The continuation of the AI investment bubble has me confused and a little concerned. Is there any way the huge amounts of resources pledged to AI companies will pay off? I don’t think so. I believe it’s happening simply because LLMs are genuinely extremely impressive and give the appearance, much more than we’ve ever had before, of true artificial intelligence and big investors hate to miss out.
Let’s think about the possible outcomes on a spectrum. On the far left is the failure case. AI just can’t do much more than it currently does. It’s extremely impressive, but it’s mostly good for some software development uses, and even there the return on investment is more mixed than it first appears. Somewhere in the middle, we have the case where AI continues to improve, but only with great expense, and useful more advanced models cost a fortune to run. There, we get benefits like new drug discovery, new math proofs, possible solutions to engineering problems and so forth. This is the world most people envisioned seven or eight years ago. Finally on the far right we have the case where AI keeps getting cheaper to run and we keep seeing meaningful improvements in capabilities.
Wherever we land on this spectrum, I think we’re in for a bad time. In the first two cases it’s pretty obvious the outcomes don’t support anywhere near the present valuation of the big players. For the last – the “optimistic” case – it’s less clear and I think that’s the most interesting and the most likely future.
The usual critique is that To justify the scale of investment, LLMs must be insanely effective. I have a different concern. Even if the AI companies can show good results and deploy useful AI at a large scale, it won’t guarantee a lot of growth.
In the short term, AI work agents destroy a lot of their advertised value, to the extent they get deployed and used at scale. They do this in two related ways: First, vast amounts of AI produced work lowers scarcity of that work, lowering the price. But more supply doesn’t automatically create more demand. Second, when a large portion of available art, music, advertising, etc. is AI generated, it makes it hard to sort out valuable work from junk, which makes these previously valuable quality signals uninteresting.
Flooding the market with supply doesn’t automatically boost demand. So what are all these companies tooling up with AI thinking? They’re thinking they get cheap labor. AI – if it lives up to current expectations – floods the market with cheap workers. There’s a market for labor just as much as there’s a market for music and art.
Increasing supply doesn’t automatically increase demand. However, too much labor for office work demand may free up resources to be used in new ways that previously weren’t worth the cost, resulting in more total use. I suppose it’s possible that once we destroy the demand for white collar labor new types of jobs (for the USA anyhow) may come back. Ever notice that hiring servants for your household is terribly expensive? No, because it’s so out of reach for the average household it’s not even a consideration except for the ultra-rich. Well, in the future we could see some growth there.
Scarcity
I made a new programming language in Rust that compiles to the Cranelift back end. The core of the project was hand written. I got it to a working state using a simple interpreter. Adding Cranelift support would be a lot of work. I did it all with Claude Code. Claude was incredibly effective. The whole thing works pretty well.
In the past this project would have been of some interest to the community. Completing a language design, parser, high performance JIT compiler is a big undertaking. Or it was. I notice that I don’t even have a lot of interest in the project myself; I didn’t really write it after all. If I can’t be bothered to care why should anyone else?
Scarcity is an under-appreciated part of what makes stuff interesting to people. A lot of people fixate on the artifact and not what it represents: Gold is interesting only because of its scarcity obviously. Less obviously the same is true for music. You can see the whole sector hasn’t grown; it’s actually gotten smaller. The point is an infinite amount of availability doesn’t translate into infinite consumption. There’s a somewhat fixed appetite. In 1995 when you couldn’t get ahold of your favorite artist’s new album for two weeks you developed some anticipation and were more than ready to put down $18 on the CD. If you can stream it whenever with your existing sstreaming service you don’t tend to care as much.
Tools like Claude Cowork make artifacts that used to require a lot of human effort cheaply and prolifically, with little oversight. I’d separate what coding agents and general work agents do into two parts. The actual work – reviewing code or large document collections, coding and testing, analysis is the work. Slide decks, spreadsheets, graphs and final reports are the work products. There’s a related discussion concerning signaling to be had but let’s just focus on scarcity for now.
AI tools can make work products in minutes that could take hours to do by hand. The actual writing is a bit cringey, but it can improve.[1] The back end intellectual work is kind of hit or miss. A lot of tasks are pretty mechanical and can be done effectively by current work agents, saving time if not brainpower. The intellectual side seems to get better and better with every release of new models. Long term I think this is where AI will really benefit everyone.
