Should I Start an Generative AI Company?

Probably not (yet?)

note: I’m going to lower the friction to what I post here. I’ll get more out, they’ll be shorter. Not full meals, but a pastry on the way to work.

I don't personally think generative AI (LLMs, image diffusion, etc.) has settled enough for any specific business idea. Similar to maybe the first year or two of the iPhone: Lots of new capability overhang, but also most ideas are dead ends. Traps like how the first great-sounding ideas with large TAMs get swallowed up by the platform (then: Apple, now: whoever-supplies-the-model). The first big social media apps got subsumed by Facebook. E.g. for LLMs, maybe you build some writing aid on top of GPT, OpenAI realizes they can absorb that functionality in GPT-N+1, and your startup is dead. You do not want to risk 2-4 years of your life on that. I spent 2 years of my twenties doing the same for social media.

I think it's time for idea labs to prototype like crazy. As

Alex Komoroske

sometimes puts it, not deep focused innovation but combinatorial exploration: Throwing different building blocks together over and over and tinkering until you find something new.

What’s new this time around with GenAI is the wildly-new uncertainty at the "platform" layer: the platform surface is not a durable API you can migrate from/to. It's a fuzzy, probabilistic “API” that absorbs some new skills with each version in unpredictable ways: So betting 2-6 years on, e.g., applying current LLMs to some-large-market is taking on an additional risk of "maybe the LLM is materially different in 6 months", and making that bet up to 12 times in a row. And it only has to come true once for you to be sunk.

Smartest people I know are figuring out what a lab would look like. Where you're playing with possible applications without building any tech too deeply on any specific capabilities. Gets you an option factory.