I'm Not Quite Sure What Anthropic Wants
Anthropic, more than any other company, has made their goal abundantly obvious. They posture about how the world needs to stop all model development while publicly doubling down on the end of software development as skilled work.[1]
Of course, in their narrow-minded naive view, this is an excellent thing for their product's sales. If AI can do everything software developers can do today, “anyone” can do “anything”, so long as they have Anthropic's shiny new poorly-conceived life-sized remote-control hand warmer.
There's one problem with that. It's the word “anyone”.
Closed source software has been quietly losing for the last five years. In the last year, it's been loudly losing. People are, by and large, sick and tired of Big Tech's bullshit, and are increasingly willing to burn their house down (metaphorically) to get away from it.
So my question is – if we are going into a world where the work pays nothing (like Anthropic says we are), and everyone can do everything (like Anthropic wants us to believe we will be able to) – how does it not logically follow that everyone will work for themselves? How does this actually benefit their investors? What's the business plan?
The cynics say “well obviously then we'll all do manual labor”. The US economy cannot exist on manual labor. If you look up the definition of “service economy” in an economy textbook it's just the link to “Economy of the United States” on Wikipedia. And if you look at that nice little page, they've grouped it by sector, and you can see that 76.4% of the US GDP is services.
So, okay, we'll poof away services. Blow up the whole tech sector, because it's not going to live in a world where creating competition takes 3 hours and you can ask an AI to yoink all the data you want. The financial sector? If Anthropic has it their way, an LLM could build a better trading algorithm in the blink of an eye, which would make the hundreds of millions of dollars spent building them completely valueless. You could already automate almost the entire real-estate sector with AI.
The US can't exist in a world with AI. It would cut our GDP in half – at least. Overnight.
So, okay, there's two “solutions” to this problem. The first is that LLMs can't cause drastic over-crowding in every industry if their access is limited to those who already have a major position of dominance in an industry. The second is to just make LLMs so catastrophically expensive that they are out of reach of people and companies who aren't already established players in their field. These accomplish basically the same goal, and they both have the same problems.
The first issue is that, unsurprisingly, China as a whole does not seem to give a single fuck about the concerns of American corporations. Their economy is vastly more diversified than ours is, and their government has much greater options for intervening in it. Ultimately, there is still no evidence – despite the fact that the Chinese models have gotten closer and closer to the frontier with every passing month – that the best labs there plan to stop releasing the weights for their models, or that they plan to stop perusing efficiency improvements completely. Restricting access isn't going to work when an equivalent model is free (as in freedom) and costs less than a coffee from any of 9 million providers.
The second issue is that on paper, what Anthropic, OpenAI and Google are doing only really work because of economies of scale. Pre-training has gotten so remarkably expensive because companies like Anthropic are not in an efficiency arms race, like the rest of the world is. They are in a size arms race. This means that training costs more every single time, and they're going to either need to raise the price or get more customers to make up for that. They can't do either of these things, though, because the API prices are increasingly subsidized (I remember a quote somewhere that the real cost of serving Mythos 5 was at least 5x what they were charging for it after release), and they're already too expensive to be worth the widely reported 0 productivity gain over just hiring a software engineer.
Basically, what I'm trying to get at is that these AI companies have put themselves in a Catch-22 situation where the more viable they try to make their product, the less viable their customers are. There's no way for both Anthropic and their corporate customers to be viable businesses at the same time. Anthropic also can't sell to consumers long term, because their product is too expensive to run, and they can't slow their roll because their product only has value if they sell it at a cost they can't afford to people who increasingly do not want or care for it.
Some people have suggested that Anthropic's true goal is to integrate every single piece of software into Claude, one way or another. This still won't work, because it ignores the existence of other, cheaper, open-weight models.
Many people have also suggested that the open-weight models are going away soon. I don't think so.
Geopolitically, the best way for China to keep America's position of power in check here is to encourage every lab to release open models. The models the Chinese labs are producing are still unlikely to be as large as the models that American labs are producing, because – despite the complaints about their efficacy – the export controls for NVIDIA cards do actually seem to work. Since they can't really compete on size, their best bet is to compete on efficiency and cost, and the best way to do that is to release open-weight models.
Given all that, I don't see how the winner here is big companies. If intellectual property is post-scarcity, it's going to demand post-scarcity economics. In that world, open-source is the clear winner.
1. On a slightly stranger note, if all of this holds water, I think it's worth taking Dario Amodei's pleas for UBI and a freeze to all model development at face value. They need it more than we do. ↩