Comment by jonplackett
7 months ago
I find this all quite baffling.
I’m pretty sure I could knock up a decent Siri clone with got4o-mini, because I already did a bootleg Alexa to power our smart home stuff after Amazon removed the privacy controls. The only hard bit was the wake word.
Siri is currently so terrible that even something like Mistral 8b could do a decent job. Whey don’t they just run something like that on their own servers instead?
Absolutely lost what is going on. Awni Hannun on Twitter works for Apple on the MLX team and is always completely up to date on all the local LLM models and their capabilities. They literally have a team of people building MLX for doing model inference on Apple Silicon.
Does someone need to send someone an email to realise you don't need a huge frontier model to do basic tool calling like this?
Not only that, but funneling every user query to Sam Altman isn’t exactly on-brand for Apple, also.
Well they have been funneling search to Google for years.
Of course, Google pays them, but it’s still funneling info to a surveillance company vs. providing a private search.
If Kagi can do it, Apple could.
They are competing against Gemini on Android, so it stands to reason that they need something on par with that. Per TFA they are still talking about running those models on Apple's own servers.
The more interesting question is how they're going to handle the 180 on all the talk about privacy.
Claude can be used via amazon bedrock technically on your own managed AWS infra...
But yeah you're not trusting Anthropic but Apple + Amazon
I dunno if thats even a win?
If the model runs on apple servers then the data isn't leaking anywhere. There's no 180 to discuss.
There's a big difference between running on a local device and running on Apple's servers, and their previous stance was that most things would be in the former category. Switching to cloud (even if it's Apple's cloud) for regular Siri stuff would be a big 180.
> They are competing against Gemini on Android, so it stands to reason that they need something on par with that
Why? What does Gemini actually do, that users actually use, that requires deep integration into the OS?
Things like notification summaries that aren't hallucinated, for example.
They could secretly make Siri slightly better in intervals. People have a low opinion of Siri anyways. No way they compare Siri to Gemini. For them, Siri might just stop sucking completely at some point and then the comparison is between the Siri of the last several weeks and the old Siri.
> Siri might just stop sucking completely at some point and then the comparison is between the Siri of the last several weeks and the old Siri.
The same thing did happen to Apple Maps, but many people still default to google (though google maps is still significantly better at finding businesses). But Apple was humiliated by the Apple Maps rollout. Siri has just been a slow-burning joke that's only really useful for setting a timer or reminder.
There's very low margin of error for a user trying something over speech (where they don't know what's actually capable). A user tends to try something once, if it fails they nearly never try it again. So now the question is, how do you get a user to try it again when you've fixed it? Alexa's approach has just about driven everyone mad "By the way, did you know..."
while everyone agrees that Siri is crap saying you can knock out a decent clone yourself completely underestimate the complexities involved