If LLMs used very little RAM and processing power, was light weight even for Mythos or Sol, would people be as opposed? Like no data centers, no memory shortages, like it was able to run on a calculator with the same performance.
If LLMs used very little RAM and processing power, was light weight even for Mythos or Sol, would people be as opposed? Like no data centers, no memory shortages, like it was able to run on a calculator with the same performance.
That’s fine to disagree with me but I think it’s a pretty generous reading of OP’s phrasing to take away that all she’s asking is “What if AI ran on your phone?” Having said that: you’re still making some huge leaps here that do approach sci-fi if not magic outright. Firstly quantum computing as a field has stalled. To compound on this, transnational technological collaboration is also slowing. What this means is that it’s unlikely we’ll see another spurt of insane growth in mobile tech that will rival a $24,000,000 rack of Vera Rubins any time soon. The processor tech we have today seems pretty well plateaued.
Locally run AI can exist, and has for a long time now. It’s very feasible, and also resource-intensive, especially on hardware that wasn’t designed for it. The issue of data centers also isn’t really removed here, just mitigated. First and most pedantic: data centers have been a thing since the internet was created. When you pull up a video on YouTube, that data is being served through a data center, it’s only against the insane demands of AI generation that our existing supply of data centers isn’t enough. YouTube doesn’t need to generate that video every time you watch it, just store one copy and fetch it for you whenever you ask. The second point is that locally run AI still needs to be trained. Even offline models need basic updates to prevent model drift and keep accurate knowledge (to be useful). This training data needs to come from… a data center. Admittedly, in possibly smaller amounts than full-access to an online generalized LLM.
Anyway, another point: there was a lot of trial-and-error to get to the cheap smartphones of today from 1985. 41 years of it. In retrospect, most people would say it was worth it (most). Will people say the same about AI? It’s doubtful. This is for the exact reasons I said above: the criticisms of AI come from when it works right! They’re criticisms against the product itself, not like those against early smartphones (“It’s too expensive!” “The battery dies too fast.”) It’s almost a given that efficiency will improve, but is this a good thing? More slop and more surveillance for cheaper? I don’t think so.