So yesterday Apple announced their long awaited platform-agnostic AI strategy for all their consumer products. They already did that 2 years ago but did not follow through. This time, they might.
CapabilitiesCapabilities
The short version is that new Siri AI features broadly fall in 2 categories imo:
- Chat with context:
In all kinds of places you will be able to chat with Siri as if it was ChatGPT. Consumers are familiar with it. And your phone has much much more context about you than chatGPT-memories or even Google Gemini, which also has a lot of context for consumers who gave all of their information about themselves for free
- Agentic
This one surprised me. Apple basically lets you vibecode shortcuts via its AI tools. https://support.apple.com/guide/iphone/use-apple-intelligence-in-shortcuts-iph78c41eaf8/ios This is almost like letting Opencode, codex, OpenClaw, Claude Code and the like just run CLI commands on your PC. I did not seem them going this far so early. And it's not even only shortcuts, the normal chat interface gets interactions that are not read-only.
Selfhosted Privacy-first AlternativeSelfhosted Privacy-first Alternative
Can my private on-prem setup compete with all this context Siri or Gemini have about our lives? Setting aside, if this is even desirable. I think yes, but it takes a lot of effort.
- You can easily deploy your own Ollama or llama.cpp or whatnot if you have the expensive electronics. Easy
- You can easily chat with this LLM using Open Webui, Libre Chat which are easy to self host. Also Opencode can be even more powerful considering its built-in CLI tools.
- Now comes the hard part. You'd have to build all of these tools. Granted, writing MCP tools isn't hard. But keeping them updated and connecting to all of your equivalent selfhosted tools is. You basically need to have your email, your journal, immich, texts already be accessible via an open API or interface and then write tools for them additionally.
IMO this is all very important for businesses and organizations and institutions. But kinda too much to ask for for private people in their freetime.
The SelloffThe Selloff
Apple Stock sold off yesterday evening and today. Hard to say how much, the whole market did too. Idk what people expected. I know some people wanted the MACs with huge cheap unified memories to come for server racks as true B2B electronics. Ofc this was not going to happen. It's a B2C company.
EU wants Apple to let users select their preferred model rather than restricting them to Apple Foundation Models.
Of course this doesn't benefit Apple, at all, but it can benefit users with local LLMs. Their opinion on EU's requests is that they "know better" when it comes to users' privacy, and would never allow a 3rd party to access them (OpenAI, X).
Completely glossing over the fact that local LLMs exist.
Is that really a feature though? Why do you need your personal email and texts exposed to an LLM (or a third party at all)? What's the benefit?
Some people might not be able to take advantage of that. I have all kinds of information in my Emails, texts, journals and photos. Maybe you do not care and that's okay.
I'm having BBQ with the neighbors this evening - Hey Siri did the neighbor go to MIT or to Berkley, I think I journaled it down last bbq summer. Planing a vacation to Italy - Hey Siri was my last trip to Rome in 2016 or 2018, I took photos for sure. I bought this exact same pair of shoes last year - hey Siri how much did it cost back then, I know it's somewhere in my emails.
To spin this a little further - I really like the idea of my home and my family having an AI that is deeply ingrained in our family culture. Everything selfhosted and everything peculiar to us. With a deep library of knowledge. All private, selfhosted, on-prem everything.
Hold up. You take notes about private information you get from your neighbours over bbq? Do they know you're a spook? (I'm only half pulling your leg here; this is why I never tell anyone anything, because people cannot be trusted.)
This functional part I like (it's about the only functional part I can possibly like though, haha.) I made a datalake for this. It is like my personal knowledge base, indexed and with embeddings. I also should disclose that I rarely use it, but that's because I'm too busy to adapt it wider.
What it does not contain: private conversations, financials, sensitive stuff like PII, contacts and so on. Everything (=knowledge) could in theory be ingested, but I have thus far only fed it research and notes on that research, and all security review work I do regarding 3rd party apps on my phones (because that's more useful than grepping markdown.)
This technical part I like. It's the only way.