comment because this article mostly talks about California, which is already one of the most expensive places to live and the most NIMBY. Every other state in the US is generally cheaper to live in. The places that are cost as much as California are just as NIMBY and heavily influenced by Californians (Hawaii) or is the cultural and financial center of the country (NYC).
> Look up property taxes, cost of living expenses, and overheads like parking, schools, etc.
I currently live in an arguably not very dense city, in the suburbs. I pay thousands of dollars in property taxes. I must own two cars to serve the whole family, for things as basic as going grocery shopping. My HOA is almost a thousand dollars a year. A couple years ago I had to replace the roof, at a cost of several thousands of dollars.
I had none of these problems when I was living in a more dense city, and on top of that, I could actually walk to the nearest coffee shop.
> Is NYC the cheapest place to live in the country?
NYC is dense because it appeals to more people, and the more people that move to the city, the more expensive it gets, precisely because there are not enough homes.
Are you assuming that less dense cities are more desirable to live in? Is Anchorage a more appealing city to live in than NYC?
There's an aphorism that I like to use every time someone tells me that Windows is the easy choice: Windows makes accomplishing easy tasks easy, and hard tasks impossible.
To this day and for the past 20 years, every time I go back to my parents' house, it's Windows tech support time. Every time, I have to go through the same routine of cleaning up all sorts of crap that make their computer slow to a crawl, even when I purposely created non privileged accounts for them. Every time, the same ritual: diagnose why the printer stopped working, why apps look pixelated, why some sites stopped working.
So, it works for my parents, and possibly works for the vast majority of people, because it has created this ecosystem where these users either depend on folks like me, or have to pay up at the repair shop. And somehow, something as simple as not breaking a critical component during an update, or prevent users from installing harmful stuff, has not been addressed properly. And that's without mentioning what Microsoft does that's very much anti-consumer, like stuffing the consumer versions of the OS with ads.
But regardless, services are extremely cheap right now, to the point where every single company involved in generative AI are losing billions. Let’s see what happens when prices go up 10x.
reply