Rumours continue to swirl that Apple is working on something dubbed AppleGPT. Whether this is a real product or service or something that will be folded into a super-powered Siri or perhaps melded into lots of apps and services remains to be seen. Clearly, artificial intelligence is a big deal but will it make a real, lasting change to how we use technology or will it become next year’s blockchain – useful but not the world-changing tech many have hyped it to be?
Tech research company has an interesting model it uses to describe the uptake of new technologies. Gartner’s Hype Cycle suggests there are early adopters who jumping new technologies and hype them up. But the enthusiasm of those early adopters gives rise to a ‘trough of disillusionment’ where the early promises are not met. If the technology can overcome the trough, there’s a steady uptake of the technology until it reaches mainstream adoption.
We are yet to hit the trough with AI, or more specifically generative AI, but there are signs we are going that way. There is some fear, uncertainty and doubt that AI will be a boon to humanity. And governments are looking at how to regulate AI, or more specifically the outcomes of the use of AI, to ensure employment is maintained and the unintended consequences of AI are avoided.
For Australians, a good example of when algorithms go rogue was seen with her Robodebt scandal. That was a system that used software to identify alleged welfare cheats. But the software incorrectly identified people, forcing them to repay finds that they were entitled to. This caused widespread distress across the country.
A model for understanding AI
We can look at AI has four distinct levels. The most basic form of AI is process robotics where systems mimic human action. We then move up to intelligent automation where systems mimic or augment human judgement.
Where it gets interesting is with cognitive automation where systems augment human intelligence. I think generative AI, like ChatGPT, sits between intelligent automation and cognitive automation.
At the pinnacle there’s artificial general intelligence. These are systems that can mimic human intelligence.
How does AI work?
An analogy for what AI systems do is trying to describe the colour green to someone that has been 100% blind for their entire life. Then consider what data you’d need to provide a computer to tell the difference between green, red and yellow.
AI systems have two broad components: the model which is the software that creates the answer. And the data, which is used to train the model.
AI experts use the terms features and labels respectively to describe the inputs and outputs of AI systems. They want to know what features they want to add to a system in order to label an output correctly. For example, I have a 5kg cavoodle and 22kg cattle dog. What features would I need to give the system to correctly label both as dogs?
So, the data used to train the model has to be of a sufficient quality so the outputs are not wrong. Or worse, unexplainable.
In the USA, AI systems have been created to assist judges with sentencing convicted criminals. Those systems have been trained with historical data. As a result, black people are typically jailed for longer than white people accused of similar crimes. Compounding this, the algorithm tries to predict recidivism rates. And it almost always underestimates white recidivism and overestimates black recidivism.
Is the model wrong? Possibly. But the bigger problem is that the features used to train the systems are delivering the wrong labels. The system is trained with historical data which is, unfortunately, tainted by historical racism in the justice system.
What does this have to do with Apple?
The big question is what are Apple’s AI ambitions? If the goal is to do all the AI processing on your device, and not at a cloud service such as ChatGPT, then those ambitions will be limited by the data available to the model.
Apple’s investment in its own processors gives it a leg up on competitors. The Neural Engine, a collection of specialised computer chips, enables devices to perform AI-related operations. Instead of using a processor made to cover generic tasks, it is specialised for AI tasks.
We already see the benefits of the Neural Engine when we use FaceID or when the system automatically detects text in images or wth the new Personal Voice feature that can replicate your voice if you’re in danger of losing your ability to speak.
Apple could look at everything from where you go to who you interact with and how you prefer to communicate to automate specific tasks to assist you.
What we do know is that Apple has a couple of related projects on the go. One is called Ajax and believed to be a platform for creating large language models (LLMs) like GPT-4 which is being leveraged by Microsoft with its Bing Chat tool. Google’s take on this is PaLM 2 model – it’s powering Google Bard.
The models created by Ajax are probably going to be used to enable AppleGPT.
When can we expect to see AppleGPT?
I would not hold my breath. Apple’s modus operandi has been to let the dust settle on new technologies and implement them when they are ready. Apple could have rushed an iPhone out years before it came to market. Instead, it took its time to ensure the hardware and software were optimised.
Similarly, we can expect Apple to only embed AI capability into its software and hardware when it believes it will work effectively and safely. I suspect that will during 2025.
My gut feeling is that AppleGPT, if that’s what it’s called, won’t be announced as an application or product of its own. More likely, it will be a capability that’s baked into other applications. And that means we may not see it as a big bang announcement. It may get some special love, like the video made with Dan Riccio when Apple introduced its unibody manufacturing process.
But it won’t exist as a separate application or service.
Anthony is the founder of Australian Apple News. He is a long-time Apple user and former editor of Australian Macworld. He has contributed to many technology magazines and newspapers as well as appearing regularly on radio and occasionally on TV.