I've been following them for years, and every year everyone says they're not worth it.
It started with "AI progress is fake." "AI will never beat human creativity." "AI can't code anything past 15k tokens." "AI will always be too expensive." "AI will never sound like a human." "AI can't even understand humor."
And now we're at "AI is only like this because they're stealing copyright." "If everyone uses AI, it will cannibalize its own outputs." "Other people can build LLMs too."
But year after year, they solve these problems. They weren't worth a hundredth of it in 2021. At this rate, they would be worth a tenth of it in 2026, and maybe a oneth of it in 2030. And that's what the VCs are backing on. If they're not, well, it converts to debt.
Google (the search engine) hasn't really had a moat either. Plenty of competitors, some better ones, but they're still around. ChatGPT is a brand name.
Yet you're ignoring that competitors, have done a terrific job at catching up in extremely small time frames.
Hell, even Nvidia 2 weeks ago released their own LLM model which is competitive with the best commercial ones.
AI is a commodity, eventually where money will be made (interacting with the software and user, from Excel to your OS like Apple Intelligence) it won't be much relevant which model you use, average Joe won't even notice.
I don't consider models a commodity though. Coffee and steel are commodities. You need some level of quality and materials, but they can be swapped out. Models are more like engines or CPUs.
Unbenchmarked things like their ability to use south Jakartan slang, writing jokes, how and when they reject input, how tightly they adhere to system input, how they'd rate a thing from 1-10. They function as a part of a complex system and can't be swapped out. I'm using Claude Sonnet 3.0 for a production app and I need a week to be able to swap it to 3.5 while maintaining the same quality. We've trained our own models and it's still incredibly hard to compete with $0.075 per million tokens just on things like cost of talent, hardware, electricity. And that speed.
The question is why not something like Anthropic?
I'd say OpenAI has other cards up their sleeve. The hardware thing Jony Ive is working on. Sam Altman invests into fusion power and Stripe; guess who's getting a discount? There is a moat, but it lies at a higher level. Other competitors are also playing other kinds of moats like small offline AI.
I have also worked on an ai chatbot for legal purposes, and the differences between models were negligible when most of what you're doing is RAG (99% of apps out there).
I have also tried the different models in cursor and again the differences were negligible with some projects and questions slightly favoring one or the other.
Which does nothing but confirm that none of those has any kind of moat really. Data and products are eventually what's gonna make money, and the models used will likely be an implementation detail like choosing a database or a programming language.
To be a commodity, there has to be a constrained supply. So-called AI is just software and the replicable nature of software means it cannot be a commodity. I'd agree on your other point that the competitors have done a great job of catching up to a point where the valuation makes little sense given the limited moat.
It started with "AI progress is fake." "AI will never beat human creativity." "AI can't code anything past 15k tokens." "AI will always be too expensive." "AI will never sound like a human." "AI can't even understand humor."
And now we're at "AI is only like this because they're stealing copyright." "If everyone uses AI, it will cannibalize its own outputs." "Other people can build LLMs too."
But year after year, they solve these problems. They weren't worth a hundredth of it in 2021. At this rate, they would be worth a tenth of it in 2026, and maybe a oneth of it in 2030. And that's what the VCs are backing on. If they're not, well, it converts to debt.
Google (the search engine) hasn't really had a moat either. Plenty of competitors, some better ones, but they're still around. ChatGPT is a brand name.