But I'd call false dichotomy on this:
> brute-forcing software may be the best solution to make money. So if you are rich and can brute force software in a dumb way, chances are that you can make a business out of it. If you're poor but know what you are doing, you cannot compete.
Seems to me like the big winners will have know-how and some money to throw at AI.
What I find interesting about this is that Anthropic should be better equipped in terms of talent and token accessibility than any other company. Yet, it seems that this doesn't actually provide a fundamental advantage in terms of writing software yet. If I compare Claude Code with Pi, then I personally would prefer Pi over Claude Code, but even if one sees it the other way, I don't think there are many serious people who would say that Claude Code is by far a better harness than Pi. But Pi has a handful of people working on it with, from what I've gathered from reading Mario's and Armin's tweets, Claude Max subscriptions. So it doesn't seem as if throwing tokens at a problem is a major differentiator in creating software as of today.
This just says "if you have money and no idea what you are doing, you can throw the money at the LLM and it will build something that you think is a new OS"? Maybe it's just running a Linux distro with a different UI... who can know if nobody looks at the code?
Back to the original argument about this reinforcing inequalities, 1000 agents working for 8 days is well within the capacity of a handful of people's $200 plan for a year.
Also you could just prompt a new AI session and ask if it all it made was a linux from scratch OS and you would get a pretty thorough investigation a lot cheaper than an maybe more accurate than a consultant trying to figure it out without being allowed to use any AI tools.
Sorry, what? So you're saying this experiment cost them $2400 in tokens?
There's also the cost of the coding environments, but you would have that with humans making an OS too.
May still be a bit off with 5.2 taking more usage (and they also didn't specify the reasoning level) but it's not going to work out to reinforcing inequalities vs the expense of hiring people to make an OS that can boot and look like anything in the clip.
So you assume 50 plans for 22 agents (not sure how you got 21), so each plan runs 0.5 agent for 24h, so 1 agent for 12h without stopping, with Fable 5.2 and with tons of context (and context-passing between the agents, and re-reading what the other agents created, etc).
I just tried Fable 5.2 on a pro subscription, I can tell you that the weekly quota doesn't last more than a few hours. Let's say 3 hours, but that's generous (and it was not at busy time where it gets more expensive, I think?). So from 1 agent running 12h per day, with pro I have 1 agent running 3h per week. I need 28 times that to get to your estimation.
You mentioned the Max subscription, where the pricing page says it has 5x more usage than Pro. So... I would need 6 times what you estimated? That's $720,000 and for my* estimation where I am careful to pass as little context as I can. Their experiment probably has context overhead, I can't estimate how much.
My point is that I wouldn't be surprised at all if you were completely underestimating the price. And now back to your last sentence:
> vs the expense of hiring people to make an OS that can boot and look like anything in the clip.
Their clip shows a super simple UI. They don't say if it wrote a kernel. They say "network", does it mean that it re-invented the whole concept? Probably not... did it reinvent the physical layer then? IP? TCP? Or did it reimplement TCP from perfectly valid examples that already exist? Or did it assemble open source networking components? Not clear at all. In any case, I am not impressed at all that an LLM can rebuild some of the most standard and widespread code there is.
I can build an OS in a few hours, the smart way. Sure, I won't be building everything from scratch. But neither did their agents. Whatever they rebuilt to get to that video clip was mostly wasted, because really everything exists already.
That is what I mean. If you can waste millions, then you can brute-force something that someone not-so-experienced could assemble in a few hours. If there is money to be made with the output of that "work", then the rich will benefit from AI. That is reinforcing inequalities: knowing better doesn't give you an edge anymore, it's a lot more productive to be richer.
And we could say "sure, it's not impressive because it's an OS, and open source OSes probably take most of the open source world, but that's just an example, right? With that kind of resources, they could build something truly innovative and useful, right? Right? And they just chose to waste it for a video clip, right?
See where I am going? By brute-forcing it with tons of resources, they managed to... copy some of the most well-established open source software out there. Being able to copy something successful without understanding how it works by throwing money at it does reinforce inequalities in my book.
As for inequality that's a matter of monetary policy not of how anthropic spends it's money.
Which is not new, but this "reinforces" it.
Edit: to be clear they spent lets say 50,000 dollars to make something that is worth 0 dollars (or negative). If there is money to be made with AI it's in serving and the ideas of the users, but the biggest users will get ripped off by smaller, and so a race to the bottom.
What I am saying is that AI makes it easier for the rich to brute-force their way into something they don't really know, while the poor who knows what they are doing is slower than the brute-force, wasteful method of the rich.
This makes the rich, richer. And the poor, poorer. Therefore it reinforces inequalities.