A white paper on Apple’s Foundation Model, the company’s homegrown LLM (large-language model) that powers Apple Intelligence, reveals two important facts: it’s the safest in design and highly competitive with both Meta’s Llama and OpenAI’s GPT-4. This seems to debunk a big myth about Apple’s AI efforts: that the company’s privacy-first philosophy would hold it back.
The Apple Foundation Model is just as capable in tests of writing and summarization compared to the top LLMs by OpenAI, Meta, Mistral AI and others. And thanks to Apple’s strict guidelines for expunging harmful content, human-evaluated tests repeatedly rank its foundation model as the safest above all the rest — by a wide margin.
This is interesting. “Ahead” seems difficult to prove, but time will tell how good Apple’s AI is.
I’m wary of these kinds of benchmarks, because they can be easily manipulated to make the product they want come out on top.
Writing and summarization alone are pretty limited compared to what generative AI is capable of. It sounds like these functions don’t include research and analysis.
Claude and ChatCPT can do research and analysis at an already high level, although often with inaccuracy that must be confirmed manually. As one example, you can give them a set of medical symptoms in detail, and tell them to provide every potential disease or syndrome that could be causing those symptoms and to explain the mechanism of each. The results will be thorough, and can include conditions that treating physicians have not considered. This provides possibilities for a physician to then rule out. Summarizing emails and providing responses that are accurate are still valuable, but aren’t comparable to what else the big LLMs are capable of.
(One exception: don’t use generative AI to do legal research, the results are completely unreliable.)
Siri knowing how apps work, and a user being able to ask Siri to do what otherwise takes many presses and swipes–that will be very nice.
