The Gist
- Anticipation fuels AI. Apple’s AI advancements at WWDC drive market excitement and stock value increases.
- AI myths debunked. Despite rapid advancements, human-level AI remains a distant reality, not as imminent as perceived.
- Apple’s AI independence. Apple develops in-house language models, emphasizing capability over collaboration with OpenAI.
The scene at Apple’s WWDC (Worldwide Developers Conference) last week was, in a way, emblematic of the times. The tech giant’s AI announcements were massively hyped ahead of the show. The products themselves were interesting, but not the “next iPhone” some expected.
Still, the market loved it, adding hundreds of billions to Apple’s market cap in just a few days.
A good part of this AI moment is built on anticipation: There’s a belief that models will keep improving, products will get better, and people and companies will buy in. We don’t know for sure where it’s all leading, but it’s heading somewhere, and that seems to be good enough. The demos should work eventually.
I spent the week in Silicon Valley visiting sources and tech companies — starting at Apple and ending at NVIDIA — to get a sense of where we are on the continuum, who’s poised to lead and how power is shifting. Much of what I learned will land in future stories and Big Technology Podcast episodes. So stay tuned. But here’s what stands out in my notebook:
Room for AI Models to Improve
It doesn’t appear generative AI will hit the resource wall anytime soon — at least according to those closest to the work. AI research houses are focusing on constraints like compute, data and energy. But they also realize there’s room to improve the current set of models by getting better at selecting the right data, fine-tuning the models and building new capabilities like reasoning. Meanwhile, incoming compute improvements should lead to more powerful and efficient training and inference.
The next 18 months will be interesting.
Related Article: Tim Cook’s AI Moment: A Pivotal Shift at Apple
Expectations Might Still Be Unrealistic
Still, the popular conversation around AI tends to portray human-level artificial intelligence as right around the corner. It’s not. The next generation of models will be impressive, but the release of ChatGPT (launched on a version of OpenAI’s GPT-3) followed soon after by the release of GPT-4 made the pace of AI development seem, to many, faster than it is. Training and fine-tuning these models takes a long time.
So, while GPT-5 and its peers will be hyped and hotly anticipated, the push toward reasoning and AI agents may be more tangible in the short term vs. sheer model size.
Related Article: WWDC 2024: Apple Intelligence, RCS Messaging — and More
OpenAI Might Be a Placeholder in Apple Intelligence
Sam Altman is a master dealmaker, but what if he’s just keeping the seat warm within the new “AI iPhone” for Google? Recently, Bloomberg’s Mark Gurman reported that Apple is not paying OpenAI for use of ChatGPT in its next generation of iPhones. Apple has also been negotiating with Google for 4-5 months for a similar placement, he told me.
The key question then becomes, who gets the default position? Google pays Apple $20 billion per year to be its products’ default search engine and, if it can figure out the economics, a similar (or smaller) deal may supplant ChatGPT with Gemini as Apple’s AI default in time. (Gurman talks more about this on Big Technology Podcast recently).
Related Article: Apple’s AI Tango: Graceful Steps at WWDC 2024
NVIDIA’s Key Ratio
I spent a wild day inside NVIDIA on recently, speaking with company leaders from morning till evening about the technology powering this moment. There will be plenty more to come on NVIDIA in the next few months here, but here’s one fun fact: NVIDIA has more software engineers than hardware engineers. There’s so much more to NVIDIA’s dominance than chips, starting with the fact that its software is core to training AI models, and the company’s headcount reflects it.
Apple Tries Small Language Models
The real surprise at WWDC was that Apple used many of its own models to power Apple Intelligence, and not OpenAI’s. In fact, ChatGPT was mostly a plugin in the company’s demos.
To make its AI experience work, Apple built a series of small language models that reside on-device. These more focused, less compute-intensive AI models are good at specific tasks like proofreading and run in concert with each other (The Verge has a good write-up).
To those watching, this demonstrated A) Apple can indeed make real progress on AI model development and, B) These smaller models can be useful to bring big ideas to life, even for the largest companies. Now, we’ll wait to see if those demos work in real life.
