Large language models—whether or not you believe they’re capable of true intelligence—will likely prove to be meaningfully helpful in both work and personal settings, once they become better trained and more predictable. Recent announcements from Salesforce and Microsoft have shown how AI models, trained both on huge piles of text and proprietary company data, can reduce busywork and increase productivity in work settings. It follows that large language models, given access to the right data, could be similarly useful in personal computing.
For example, generative AI could put a foundation of general knowledge underneath the more specific informational and task-oriented data that voice assistants like Apple’s Siri typically call up. It might give Siri a basic knowledge of how the world works, as seen in ChatGPT, and therefore a better framework for understanding and helping users. A better understanding of the underlying meaning behind a user request to lead to more useful, on-point responses.
Indeed, it may not be long before Apple is forced to reckon with users who are wondering why ChatGPT seems so much smarter than Siri.
PULLED INTO AN ARMS RACE
Apple has stayed out of the generative AI discussion so far, mainly because the technology is not seen as directly disruptive to its core businesses (hardware)—at least not in the way that generative AI could be disruptive to Google’s core Search business, or to Microsoft’s core productivity apps business. Also, Apple is known as a “fast follower”; it likes to wait until new technologies have matured, then jump in with its own Apple-flavored version. So far, the company has treated AI as an enabling technology that it deploys behind the scenes to make its devices and apps work better (the image selection and editing features in its Photos app, for example).
But those old habits may not work with generative AI. The technology, for better or worse, may be the Next Big Thing, with transformative power on the order of social media or mobile computing. It may be too important to keep on ice in an R&D lab or to cast in a behind-the-scenes role.
Generative AI has already caught the industry’s—and the public’s—attention starting with the public release of OpenAI’s ChatGPT. Google was on a path to gradually work generative AI into its existing services, but as the public (including investors) came to understand the potentially transformative power of the generative AI, the company moved quickly to deploy the tech. Google’s first generative AI product was a public-facing chatbot called Bard; it made no announcement at all when, many months ago, it quietly began using generative AI to analyze the intent of search queries.
Apple may be subject to some of these same pressures. It may be forced into the generative AI arms race simply because not showing up could paint a picture of a company that is slow to recognize game-changing ideas and weak in R&D.
Apple has responded to similar pressures before. The HomePod, which represented Apple’s entrance into the smart speaker market, has not been a hugely successful product by Apple’s standards, but the company felt that ambient voice assistant technology (à la Alexa) was relevant to the future of computing, so it entered the smart speaker category.
Ambient voice computing hasn’t been exactly transformative—mainly because the Alexas and Siris of the world are simply not smart enough. Large language models may usher in the second major phase of voice assistants, and Apple may be wise to do whatever it can short of a complete rebrand to make sure the world knows that Siri now leverages the cutting edge in generative AI.
But adding generative AI to its assistant may be harder for Apple than for its big tech peers. The company’s dogged focus on privacy has been an effective cudgel against Meta and Google in the social networking and advertising realms, but it’s a serious impediment to building an assistant that harnesses the goodness of generative AI. In order to transform Siri—to make it smarter and more human-feeling—via a large language model, Apple would have to open up access to a broad swath of public information (from the web) and, more importantly, personal user information: their communication style, plans, priorities, preferences, health stats, tastes, and relationships.
Apple has so far been willing to train models on this type of data only if the model can run solely on the user’s device (versus in the cloud) and if the data remains invisible to Apple itself. Large language models require a lot of computing power to run, and while major efforts are underway to make the models smaller and more efficient, that work may not have progressed to the point where such a model can run on an iPhone.
Apple is very likely working to find a way to deploy its own flavor of generative AI, and there’s evidence that the work has intensified since the arrival of ChatGPT late last year.
Even though the company has slowed hiring, it’s currently advertising almost 100 jobs under the heading of “Machine Learning and AI: Natural Language Processing and Speech Technologies.”
Nineteen of those roles are located in Seattle, where Apple is known to have centered a good part of its AI R&D. Seven of these roles contain the word “Siri” in their titles, including the three most recent listings from late February.
One of the open listings calls for a candidate with experience in “NLP, personalization, recommendation, federated learning [a way of training models while anonymizing training data], and model compression [making models that require less compute power], that help power products including: Siri, Spotlight, App Store, Apple Music, and much more at Apple.”
Thirty-six of the NLP job postings are for “annotation analysts” in places such as Beijing, Barcelona, Cork (Ireland), and Singapore. These jobs entail analyzing the Siri data of current users (in various languages) who have “opted in” to sharing their interactions with Apple. Apple has analyzed such data for years, but the same skill set could be applied to labeling training data for large language models.