Hacker News Reader

> Are local models anywhere close to gaining enough capability and traction to do it in-house?

Cloud providers will always be able to offer more performance and more powerful options.

I don't know, I could envision a situation where investing enough capital in hardware to be able to generate "infinite" tokens for free (or close to that) might open up possibilities not economically practical otherwise. Especially once the incumbents run out of hype / investor money to burn and start raising their prices to push revenue in line with their enormous expenditures.

Also, presumably at some point far in the future we'll reach a technological asymptote and factors like latency may start to play a bigger role, at least for some applications.

I grant that training data is crucial distinguishing factor that may never become competitive in-house.

Reply on news.ycombinator.com