iOS 18 could be loaded with AI, as Apple reveals 8 new artificial intelligence models that run on-device

Apple has released a set of several new AI models that are designed to run locally on-device rather than in the cloud, possibly paving the way for an AI-powered iOS 18 in the not-too-distant future.

The iPhone giant has been doubling down on AI in recent months, with a carefully split focus across cloud-based and on-device AI. We saw leaks earlier this week indicating that Apple plans to make its own AI server chips, so this reveal of new local large language models (LLMs) demonstrates that the company is committed to both breeds of AI software. I’ll dig into the implications of that further down, but for now, let’s explain exactly what these new models are.

The suite of AI tools contains eight distinct models, called OpenELMs (Open-source Efficient Language Models). As the name suggests, these models are fully open-source and available on the Hugging Face Hub, an online community for AI developers and enthusiasts. Apple also published a whitepaper outlining the new models. Four were pre-trained on CoreNet (previously CVNets), a massive library of data used for training AI language models, while the other four have been ‘instruction-tuned’ by Apple; a process by which an AI model’s learning parameters are carefully honed to respond to specific prompts.

Releasing open-source software is a somewhat unusual move for Apple, which typically retains quite a close grip on its software ecosystem. The company claims to want to “empower and enrich” public AI research by releasing the OpenELMs to the wider AI community.

So what does this actually mean for users?

Apple has been seriously committed to AI recently, which is good to see as the competition is fierce in both the phone and laptop arenas, with stuff like the Google Pixel 8’s AI-powered Tensor chip and Qualcomm’s latest AI chip coming to Surface devices.

By putting its new on-device AI models out to the world like this, Apple is likely hoping that some enterprising developers will help iron out the kinks and ultimately improve the software – something that could prove vital if it plans to implement new local AI tools in future versions of iOS and macOS.

It’s worth bearing in mind that the average Apple device is already packed with AI capabilities, with the Apple Neural Engine found on the company’s A- and M-series chips powering features such as Face ID and Animoji. The upcoming M4 chip for Mac systems also appears to sport new AI-related processing capabilities, something that's swiftly becoming a necessity as more-established professional software implements machine-learning tools (like Firefly in Adobe Photoshop).

In other words, we can probably expect AI to be the hot-button topic for iOS 18 and macOS 15. I just hope it’s used for clever and unique new features, rather than Microsoft’s constant Copilot nagging.

You might also like…

TechRadar – All the latest technology news

Read More

I created an AI app in 10 minutes and I might never think about artificial intelligence in the same way again

Pretty much anything we can do with AI today might have seemed like magic just a year ago, but MindStudio's platform for creating custom AI apps in a matter of minutes feels like a new level of alchemy.

The six-month-old free platform, which you can find right now under youai.ai, is a visual studio for building AI workflows, assistants, and AI chatbots. In its short lifespan it's already been used, according to CEO Dimitry Shapiro, to build more than 18,000 apps.

Yes, he called them “apps”, and if you're struggling to understand how or why anyone might want to build AI applications, just look at OpenAI's relatively new GPT apps (aka GPTs). These let you lock the powerful GPT-3.5 into topic-based thinking that you can package up, share, and sell. Shapiro, however, noted the limits of OpenAI's approach. 

He likened GPTs to “bookmarking a prompt” within the GPT sphere. MindStudio, on  the other hand, is generative model-agnostic. The system lets you use multiple models within one app.

If adding more model options sounds complicated, I can assure you it's not. MindStudio is the AI development platform for non-developers. 

Watch and learn

MindStudio

Choose your template. (Image credit: MindStudio)

To get you started, the company provides an easy-to-follow 18-minute video tutorial. The system also helps by offering a healthy collection of templates (many of them business-focused), or you can choose a blank template. I followed the guide to recreate the demo AI app (a blog post generator), and my only criticism is that the video is slightly out of date, with some interface elements having been moved or renamed. There are some prompts to note the changes, but the video could still do with a refresh.

Still, I had no trouble creating that first AI blog generator. The key here is that you can get a lot of the work done through a visual interface that lets you add blocks along a workflow and then click on them to customize, add details, and choose which AI model you want to use (the list includes GPT- 3.5 turbo, PaLM 2, Llama 2, and Gemini Pro). While you don't necessarily have to use a particular model for each task in your app, it might be that, for example, you should be using GPT-3.5 for fast chatbots or that PaLM would be better for math; however, MindStudio cannot, at least yet, recommend which model to use and when.

Image 1 of 2

MindStudio

Connect the boxes (Image credit: MindStudio)
Image 2 of 2

MindStudio

And then edit their contents (Image credit: MindStudio)

The act of adding training data is also simple. I was able to find web pages of information, download the HTML, and upload it to MindStudio (you can upload up to 150 files on a single app). MindStudio uses the information to inform the AI, but will not be cutting and pasting information from any of those pages into your app responses.

Most of MindStudio's clients are in business, and it does hide some more powerful features (embedding on third-party websites) and models (like GPT 4 Turbo) behind a paywall, but anyone can try their hand at building and sharing AI apps (you get a URL for sharing).

Confident in my newly acquired, if limited, knowledge, I set about building an AI app revolving around mobile photography advice. Granted, I used the framework I'd just learned in the AI blog post generator tutorial, but it still went far better than I expected.

One of the nice things about MindStudio is that it allows for as much or as little coding as you're prepared to do. In my case, I had to reference exactly one variable that the model would use to pull the right response.

MindStudio

Options include setting your model’s ‘temperature’ (Image credit: MindStudio)

There are a lot of smart and dead-simple controls that can even teach you something about how models work. MindStudio lets you set, for instance, the 'Temperature' of your model to control the randomness of its responses. The higher the 'temp', the more unique and creative each response. If you like your model verbose, you can drag another slider to set a response size of up to 3,000 characters.

The free service includes unlimited consumer usage and messages, some basic metrics, and the ability to share your AI via a link (as I've done here). Pro users can pay $ 23 a month for the more powerful models like GPT-4, less MindStudio branding, and, among other things, site embedding. The $ 99 a-month tier includes all you get with Pro, but adds the ability to charge for access to your AI app, better analytics, API access, full chat transcripts, and enterprise support.

Image 1 of 2

MindStudio

Look, may, I made an AI app. (Image credit: MindStudio)
Image 2 of 2

MindStudio

It’s smarter than I am. (Image credit: MindStudio)

I can imagine small and medium-sized businesses using MindStudio to build customer engagement and content capture on their sites, and even as a tool for guiding users through their services.

Even at the free level, though, I was surprised at the level of customization MindStorm offers. I could add my own custom icons and art, and even build a landing page.

I wouldn't call my little AI app anything special, but the fact that I could take the germ of an idea and turn it into a bespoke chatbot in 10 minutes is surprising even to me. That I get to choose the right model for each job within an AI app is even better; and that this level of fun and utility is free is the icing on the cake.

You might also like

TechRadar – All the latest technology news

Read More

Threat Intelligence Necessities

Without threat intelligence, security teams tend to prioritize vulnerabilities based on their severity. Bad actors know this, which is why they often exploit medium severity vulnerabilities. Conversel…

Articles RSS Feed

Read More