Thanks! It’s a good read and I like the idea of a private cloud compute (PCC) system, but that doesn’t mention anywhere that ChatGPT will be running in that PCC system (if you were trying to imply that).
And while OpenAI could implement something similar to PCC, I haven’t seen them announce that anywhere either.
I’d say the proof is on Apple to show that it’s being done on-device or that all processing is done on iCloud servers.
You’re saying that OpenAI is just going to hand over their full ChatGPT model for Apple to set up on their own servers for free?
But from the article itself:
the partnership could burn extra money for OpenAI, because it pays Microsoft to host ChatGPT’s capabilities on its Azure cloud
I get it if they created a small version of their LLM to run locally, but I would expect Apple to pay a price even for that.
I think you may be confusing this ChatGPT integration with Apple’s own LLM that they’re working on… Again, from the linked article:
Still, Apple’s choice of ChatGPT as Apple’s first external AI integration has led to widespread misunderstanding, especially since Apple buried the lede about its own in-house LLM technology that powers its new “Apple Intelligence” platform.
What? No. I would rather use my own local LLM where the data never leaves my device. And if I had to submit anything to ChatGPT I would want it anonymized as much as possible.
Is Apple doing the right thing? Hard to say, any answer here will just be an opinion. There are pros and cons to this decision and that’s up to the end user to decide if the benefits of using ChatGPT are worth the cost of their data. I can see some useful use cases for this tech, and I don’t blame Apple for wanting to strike while the iron is hot.
There’s not much you can really do to strip out identifying data from prompts/requests made to ChatGPT. Any anonymization of that part of the data is on OpenAI to handle.
Apple can obfuscate which user is asking for what as well as specific location data, but if I’m using the LLM and I tell it to write up a report while including my full name in my prompt/request… that’s all going directly into OpenAIs servers and logs which they can eventually use to help refine/retrain their model at some point.
I’m sure you understand this, but anonymized data doesn’t mean it can’t be deanonymized. Given the right kind of data, or enough context they can figure out who you are fairly quickly.
Ex: You could “Anonymize” gps traces, but it would still show the house you live at and where you work unless you strip out a lot of the info.
http://androidpolice.com/strava-heatmaps-location-identity-doxxing-problem/
Now with LLMs, sure, you could “anonymize” which user said or asked for what… but if something identifying is sent in the request itself, it won’t be hard to deanonymize that data.
I think you meant to say:
“China has enough air pollution that the smog from there blows over to Los Angeles”