alt Hacker News> If more people could tinker with local models and see what they can do, I think there would be far less belief that only the big two/three hold the keys to the kingdom and far more that the future is a bit more distributed
Eh, I've played with them. They take way more babysitting to do the work reliably, and I have to much-more closely monitor the outputs for hallucinations.
They'll get there. But it might take a decade or so for (a) consumer hardware and (b) SOTA distilled open-source models to converge with the closed-source stuff, and that's enough time for both a slow deflation and plenty of profits to be made.
I was specifically referring to consumer AI, not agentic coding, and I really think Apple have a good chance of demonstrating smaller on-device LLMs plus traditional symbolic AI work (even just rules systems like evolved forms of Automator and Shortcuts in harnesses) can do much of what consumers want.
The current industry obsession with agentic coding is a giant red herring, in a way. Or at best a shiny thing. The pitch from OpenAI for their extraordinary valuation has always been absolutely all-encompassing in terms of utility and market. But we've stopped talking about that.
I don’t think these vast LLMs will be the dominant approach in a decade, frankly, because the performance gap between a 35B MoE and these absolutely gigantic models just is not proportional to the scale difference, but at this point we are just pitching belief against belief.