MythoTech AI News: 6/11/24

Apple rolled out the beginnings of their GenAI strategy yesterday at WWDC. I'm not normally their biggest fan, but I must admit the direction they're going with "Apple Intelligence" is smart. Here's why:

1️⃣ The overall architecture. They're using three tiers of Large Language Models - the first is completely on-device, small (3b parameter), and fine-tuned to be able to execute specific tasks in iOS 18. Basically, it's a better version of Siri that can talk to you, see your screen, understand your phone's data, and take basic actions - all completely offline. The second tier is escalated to for more complex tasks and runs on Apple's private cloud. The third tier is GPT-4o, but it only escalates here with direct verbal approval from the user, and is a special partnership with OpenAI where they will not retain or train on this data. This is a great approach - start with the cheapest model with the strongest privacy protections, and then work your way up towards stronger models only when it is needed.

2️⃣ The privacy strategy. Apple is contrasting themselves to their major competitors by focusing on data privacy that goes beyond just a promise. They're committing to doing more in the open source space to ensure transparency in how they're collecting and using the data, which is commendable. And it's very easy to contrast this to Microsoft's Recall feature for their new Copilot+ PCs which has made lots of headlines in the past week for being a privacy "nightmare". Recall is constantly taking pictures of the desktop for the AI to process, yet according to a security researcher it is storing them unencrypted on the device.

3️⃣ Potentially kill two birds with one stone. The biggest hurdle to LLM adoption for the average person is finding how to integrate it into basic, mundane life. And the biggest hurdle to adoption for virtual assistants like Siri was a lack of capability and consistency. This strategy could POTENTIALLY cross both hurdles, if Apple can get it to consistently solve everyday problems for enough iPhone users.

4️⃣ The up-sell is baked-in. Of course, Apple is primarily a hardware company, and this is a natural win for them on that side as well. This functionality is initially only rolling out to their top-end phones, most likely because those are the only ones with enough processing and RAM to support running a local LLM in the first place. But I'd be willing to bet that the iPhone 16 generation comes with enough compute capacity in every model - and another price hike would not be surprising, of course. And future generations can easily continue making clear improvements to the on-device AI as things continue to get more and more efficient and powerful over time. That might end up changing the overall strategy for iPhone and other iOS hardware if this proves successful!

---

Is Apple now positioning itself as the privacy-focused AI leader? Are you excited for on-device AI or concerned about it? Join the discussion on LinkedIn!