Apple’s cautious but clever AI use
Would you like ChatGPT with that? (Apple)
The big announcement at Apple’s WWDC this week may have seemed underwhelming at first glance, but is a very smart move the closer you look: Yes, we’re finally getting a calculator on the iPad a decade after the device was first launched.
The company is also cleverly avoiding most of the issues with hallucinations, privacy and over-promising and under-delivering on the hype that plague most AI products (see Adobe, Microsoft and Google’s recent AI debacles).
Instead, Apple is using compressed, on-device, open-source derived models, which cleverly hot swap adapters when required that have been fine-tuned to specialize in one particular task, whether it is summarization, proofreading or auto-replies.
The idea is to increase the chance that most AI tasks can be completed successfully and privately on the device itself — and, of course, to finally provide a compelling reason to upgrade your phone or iPad.
More difficult queries are sent using anonymized, encrypted data to a medium-sized model on Apple’s servers (which doesn’t store the data) and the most complex tasks involving writing or synthetic reasoning are sent on to ChatGPT after you’ve given it permission. OpenAI can’t store your data either.
Based on the presentations, the privacy and functionality aspects seem very well thought out, although we won’t find out for sure until September. Apple didn’t build the first smartphone, but it came up with one of the best versions of it with the iPhone. It will be interesting to see if its cautiously optimistic approach to AI enjoys similar success.
Apple can do some things well but isn’t promising the earth. (Apple)
Google AI is still stuck on glue
Like a snake eating its own tail, it seems that all of those articles about how stupid Google’s AI Overview answers were have just been making the answers worse. You may…
..
