Hi guys and welcome to TechCrunch’s regular AI newsletter.
This week, Apple was within the highlight within the AI sector.
At the Worldwide Developers Conference (WWDC) in Cupertino, Apple unveiled Apple Intelligence, its long-awaited, ecosystem-wide foray into generative AI. Apple Intelligence enables an entire host of features, from an improved Siri to AI-generated emojis to photo editing tools that remove unwanted people and objects from photos.
The company promised that Apple Intelligence can be developed with security at its core and enable highly personalized experiences.
“It needs to know you and be anchored in your personal context, like your every day routine, your relationships, your communications and more,” CEO Tim Cook noted during Monday's keynote. “All of this goes beyond artificial intelligence. It's personal intelligence and the subsequent big step for Apple.”
Apple Intelligence is typical Apple: It hides the technical details behind obvious, intuitively useful features. (Cook never once uttered the phrase “big language model.”) But as someone who writes concerning the dark side of AI for a living, I wish Apple can be more transparent—just this once—about how the sausage is made.
Take Apple's model training practices, for instance. Apple disclosed in a blog post that it trains the AI models that power Apple Intelligence using a mixture of licensed datasets and the general public web. Publishers have the choice to opt out of future training. But what in the event you're an artist who desires to know in case your work was included in Apple's initial training? Tough luck—there's no word on that.
The secrecy might be for competitive reasons. But I believe it's also to guard Apple from legal challenges—particularly challenges related to copyright. Courts have yet to make your mind up whether vendors like Apple have the best to coach on public data without compensating or naming the creators of that data—in other words, whether the fair use doctrine applies to generative AI.
It's just a little disappointing to see Apple, which frequently presents itself as a champion of sensible technology policy, implicitly support the fair use argument. Behind the veil of selling, Apple can claim to pursue a responsible and measured approach to AI, while it might have trained on developers' works without permission.
Somewhat explanation would go a good distance. It's a shame we didn't get one – and I'm not confident we'll get one any time soon unless a lawsuit (or two) comes along.
News
Apple's key AI features: Your humble servant has summarized the important thing AI features Apple announced this week through the WWDC keynote, from improved Siri to deep integrations with OpenAI's ChatGPT.
OpenAI is hiring executives: OpenAI this week hired Sarah Friar, the previous CEO of hyperlocal social network Nextdoor, as chief financial officer and Kevin Weil, who previously led product development at Instagram and Twitter, as chief product officer.
Mail, now with more AI: This week, Yahoo (TechCrunch's parent company) updated Yahoo Mail with recent AI features, including AI-generated email summaries. Google recently launched an identical generative summarization feature – nevertheless it's behind a paywall.
Controversial views: According to a recent study from Carnegie Mellon University, not all generative AI models are the identical – especially in the case of coping with polarizing topics.
Sound generator: Stability AI, the startup behind AI-powered art generator Stable Diffusion, has released an open AI model for generating sounds and songs that it says was trained exclusively on royalty-free recordings.
Research paper of the week
Google believes it could possibly create a generative AI model for private health – or no less than take initial steps in that direction.
In a brand new paper featured within the official Google AI blogResearchers at Google are lifting the curtain on the Personal Health Large Language Model, or PH-LLM for brief – a refined version of one among Google's Gemini models. PH-LLM is designed to supply recommendations for improving sleep and fitness, including by reading heart and respiratory rate data from wearable devices resembling smartwatches.
To test PH-LLM's ability to supply useful health advice, researchers conducted nearly 900 sleep and fitness case studies with subjects across the United States. They found that PH-LLM provided sleep recommendations that weren’t quite pretty much as good because the recommendations made by human sleep experts, but still not quite pretty much as good.
The researchers say PH-LLM could help contextualize physiological data for “personal health applications.” Google Fit springs to mind; I wouldn't be surprised if PH-LLM eventually powers a brand new feature in a fitness-focused Google app, whether Fit or something else.
Model of the week
Apple has written quite a bit on its blog concerning the recent on-device and cloud-based generative AI models that make up the Apple Intelligence Suite. However, despite the length of this post, it reveals little or no concerning the capabilities of the models. Here is our attempt to investigate it:
The unnamed on-device model that Apple is highlighting is small, little doubt so it could possibly run offline on Apple devices just like the iPhone 15 Pro and Pro Max. It comprises 3 billion parameters — “parameters” are the parts of the model that essentially define its ability to do an issue, like generating text — making it comparable to Google's on-device Gemini model, Gemini Nano, which is available in 1.8 billion parameter and three.25 billion parameter sizes.
The server model, meanwhile, is larger (how much larger, Apple isn't saying exactly). What we do know is that it's more powerful than the on-device model. While the on-device model performs on par with models like Microsoft's Phi-3 mini, Mistral's Mistral 7B, and Google's Gemma 7B within the benchmarks listed by Apple, the server model “compares well with OpenAI's older flagship model, GPT-3.5 Turbo,” in response to Apple.
Apple also says that each the on-device model and the server model are less vulnerable to getting out of hand (i.e., spewing toxicity) than similarly sized models. That could also be so—but this writer reserves judgment until we’ve a probability to place Apple Intelligence to the test.
Grab bag
This week marked the sixth anniversary of the discharge of GPT-1, the predecessor of GPT-4o, OpenAI’s latest flagship model for generative AI. And while Deep learning could also be reaching its limitsit's incredible how far the sector has come.
Consider that it took a month to coach GPT-1 on a dataset of 4.5 gigabytes of text (BookCorpus, which comprises about 7,000 unpublished novels). GPT-3, which is sort of 1,500 times larger than GPT-1 by parameter count and significantly more sophisticated within the prose it could possibly generate and analyze, took 34 days to coach. How does that scale?
GPT-1's training approach was groundbreaking. Previous techniques relied on huge amounts of manually labeled data, which limited their usefulness. (Manually labeling data is time-consuming—and tedious.) But that wasn't the case with GPT-1; this system trained totally on data to “learn” the way to perform a series of tasks (resembling writing essays).
Many experts consider that we’ll not see a paradigm shift as significant as that of GPT-1 within the near future. But on the other hand, the world didn’t see GPT-1 coming either.
