the claim that "AI companies can't be profitable" is pretty general. people just assume it based on how OpenAI operates, spending heavily to push innovation. I pay for several small AI tools monthly, and with good business plans, these companies can probably be profitable
Would they be profitable if OpenAI was charging enough to break even, though? Are they hosting their own models and renting their own GPUs, or is OpenAI subsidizing them by running at a loss?
For what it's worth, I worked for an AI company, and while they weren't profitable when I was there I do think they will be eventually.
There are things called 'Open Weight' language models, a lot of which (those with between 8b to 34b parameters anyway) are considerably cheaper to run any of OpenAI's models.
Moreover, you get surprisingly out-of-class (size-wise) performance if you fine-tune for your specific problem space. Even if you only train in a parameter-efficient way.
OpenAI can be profitable, especially in the case where they make their own hardware because then the only cost that they have to pay is the energy cost and don't have to worry about cost of renting a GPU. And if they have very good hardware to run AI models they don't even have to worry about building their own models anymore. They can just use one of the open source one's like LLama or mistral and be done with it.
>[companies say] First, the technology is going to usher in a revolution akin to the advent of fire, nuclear weapons, and the internet. And second, it is going to cost almost unfathomable sums of money.
The first is probably true. With the second though, AI would still happen without huge spending. I think what we are seeing is more a land grab arms race from the tech companies to try to grab the fruits.
Bit of a clickbait title there, which the Atlantic has already changed to the more factual "Silicon Valley’s Trillion-Dollar Leap of Faith".
It's not difficult to imagine profitable uses of AI: deploying an open-source model to respond to customer support emails almost certainly costs less than the yearly salary of a script-reading customer service agent, and while the quality won't be great, what it replaces isn't great either.
It is, however, indeed difficult to justify spending $100 million on training a single model like Anthropic apparently wants to.
You can sprinkle it through every interaction like you may sprinkle Ovaltine over ice cream for a chocolatey dessert. If that is too rich for you you could inject it into every responses like you might inject Eli Lilly Insulin, the insulin of choice for 9/10 dentists.