Loading player
Last September the technology news site The Information revealed that Apple was investing “millions of dollars a day” in artificial intelligence with the aim of accelerating the development of chatbots (like ChatGPT) and other tools for generating (or editing) images and videos. Apple has created a specific internal division which includes sixteen people, including several former Google employees, led by John Giannandrea, also formerly of Google, who in 2018 was hired by Apple to improve Siri, the voice assistant of the company.
Investments aside, however, Apple does not seem to have adopted a precise strategy on AI, while most of the large technology companies have been moving for some time: Microsoft has a close alliance with OpenAI; Meta is inserting AI services into many of its products; Amazon is investing “deeply in the sector,” according to its CEO Andy Jassy; and Google has long been considered the most cutting-edge company in the industry (despite being surprised by the success of ChatGPT).
In March Bloomberg revealed that Apple would have “open negotiations” to use Gemini – Google’s artificial intelligence – in the next iPhones, and would also consider using ChatGPT. A similar decision would be new for Apple, which has always preferred to develop its own software and build closed and well-protected environments. In this regard, the company is developing its own large language model (Large Language Model, or LLM), called Ajax GPT. Ajax was created for internal use at Apple and is still under development: always second Bloombergis more powerful than GPT-3.5, the language model on which the initial version of ChatGPT was based (OpenAI has since released GPT-4, considered a major advance over GPT-3.5).
Apple’s AI strategy will likely be clarified during WWDC, its annual developer conference next June. Tim Cook himself, the company’s CEO, recently said that he considers generative AI “a fundamental opportunity for our products” and assured that Apple will announce more “in the coming weeks”.
– Read also: The enthusiasm for artificial intelligence is waning a bit
Apple’s perceived lag in AI adoption is partly related to the company’s emphasis on protecting its users’ data, particularly in contrast to Android, which Apple has called “a huge tracking device.” Apple has even banned its employees from using ChatGPT (and Microsoft’s Copilot service) and a similar decision has also been taken by Verizon, Samsung, Bank of America, Deutsche Bank and Goldman Sachs, among others, because the conditions of use of ChatGPT clearly state that prompts, i.e. user requests, can be used to “train” OpenAI’s linguistic models (unless the user refuses this point by opting out). The risk is that employees of these companies discuss work issues with the chatbot, revealing company secrets to OpenAI.
The attention to privacy dear to Apple, however, does not go well with the functioning of generative artificial intelligence, which, at least at this point in their technological development, requires that each request be processed remotely, in a server of some company, and not inside the starting device. This involves transferring a lot of the user’s personal data from the device to an external server, most likely operated by a company other than Apple, and increases the risk to user privacy, a point that Apple has insisted on for years.
All the generative AI mentioned so far are in fact based on large linguistic models, which, as their name suggests, are very “heavy” and require considerable infrastructure to function. The development of generative AI, in fact, has highlighted the importance of technological infrastructures in the production of chatbots and similar services. This is especially true for GPUs (graphics processing units), a type of processor widely used in training language models; among all the GPUs produced by Nvidia, which in the last year has become one of the most discussed companies in the world.
– Read also: Nvidia’s formidable stock market success
According to many analysts, Apple therefore wants to focus on smaller language models, capable of functioning completely within a device. In this regard, last March it acquired Darwin AI, a Canadian startup that develops technologies to make AI “smaller and faster”. In April it then presented OpenELM (“Open-source Efficient Language Models”), a group of eight small open source language models, designed to work entirely within a device. They are experimental models that are mostly used by Apple to understand whether the concept behind these models is viable, and in what way. In light of these initiatives, the negotiations with OpenAI and Google appear to be temporary remedies, while waiting for the small models to be powerful and reliable enough.
To achieve this, Apple is investing heavily in research, as demonstrated by the many patents registered by the company, which the site The Verge analyzed. In one of these, an attempt is made to solve the problem by using SSD memories instead of RAM ones to contain the information relating to a linguistic model: in this way, the models run faster and more efficiently, because they «use the storage space more cheap and available on the device.” Apple researchers also created a system, called EELBERT, that can compress a large language model into a much smaller space without excessive performance losses. Google’s BERT language model, for example, has been shrunk 15 times with a qualitative reduction of 4% (but with some latency problems, a term used to indicate the time taken by data to transfer across a network, and therefore the delay in communication).
Not all AI is made for conversation, though. In recent years, smartphone models such as the Samsung Galaxy S20 or the Pixel 8 and 8 Pro have stood out for their ability to modify photographs even radically, cutting out or removing elements or changing colors. One of the products developed by Apple works just like an image generator (similar to DALL-E or Midjourney) where users can edit the content produced without needing to write any other prompts. One of the papers produced by the company also describes the possibility of using written prompts to modify photographs taken by users, asking the AI to “change the color of the sky” or “make my face less strange”, for example. Apple also seems intent on applying AI to the Health app – to analyze the many biometric data of users it has available, also thanks to the Apple Watch – and Apple Music, developing a system to separate the various tracks that make up a song.
– Read also: Is this photo a real photo?
However, the center of Apple’s AI activities remains Siri, thanks also to innovations such as STEER (Semantic Turn Extension-Expansion Recognition), a system that helps the voice assistant manage series of questions, recognizing when the user is asking for clarification on a question already asked or when asking a new one. Speaking of Siri, Apple was the first major company to launch a voice assistant, in February 2010, although initial expectations have not been met since then. Siri was one of the last projects handled by Steve Jobs, co-founder and CEO of Apple until 2011, shortly before his death the same year, but his progress was immediately disappointing. Already in 2012, about two years after the launch of Siri, some Apple employees complained about the state of the product, claiming that Jobs “would go crazy” if he had seen the limited progress made by the voice assistant.
This could therefore be the perfect opportunity to perfect and enhance the service, equipping it with linguistic models capable of generating much more complete and sophisticated responses. After all, Siri was born as a startup and was acquired by Apple in 2010, at the behest of Jobs. In an interview at the time, journalist Walt Mossberg asked Jobs about Siri, which he described as a web search service. “No, it’s not in the research sector,” Jobs corrected him, “it’s in the field of artificial intelligence.”
