I could only skim to verify an automated summary, but if the takeaway is "these AI giants are doomed" then he's right.
The future, probably within 10 years, is most tasks being handled by small on-portable-device models (7B parameters or so; see Apple's Intelligence thing), a middle ground of workhorse models (pushing closer to 30s and 70s) running on more capable ML-focused chips in laptops and workstations, and home and office servers for the biggest professional users running on dedicated servers.
And then there's the apps. Whoever makes the "Stripe for generative AI" with multiple models with different levels of data provenance, security, SLA, etc for different use cases tied together with support for custom fine tuning stands a good chance of sweeping the market post-collapse.
I'm not on iOS these days, can you elaborate how "Apple's Intelligence thing" has solved the 7B-sized on-portable-device model for everyday users?
My understanding of the zeitgest on HN about Apple Intelligence was definitely not leaning towards "they nailed it". Not even in the ballpark of "promising", I'd say.
I didn't say they nailed it. I don't know since I can't run it, but ten years is a long time for any technology. I toyed with the 7B Mistral through MLC Chat and, while slow, the responses were good.
The Llama-3.2-3B-Instruct it comes with is fast but sometimes takes questioning to get accurate answers.
The older Phi variant it has was thorough and accurate, Phi's selling point, but made my phone run hot being too thorough.
The future, probably within 10 years, is most tasks being handled by small on-portable-device models (7B parameters or so; see Apple's Intelligence thing), a middle ground of workhorse models (pushing closer to 30s and 70s) running on more capable ML-focused chips in laptops and workstations, and home and office servers for the biggest professional users running on dedicated servers.
And then there's the apps. Whoever makes the "Stripe for generative AI" with multiple models with different levels of data provenance, security, SLA, etc for different use cases tied together with support for custom fine tuning stands a good chance of sweeping the market post-collapse.