Actually, really liked the Apple Intelligence announcement. It must be a very exciting time at Apple as they layer AI on top of the entire OS. A few of the major themes.
Step 1 Multimodal I/O. Enable text/audio/image/video capability, both read and write. These are the native human APIs, so to speak.
Step 2 Agentic. Allow all parts of the OS and apps to inter-operate via “function calling”; kernel process LLM that can schedule and coordinate work across them given user queries.
Step 3 Frictionless. Fully integrate these features in a highly frictionless, fast, “always on”, and contextual way. No going around copy pasting information, prompt engineering, or etc. Adapt the UI accordingly.
Step 4 Initiative. Don’t perform a task given a prompt, anticipate the prompt, suggest, initiate.
Step 5 Delegation hierarchy. Move as much intelligence as you can on device (Apple Silicon very helpful and well-suited), but allow optional dispatch of work to cloud.
Step 6 Modularity. Allow the OS to access and support an entire and growing ecosystem of LLMs (e.g. ChatGPT announcement).
Step 7 Privacy. <3
We’re quickly heading into a world where you can open up your phone and just say stuff. It talks back and it knows you. And it just works. Super exciting and as a user, quite looking forward to it.
“and it just works”
has he even used an llm before?
He sort of invented it, so you have to think he’s commenting on the concept here, not the implementation.
I have tried a lot of medium and small models, and there it just no good replacement for the larger ones for natural text output. And they won’t run on device.
Still, fine-tuning smaller models can do wonders, so my guess would be that Apple Intelligence is really 20+ small and fine tuned models that kick in based on which action you take.
An LLM has no comprehension of what it says. It’s just a puppy that is really good at performing for treats. This will always yield nonsense a meaningful proportion of the time.
I don’t care how statistically good your model can be under certain constraints and inputs. At the end of the day, all you’ve done is classically condition your computer.
It goes a tad bit beyond classical conditioning… LLM’a provides a much better semantic experience than any previous technology, and is great for relating input to meaningful content. Think of it as an improved search engine that gives you more relevant info / actions / tool-suggestions etc based on where and how you are using it.
Here’s a great article that gives some insight into the knowledge features embedded into a larger model: https://transformer-circuits.pub/2024/scaling-monosemanticity/