The Other Shoe
Well, that was fast.
In last week's blog post, I wrote about how I thought that, after the release of Fable 5, I'm not sure how Anthropic could possibly get to a viable business model. I think this wasn't expressed as clearly as it could have been (I wrote it at 3 in the morning), so let me restate the premise:
- Anthropic, OpenAI, and every other AI company built their business model around selling tokens both business-to-business and direct-to-consumer, at varying levels of stability. Increasingly, they've doubled down on business-to-business sales to software corporations, because historically this has been where LLMs have performed strongest.
- Fable 5 represents a significant jump in real-world AI capability. It performs much better on benchmarks than any other model, obviously, but the bigger deal is that it finally gave some actual weight to Anthropic's claim it would be able to lower the bar of entry for programming to “you can describe what you want to make”.
- If the bar is this low, then over-competition in the software industry is inevitable. If everyone is able to develop any application over a short time span, it's not likely that anyone will be able to hold a grasp over any sector for any real amount of time.
- If over-competition leaves the software industry non-viable, the only thing Anthropic will have accomplished is destroying an entire sector. They will not be able to sell their products to customers who do not exist.
I proposed that this left Anthropic and company with three options:
- Option 1: Pump the price of LLM access so high as to be unreachable to anyone who isn't already an established player. The reasons this one won't work are fairly obvious: firstly, the value of an LLM subscription has to be greater, in essence, than a the value of a programmer, including the fact that you still need someone to prompt the model. Secondly, it is exceedingly unlikely, due to China's geopolitical position, that open-weight models are going anywhere. If they continue at the current pace of development, sure, they'll be worse. I don't think it'll matter.
- Option 2: Try to replace all of their customers with their own product, and lock down the model to only Anthropic employees. This requires an absolutely insane amount of people to do the prompting anyways. It's not cost effective, it's incredibly risky, and it's still undermined by open weight models because it relies significantly on the fact that access to a comparable model is limited entirely to people inside an AI company. The world this is asking for is a world where the AI companies develop and own all software, and in a world with open weight models I just find this categorically unlikely.
- Option 3: Put frontier models behind a whitelist. I surmised that this was unlikely to work because it would destroy the economies of scale necessary to justify R&D costs, which means it would effectively result in Option 1 anyways. And again, this assumes that the big American companies would be able to completely control all access to close-to-frontier models, which, as I'm sure you've understood by this point, I don't think is possible.
One of the key arguments that someone made against this is that surely someone at every single AI company knows this, and if they did they would have stopped.
I agree, for the most part. I think that Sam Altman and Dario Amodei (the CEOs of OpenAI and Anthropic, respectively) have both done the calculus on this and are both aware of this. Altman seems to think that this will inevitably destroy every single service economy on the planet, and thus the solution will be to (somehow) enact global UBI and just give everyone AI credits in the hope that someone will find a way to come up with a viable business model. I think this is a pretty bad plan and it's one he's gotten a lot of flack about.
Amodei's idea, fortunately, isn't so bad. Just about every time the man's been interviewed, he's said that he'd like a globally coordinated freeze on AI development. His argument, which is one I agree with, is that this is not something that the AI companies are able to do independently due to market pressures, and even an agreement between all major players will inevitably fail once a new player enters.
Many people have called this anti-open-research, hypocritical to the extreme, or just the CEO posturing for PR. I don't think it is.
Because I do, genuinely, agree with him. The point is, essentially, that yes, we've opened Pandora's box. We can't close it. But collectively, as a species, we can agree to not open another one. And the second one will break the entire world as we know it, so it's in the best interest of everyone, even the AI companies, to not open that box.
He's not proposing “I got my bag, get bent”. He's proposing nuclear arms treaties for AI development.
That being said, I didn't think it was likely to happen. I'm still not sure it is, but there's been some interesting developments in the last few days that I think are worth considering.
Today, OpenAI announced GPT-5.6 Sol. It's supposed to significantly leapfrog Mythos. And just like Mythos and Fable, the US government has come bearing down on OpenAI and demanding that they limit access to trusted third parties approved by the US government.
It seems like someone at the Department of Commerce has finally figured out what the hell is going on.
I don't think that the general public is likely to ever see these models again. I think the US government is going to voluntell us into Option 3. And I think they're going to “fix” the issues by import-banning Chinese models, and by simply not caring about the viability of OpenAI and Anthropic.
This is too little too late, and if the goal is to stop the US economy from falling apart, this isn't going to work. It's covered in too much classically American “the rest of the world will never catch up to us” thinking. This will buy some time, sure, but I don't think the US government in it's current capacity is thinking “Yes, we'd like START but for AI! We'll start drafting the treaty tommorow!”. I think that it's more so “surely this will accelerate American companies past every single other company on the planet”.
There's not much more to say about this. I hope I'm wrong, but I think that the American solution, if I'm right about what it is, is going to just create a ton more economic inequality before backfiring on everyone in the entire country, wealthy or not. It's the classic shortsightedness that plagues modern economic planning, but this one's got real consequences that might not be fixable, even after everything's said and done.