Pessimistic, maybe, depending on one’s perspective, but hardly authoritarian. No one’s stopping you from doing as much hand-coding as you want.
1 Like
It would perhaps be useful if you didn’t try to vibe code your argument. But as near as I can tell, you’re linking to an article about something the CEO of a music streaming service said on an earnings call, and are trying to tacitly assert a) that it should be taken fully at face value, b) his claims about what some of their engineers have done over the last couple months are representative of a “new normal” that can be both extrapolated along all axes…that is, that what they say they done with some of their engineers over the past couple months should be extrapolated out into a prediction about what they’ll be doing indefinitely with all their engineers, and c) that they’re representative of the industry on the whole and we should extrapolate our extrapolations about what some undisclosed fraction of the engineers at one company have been doing out to literally all of programming in perpetuity.
I don’t trust comments on an earnings call like I would a technical analysis, because they are primarily a marketing thing intended to help the share price, so saying your company is really doing well with the Next Big Thing everyone is talking about can’t really be taken as a technical evaluation of the technology. I don’t see why Spotify should be considered a bellwether…I can’t think of any other time I’ve thought “As goes Spotify so goes the world” or anything like that. In 2023 Spotify was all in on blockchain. Should we make similar projections on industry adoption of crypto as a result? And I don’t think Spotify represents all, or even a meaninful cross-section of, requirements for code in general. What about kernel development? Med tech? Fields with regulatory compliance issues like those than handle medical or financial data? Code with real-time requirements? Very low fault tolerance? Things like avionics, ATC, medical devices, even things like ECUs and HVAC systems, power grids, and so on. And if the assumption is “just build LLMs for them”, then I don’t think there’s a straight line from where we are (“admirable” success is code that doesn’t even compile) to there without a whole lot of speculative steps.
In other words I disagree with literally all of the assumptions. So I guess to copy the style here, “Yeah, seriously.”.
There’s nothing wrong with speculating, but I want to take a moment to emphasize that what I’ve been doing in this thread is trying to minimize the amount of speculation I have to do when talking about this technology. That is, talking about what it can actually, testably, do.
So I’ll just say that what you’re relying on here isn’t just an extrapolation from what gen AI is currently capable of doing, as we started talking about in this thread, but instead are imagining some other, entirely different mode of usage and development. In other words, what you’re imagining here quite literally isn’t an LLM. It’s a different class of AI.
Like I said, there’s nothing inherently wrong with speculation, but I do think it’s worthwhile to at least be aware of the difference between “here what this thing does, here’s what you might get if you just tweak the sliders a little” and “here’s some other technology that I can imagine that might one day be invented with the current technology as its ancestor”.
I think that for any Singularity-ish, end-of-history speculations about AI to be even vaguely plausible they’d have to be accompanied with at least some napkin math about, say, how many kW-h it takes to replace a single mid-level programmer, what data center footprint it will take, how much that costs in up-front hardware costs and what the day-to-day burn rate is, and so on.
I also think there needs to be serious consideration of what the actual business model is. Right now AI models are not profit centers, they’re loss leaders. And the companies building them are spending many multiples of their annual revenue to build out capacity in the hope that the next big breakthrough, whatever that is, will lead them back into the black. I think whether or not that is even going to happen is an open question, but even if we accept that it will, what does that lead to? In your hypothesized future it is more or less every company putting their ability to function behind someone else’s paywall.
9 Likes
FWIW, in the “engineering alumni” group for [huge tech company], whenever I see a post polling about AI usage, approximately everyone who responds says AI is being used for 90%-100% of coding wherever they’re working now.
“The last couple months” are worth extrapolating from, IMO, because the models have gotten a lot better in the last couple months. I don’t think anyone expects them to get worse, so it’s hard to imagine why those numbers would go down.
Isn’t that where most such companies already are with other infrastructure providers? Every company that has a website—or uses email, messaging, teleconferencing, file sharing, e-signatures, etc.—and isn’t running it out of their own datacenter is reliant on someone else who’s selling it to them.
Well, currently the big AI companies are operating at a loss to drive adoption. That’s not a sustainable business model in the long term, or even the medium term. I don’t foresee the effectiveness of any individual model going down, but I can easily imagine the effectiveness of $100/month worth of access going down, once the companies start trying to actually make a profit. (Or whatever benchmark of cost you want to use; I don’t know the current going rates.)
ChatGPT is a household name by now, but OpenAI lost $12 billion in the last quarter alone. Sora, which they’re charging users 10¢/second for, is costing them $15 million per day. The whole industry is currently propped up by speculation and venture capital, but once that runs out, where else will they get the money to keep things running except their users’ pockets?
Or to put it differently: once companies are onboard with firing all their programmers and replacing them with LLMs, I expect the LLM companies to start charging approximately a programmer’s salary for them. Disruptive new tech always starts out cheap, but compare Netflix to cable TV now, or an Über to a taxi, or an AirBNB to a hotel.
Right, but AWS hosting isn’t costing Amazon dramatically more to operate than it costs me to rent from them. OpenAI’s own reports (according to what I’m reading online) don’t predict turning a profit until 2029.
