[everdrive]

An interesting point I don't think I've seen someone make -- people compare the LLM revolution to other technical revolutions. You don't need to worry about skill decay in the same way that you don't know how to bake bread from unprocessed wheat, or you don't know how to build a loom, etc.

But local models aside (which no matter the protests from HN, will only be available to the technically savvy few) all of these LLMs are a service, so, the company could degrade the service, they could charge more than you're willing or able to pay, they could ban you. They could disable your account with no meaningful way appeal or seek support. LLMs could look at lot more like the scenario in this thread than something like not knowing how to make your own shoes.

[kbelder]

It might settle into a situation where cutting edge LLMs are a service, while older and smaller LLMs are self-hosted. So you are not at risk of being cut off, but of being degraded.

[donmcronald]

I hope you're right. I played around with a bunch of AI stuff recently and that's kind of the conclusion I came to. Use local AI for mission critical stuff, if you're confident in that, and use the SOTA models for reviewing.

Tap the latest general knowledge for asking "could this be improved", but make the improvements with local systems and models. But then the obvious problem becomes finding new data to train the AIs. In my opinion, there's no way their plan doesn't involve stealing from everyone to keep training, so is it really going to be safe to use the cutting edge models at all?

[donmcronald]

If they manage to build good memory systems, people will stop keeping personal docs and rely on the “AI” for everything. Imagine 20 years from now when people don’t even have copies of the recipe to bake bread and then you’ll see what the goal is.

[throwaway290]

And then in future if you try to build something to reverse the situation your coding llm becomes stupid and your psychologist llm recommends you some blue pills.