Apple has not discussed AI extensively, but it has been actively developing many projects.
One might assume Apple is behind in AI. Since late 2022, when ChatGPT became very popular, many of Apple’s rivals have rushed to keep up.
Although Apple has mentioned AI and launched some AI-based products, it appears they are just testing the waters instead of fully committing.
Recently, rumors and reports indicate that Apple has actually been waiting for the right moment to act. In the past few weeks, it’s been said that Apple is discussing with both OpenAI and Google to enhance its AI features. Additionally, the company is developing its own AI model, named Ajax.
When you examine Apple’s published AI research, you start to see clues about their AI strategy. Clearly, predicting products from research papers is not very precise — the path from research to the market is indirect and challenging.
However, you can gain an understanding of what the company is considering — and how its AI features might function when Apple discusses them at its annual developer conference, WWDC, in June.
Smaller, more efficient models
It seems we both want the same thing: an improved Siri. And it appears that a better Siri is on the way! Many believe, both within Apple’s research and globally in the tech industry, that big language models will quickly enhance virtual assistants, making them smarter.
For Apple, achieving a better Siri involves speeding up these models and ensuring they are widely available.
In iOS 18, Apple intends to operate all its AI features on a model that works entirely offline on the device, according to a recent Bloomberg report.
Creating a versatile model is challenging even with a network of data centers and thousands of advanced GPUs — it’s significantly more difficult using only the components within a smartphone. Therefore, Apple needs to be innovative.
In a paper titled “LLM in a flash: Efficient Large Language Model Inference with Limited Memory” (which, despite its dull title, is quite fascinating), researchers developed a method to store a model’s data on the SSD instead of the usual RAM.
The researchers reported that they could run language models up to twice the size of the available DRAM on the SSD, which sped up the processing times by 4-5 times on CPU and 20-25 times on GPU. They discovered that using the most cost-effective and readily available storage on devices allows the models to operate more quickly and efficiently.
Apple’s researchers have developed a system called EELBERT that significantly reduces the size of a large language model without greatly affecting its performance.
Their version of Google’s Bert model was compressed to just 1.2 megabytes, making it 15 times smaller, with only a 4 percent drop in quality. However, this compression did result in some delays.
Overall, Apple is addressing a fundamental challenge in the model world: larger models are more effective and useful, but they also become more cumbersome, consume more power, and operate more slowly.
Like many others, Apple is seeking the ideal balance among these factors while also attempting to achieve the best possible outcome.
Siri, but better
When discussing AI products, we often focus on virtual assistants — tools that can provide information, remind us of tasks, answer questions, and handle tasks for us. Therefore, it’s not surprising that much of Apple’s AI research centers around one main idea: what if Siri was exceptionally good?
A team of researchers at Apple has been exploring a method to activate Siri without a wake word; rather than waiting to hear “Hey Siri” or “Siri,” the device could potentially recognize when you are speaking to it.
The researchers have admitted that this task is much harder than simply detecting a wake word because there might not be a specific phrase to signal the start of a command.
This challenge might explain why another research team has developed a technology to better recognize wake words. Additionally, another study focused on training a model to understand uncommon words, which virtual assistants often struggle to recognize.
In both scenarios, the benefit of using a large language model (LLM) is its ability to process a lot of information quickly.
For example, in the wake-word study, the researchers discovered that allowing the model to analyze all incoming sounds, rather than filtering out what might seem unnecessary, resulted in a more reliable detection of the wake word.
Once Siri recognizes your voice, Apple is also focusing on enhancing how it understands and responds. In one study, they developed a system called STEER (short for Semantic Turn Extension-Expansion Recognition) which helps the assistant determine whether you’re continuing a conversation or starting a new topic.
Another study uses LLMs to better interpret “ambiguous queries” so that Siri can understand your intent regardless of how you phrase your words.
They noted that in uncertain situations, smart conversational agents should proactively ask questions to clarify their understanding, thereby solving issues more effectively. Another research effort aims to make assistants provide responses that are concise and easier to understand.
AI in health, image editors, and your Memojis
When Apple discusses AI publicly, it often emphasizes its practical applications rather than its technical power.
Although there is a strong emphasis on improving Siri, especially as Apple aims to rival devices such as the Humane AI Pin, the Rabbit R1, and Google’s integration of Gemini across Android, Apple also sees various other practical uses for AI in everyday life.
One key area Apple is exploring is health: theoretically, large language models (LLMs) could help sift through vast amounts of biometric data from your devices to provide clearer insights. Apple has been studying ways to gather and organize your movement data, use gait recognition and your headphones for identification, and analyze your heart rate information.
They have also developed and released “the largest multi-device multi-location sensor-based human activity dataset” after collecting data from 50 participants using multiple sensors on their bodies.
Apple also views AI as a tool for creativity. In one study, after interviewing several animators, designers, and engineers, they developed a system called Keyframer. This system allows users to start with a design prompt and then use tools to modify and perfect the design interactively. This type of creative interaction could potentially be incorporated into everything from the Memoji creator to Apple’s professional design software.
Over time, it’s likely that Apple will emphasize these features, especially on iOS. Some features will be integrated into Apple’s own apps, while others will be made available to outside developers through APIs. (The recent Journaling Suggestions feature is a good example of how this might be implemented.)
Apple has long boasted about its superior hardware compared to typical Android devices; combining this with on-device, privacy-centered AI could set them apart significantly.
We might be getting ahead of ourselves, but imagine how this could work with other innovations Apple is developing.
A Siri that understands your needs, combined with a device that comprehends everything on your screen, could result in a phone that essentially operates itself. Apple wouldn’t need extensive integrations; it could simply manage apps and select the appropriate options automatically.
All of this is still just research, and making it all function smoothly by this spring would be an unprecedented technical feat. (You’ve used chatbots — you know they aren’t perfect.) However, I’m confident we’ll hear some significant AI announcements at WWDC.
Apple’s CEO Tim Cook hinted at this in February and almost confirmed it during this week’s earnings call. Two things are very clear: Apple is actively competing in the AI space, and this could lead to a complete transformation of the iPhone. You might even find yourself using Siri by choice, which would indeed be a remarkable achievement.
What we think?
I think Apple’s new AI research is exciting. They’re working to make Siri better and adding new AI features to iPhones.
They focus on privacy, so all AI runs on your phone, not in the cloud. This means Siri could work faster and understand you better. I can’t wait to see what Apple announces at WWDC. I hope these changes make Siri smarter and more helpful.