4 Likes
Even at that price, it would probably be a good investment: one programmer with AI will likely be a lot more productive than two without. (I’m certainly more than twice as productive with AI than without, and I don’t think I’m exceptionally good at coding with AI or bad at coding by hand.)
To put that estimate into perspective, the cost comes out to 13 cents per second ($1.30 per video). I think a lot of customers would still be willing to pay $2 per video instead of $1, which would leave OpenAI with 70 cents of profit.
I agree. I was specifically responding to the point about companies “putting their ability to function behind someone else’s paywall”, which is a thing companies do all the time today.
This is already happening. Companies paying for tokens rather than fixed usage plans (e.g. Claude Max x20) are finding they are spending this kind of money. That this is not profitable either should worry everyone about the “what comes after” stage.
1 Like
It definitely worries me at least. I’ve recently gotten involved in the KoboldCPP community because, admittedly, the fact that a powerful technology like this is mostly driven by large corporations which are hemmorhaging money scares me.
Google and Anthropic do seem to be doing alright, OpenAI is beginning to show cracks. As such, with the technology now having developed, there’s arguments to be made for learning how to use the smaller, more focused models on local hardware combined with open source software driving them.
The risk of enshittification is large, and it helps to use these new (very powerful) tools responsibly. In other words, never forget how to do the thing manually. Dependency is, and I know this sounds dramatic, dangerous
2 Likes
Actually it would likely not.
The $15 million per day figure quoted is the current cost of running the service.
Additional to that will be the costs associated with:
- the repaying the many 10’s (100’s?) of Billions in “loans”, that were used to build the models & data-centers.
- the payment of any relevant Taxes, once income is being generated.
- saving for future capital expenditure.
Once the real ongoing costs are known, and the owners have determine what level of returns they expect to receive, then they can determine the level of markup they want and what the future price(s) of the service will be.
4 Likes
I recently saw an analysis of the current state of OpenAI. They basically had the advantage of being first to market, and thus getting market share early. But, parties like Anthropic and OpenAI have since caught up and are slowly eating away at their market share. In many ways, Claude and Gemini have proven to have better models at many tasks.
What is notable is that investment parties are beginning to pull out of continued investments into OpenAI, because they’re beginning to lose trust. So either they quickly become profitable, or we see them concede their top spot to the other big names.
Side note, it’s wild that this seems to have shaped into the central thread on discussions regarding (ethical) AI usage and development. I have to admit it pleases me to see how civil the discussion has remained, despite the forum being a place mostly full of creative people and programmers, an intersection of skills that could be considered ‘threatened’ by these developments.
Then again, I consider myself both an artist and developer, but I have kind of been enjoying the technology too much to form a hostile opinion towards it. It’s genuinely been helpful in doing tasks I least enjoy in my workflow. I don’t know how this is to anybody else.
Either way, I love how the forum as a whole does seem to have an appreciation of the craft itself at least, because losing that fascination means losing an understanding of the thing. It’s why I recently very much enjoyed visiting a “repair café” in the neighbourhood, as it was full of technical people who knew and were willing to help repair hardware, clothing, etc. But it was equally notable how many of those people were already retired.
4 Likes
It would, by definition. You seem to be splitting hairs between gross profit and net profit.
Yeah sorry I can go down the flight of fancy path.
The reason I’m optimistic is inventions like this:
I really don’t think these big data centers are going to last and some may never get off the ground.
btw - sent you a dm here on intfiction. Any chance you read it?
This discussion reminds me of the old saying: “You can’t have Good, Fast and Cheap, but, you can pick Two.”
It seems that in the AI age of software generation, everybody gets to pick only One, especially when they think they get all Three.
Still, being able to pick One of of Three is better than before, when it’s usually Zero out of Three.
2 Likes
I’m currently working on a game. And although I’ve been experimenting with Claude Code, I’m intentionally not working on the game in Claude, because I don’t want to have to add any disclaimers to it. In other words AI will be involved in 0% of the coding.
I agree that we’ll probably rapidly reach a point where all some close approximation of all code at large tech companies would have to carry an AI disclaimer. What constitutes AI “being used for” coding seems to be fairly malleable at this point, and what people will report as AI doing “100%” of the work seems to be kinda flexible, too.
But I’m not trying to split hairs on the subject. My understanding was that the Spotify story was intended to be a counterexample for my observation that LLMs are weak at reasoning outside of their corpus (and therefore we better hope they get better at it or that we have all the languages we need, if we’ve reached the end of the era of humans writing “original” code). That is, the “seriously?” was incredulity that I should suggest that current LLMs couldn’t produce something like Rust, because Spotify apparently has some bespoke thing. I could be mistaken, though.
To put my skepticism in perspective, earlier this month an engineer at Anthropic reported burning through around $20k worth of tokens to build a compiler that could compile the linux kernel (for a specific architecture):
That’s an article that summarizes the crunchy bits; it links to the original blog post announcing the result.
I think it’s a fairly informative illustration of the difficulty of even making unambiguous statements about whether AI “works”. In that I’m absolutely willing to concede that it is unimaginably beyond the kind of results anyone could have reasonably expected until very recently, but saying the compiler “works” involves an awful lot of nuance that literally nobody would think of including in literally any other context.
