On: Most AI startups are doomed

[gimili]

I think it is more complex than the author thinks.

There are clearly defensible aspects for ai startups. Specifically I think these are: a) in-context and collaborative features (since working alone with ai through a chat box is unlikely the only way we will interact) b) gated knowledge/data (since commonly available technology can be leveraged with unique data) c) edge computing and offline usecases won't be the center piece for many classical companies and therefore can be very well exploited.

I wrote up a framework to assess LLM powered Startups/Ideas here: https://assistedeverything.substack.com/p/the-three-hills-mo...

[neptudemon → gimili]

Doesn't (a) fall into the bucket of UI, i.e., something that can be easily copied?

Agreed on (b) - I think this is anyone's best shot at a moat.

Curious to see how (c) evolves. It's unclear to me whether the future of these things are running locally or whether we'll all continue hitting remote APIs

[gimili → neptudemon]

I don't think (a) is a pure UI thing.

Think of the difference of using a single-user application to e.g. make mockups for websites or a collaborative environment like figma, in which you happen to also be able to have AI collaborate with you. Very different usecases and solving collaboration workflows, etc. is non trivial.

I guess for (c) both things will exist. Local will be done for 2 reasons: - data sovereignty (e.g. companies wanting to have applications that are purely trained/fine tuned on their own data; but that improvement is not shared) - privacy (anything from an AI having access to all your email and calendar up to having intimate "friendships" with AI)

[j-wang → gimili]

I think various of those aspects you call out here, I do as well. The specificity of the application is fairly key, whether it comes through proprietary data or application-specific stuff or simply business-lock-in.

Interesting hill analogy—I do broadly agree with the areas.