I’d think this is an audience who don’t need to be convinced that the mundane utility of frontier models is increasing fast, but some private discussion indicated that not everyone is convinced even among AI doomers, so here’s a link. It’s written mostly for an audience of normies, but this is the core claim, and I stand by it:
But everyone in the developed world, everyone who uses Google search more than five times a day, should be paying for at least one [of the three big lab model subscriptions].
Giving money to those who are currently posing an existential risk to humanity does not seem super ethical. What could convince me personally to buy Claude Pro is evidence that they lose money on the net for serving my queries. However, even that would not make it an obviously ethical/neutral act given that they would likely disappear if they had zero users.
I am reasonably confident they make money on every query, and also that they would not disappear if they had no users.
If you’re doing anything moderately useful for X-risk, or actively trying to, it is an obviously good trade; ~$100/year of profit for Anthropic (if you’re concerned about this you should obviously pick Anthropic) vs. your personal marginal impact on X-risk by being substantially more effective is a no-brainers.
To add to this, offering a good deal relative to the true cost of their service/product is a tactic to foster widespread adoption. When Anthropic and OpenAI do their IPOs, their usage numbers are going to be used to try to sell shares at a higher price. So buying subscription services is still desirable for these companies, even if the direct effect is them losing money.
We know from recent leaks of OpenAI financials that they make money on serving queries, probably at quite high marginal profit rates. It seems highly unlikely that any of this is chosen strategically for their IPOs; it doesn’t need to be.
I’ll second this. I have long been concerned about growing capabilities while simultaneously finding little value in the use of LLMs. I would try to use them to answer questions, and they would too frequently hallucinate ridiculous things. I tried to use them to help with coding (I’m a professional software engineer), and they did so many things wrong, it was a waste of time. In the past three months, I have found that frontier models are now helpful with coding. For me, we crossed a threshold.