The value proposition to the individual is that work agents save time and even give the worker capabilities they may not have had: Preparing nice presentations, coding software etc. Also saving time on tasks they can do but would rather not. To firms, the value is that they can do more with lower spending: Fewer workers get more done per worker and fewer specialists are needed. For investors, it’s that whoever has the best AI displaces lots of traditional software as a service , and automates whole new classes of tasks that SAAS didn’t touch.
In each case the value proposition makes sense in isolation, but if you imagine a world where software is super cheap, reports and presentations can be summoned in minutes from the machine and so on, it looks more like printing (too much) money. Your economy is only so large. Printing more bills just lowers the value of each bill. Like if one person finds a suitcase full of money it’s great for him, but if everyone does you’re back to square one.
If AI tools really produce good output and decision making for a given quantity of human effort, they will increase productivity by definition. But right now it seems like they’re being used as a one to one replacement for current work, like AI note takers, AI generated presentations, and so on. Really who cares about power point presentations when they’re everywhere? There isn’t an endless appetite for most of this junk. So the growth will be only incremental when it comes to near term uses of AI work agents. And that’s if the outputs are total replacement level quality.
The best you can say is that AI work tools remove a bottleneck if you have unmet demand. If a business was constrained by how many workers they could afford or how fast they could produce, now they can do much more. Again though, this only applies in a world where competitors don’t have access to the same technology. In the real world, where their competitors have access to the same technology it’s a race to the bottom. If demand stays the same, what ends up getting cut is jobs.
With reduction of human employment we see the final hurdle to making huge returns: The simple fact that with nobody employed, the customer base disappears. It’s a glib sounding point but there has been a serious look at the idea. As more and more companies see real benefit using AI to replace workers, their competitors are forced into the same choice.
We can take this one more step, applying the argument to access to AI itself. AI will get cheaper and cheaper and how can any one company keep a monolopy on it? Anthropic, Open AI and Google all seem to be developing models at a similar pace. It’s hard to see how one will win AI and swallow up the whole economy. On top of that, open weights models keep getting better, and cheaper to operate. Even if AI technically exceeds expectations, who’s going to collect the profits if it’s trivial to run a model in your basement? To be clear, I think the big AI labs will continue to have the best AI and will make a lot of money, just nothing like what their valuations today suggest.
So to sum up, it’s not clear that what AI produces has a huge appetite waiting to be filled, it’s not clear the current AI companies can even capture a lot of that growth, and the growth that happens won’t be transformative in any case.
Signaling
Signaling is a big reason fancy Power Point presentations are worth anything in and of themselves. The ideas communicated sometimes matter too, but the audience expects a bare minimum of effort and the fact someone knew how to put together the slides and took the effort to do so is a signal. It’s kind of like decent resumes in the old days. It’s a superficial but useful signal. This concept extends to most of the work products AI work agents can make. They completely scramble the proof of work signal – “proof of work” meant generically here, as in did anyone put in the work?
since AI agents can do some forms of intellectual work really well, even the ideas behind the work products get devalued. When you get past the window dressing of a presentation or grant proposal, you still have the underlying concepts they’re trying to communicate. When you see someone spent two years researching and cataloging some obscure topic, delivered by a crappy presentation, it tells you at least one person cared enough to put in that research effort and it might be worth taking it seriously even though their slides suck. Not any more.
Right now, new, conceptually interesting software projects are being launched at a crazy high rate. Which ones are worth investigating? [2] Currently it’s still useful to ignore pure-AI generated software when you spot it, but for how long will we be able to tell the difference? All that’s saving us at the moment is the sprinkling of AI cringe writing. That will change. Right now you can spot most pure-AI projects by the unmistakable AI style: Use of emogi in README files, excessive bullet-points, “Not just X, it’s Y” constructions. In any case, aside from caring how a project got built, the sheer quantity of projects makes the value of any single one miniscule.
It does make me wonder why in the long run AI work agents should even make presentations, spreadsheets, reports and front-end software. All these artifacts only have value if there’s someone on the other end to read them. Whole departments may collapse into tighter more efficient units. It’s silly to have AI agents emailing each other and scheduling meetings with each other after all.
Long term prospects
Longer term, the value of AI will be in places where a larger amount of intellectual work is valuable and was never available before. The good version of this is home assistant software that actually understands your request to turn off the lights and can carry out multi step complex instructions. The bad side is super effective cheap mass surveillance.
There are some other good uses such as reviewing archives of scientific papers at a huge scale to produce new insights, reviewing safety critical code for bugs and other places where skilled human attention is unavailable.
- Earlier LLMs like GPT-2 and Gemma2 actually have a more human and varied writing style, trained out of them in later versions in favor of predictability. We know they’re capable of it in theory.