Like if you just put it up as an anonymous git repo and announced it as a new compiler I don’t think people would be shy about saying it was just fundamentally busted. That it really doesn’t do what it was designed to do. Nobody would, like, start to preferentially use it instead of whatever compiler they might have been using already. It is entirely because it was produced by a novel means that it deserves any attention at all, like an ambitious science fair project by an especially precocious elementary school student.
And AWS capital expenditures for AI capacity absolutely dwarf capex spending for cloud infrastructure prior to the AI boom. This year they’re estimating they’re going to spend over $200 billion. For comparison that’s more than the three biggest cloud providers spent on capex combined in 2023.
3 Likes
Hi, there is a misunderstanding. The ‘seriously’ message was not directed at you but at the ad hominem message from the person I quoted. I don’t know why it also picked up your quote.
And to be clear, I never referred to programmers as code typists. What I mean by the ‘day of the code typists are over’, is that the value in the process of typing out the code (which is time consuming), are long gone.
We went from ‘ai code completion is quite clever’ to ‘jeez this is now really good’ in record time last year. In the past weeks we went from ‘code produced needs frequent manual tweaks’ to ‘it rarely needs a fix’. Things are happening so fast that it looks like the hockey stick effect is here and we are close to a phase change.
Over the decades, I’ve seen so many technical innovations pitched at business management as doing away with the need for programmers. Inevitably, though, they hire programmers to deal with the latest innovation in data mining because they still can’t be bothered to learn the tool or apply it properly to the problems they want to solve.
I suspect LLMs, as fascinating as they are, are much the same. The sorts of analytical minds that are good at programming are also good at analysis and problem solving and the need for that will never go away as long as managers are too impatient or busy to sit down quietly and think deeply.
I do see LLMs and voice recognition pretty much eliminating most customer service jobs, though. So, there is going to be some sort of economic realignment. Hopefully that means universal basic income and a shorter work week.
My biggest concern with LLMs is that they talk a good game, but will happily lie to you. Who wants an employee like that? Frankly, the lack of trust is a dealbreaker for me, more so than the risk for plagiarism.
With regard to plagiarism, I think about collage artists and remix culture. LLMs seem like they are in the same category. Therefore, the results may fall within fair use sometimes (most times?) It really depends on the output. The LLM itself is not creative, but the human who wrote the prompts and curated the results is creative. This is not to say the creative folks who made the original corpus that trained the LLMs should not be compensated; they certainly should be.
I recently corresponded with someone who knows nothing about programming and is trying to program a video game with an LLM. He described his process of iterating with the LLM and it sounded so incredibly frustrating. Learning to actually program seemed like the more efficient path, honestly.
In summary, I neither love nor hate LLMs. I am, however, very skeptical.
I would have completely agreed with this comment prior to December.
In two months the tech has made an insane leap in credibility.
2 Likes
I’ll describe an example. I bought a Mac-mini in November with the intention of using Claude to help me ramp up iOS development. I started down the path of a lesson plan, but Xcode and Swift are not trivial things to learn.
I pivoted and asked Claude to just do everything I needed for an app I wanted to build.
That app is now in Test Flight and includes a Rust backend and a relatively complex UX, all of which was coded by Claude.
Prior to December I believe this would have been impossible.
Ah, no worries. Thanks for clarifying.
1 Like
I really disagree, and I’m not sure if the disagreement is “just” semantic or if we’re modelling the situation differently.
I guess if you want to say that literally any use of a service provider qualifies then yeah, but I don’t think that reflects the situation with genAI.
There just aren’t that many competing offerings, and they’re really not all more-or-less equivalent in the sense that cloud providers, for example, are. Spinning up a new cloud provider (which I’ll consider as a proxy for how the market one the whole might react to pricing changes) is much more plausible than spinning up a new coding LLM that would be competitive with the existing model (unless all the existing examples are just doing something specularly wrong and so their costs are astronomical compared to what a competitor could be doing). We know what a profitable cloud hosting provider looks like, but we have literally never seen one for AI coding (and as far as I know none of the existing crop expect to see a profit for several years. and it’s worth pointing out they have enormous incentives to being as optimistic as possible on this front).
But bottom line, if you’re extrapolating out from where we are to a future where literally all programming is done by genAI, you’re looking at a future where control of more or less literally all technology rests in the hands of approximately three corporations.
And for the record I think we already have too much control in the hands of two few entities. But having literally all of tech going through whoever controls the LLMs is an approximately maximally dystopian outcome.
4 Likes
I’d say we’re basically in that same boat already with cloud hosting providers. Having seen some of what goes into providing cloud hosting, I don’t share your optimism about how easy it would be for someone else to spin up a competitor.
Nor do I agree that cloud hosting providers are “more-or-less equivalent” in a significant way that coding AI assistants aren’t. Choosing between AWS and GCP is more of a commitment than choosing between Claude Code and Codex, for example; you can change models with two clicks in Copilot or Roo, but switching cloud providers is a major undertaking.
1 Like