Apple's Unified Memory powers Kiraa AI's rapid automation

Talk to most people working with AI and they'll almost all tell you the same thing. To get the best performance from AI you need the latest GPUs from Nvidia. But an Australian start up is bucking that trend and creating high-performance AI and automation tools on a completely different set of chips.

Kiraa AI is building AI tools that are high performance and can run on hardware that almost any business can afford, allowing them to reap the benefits of AI without costly token spends or rely on cloud services. The key to doing this lies in the architecture of Apple Silicon – the class of chips Apple designed and come as standard in every Mac.

Errol Brandt, co-founder of Kiraa AI, says the benefits come from Apple's Unified Memory. The CPU, GPU and memory reside on a single chip enabling faster movement of data between each different type of core.

"The big play is unified memory," explained Brandt. "If you look at what everyone's doing today, they're using GPUs and throwing the whole box and dice at a problem for LM [language model] training. But in a workflow situation or an enterprise, you have a mixed compute requirement. You're reading files, you're going through lists, some of those need a CPU, others need a GPU and some need a neural engine. Unified memory means we put an object in memory and just apply the operations to it."

Brandt explained that by not moving data between separate CPU, GPU and Neural Engine memory it's much faster and generates much less heat. He added that the performance is orders of magnitude better than other platforms.

"We're getting workloads that would take days to run in hours and minutes," he added.

One of the first successes Kiraa AI had with its approach came with a manufacturing company. As a former CFO, Brandt was acutely aware of the workloads finance teams face at the end of each month. For many organisations, month-end processing can take several days to complete. Kiraa AI reduced that process to an hour.

And while Kiraa AI typically recommends that clients use a Mac Studio, Brandt has the software running on an iPad. As a self-confessed geek, he has also developed a version of the software that he demonstrated working on an iPhone. And he's working on a side project to run the software on an Apple TV.

While Kiraa AI does use LLMs and other AI techniques in its software, Brandt said it's important to understand not every problem needs an AI solution.

"I don't think that LLMs are the right technology. People are putting their faith into token prediction machines. There are brand new architectures out there which are much more suitable for reasoning and advanced decision making than simply predicting the next tokens," he said.

Other forms of automation still have a place and he added that human oversight remains important. Rather than using AI to control the workflow, which is what is advocated by proponents of agentic AI, Kiraa AI's approach is to use AI to populate parts of the workflow.

"We can't just trust agents. What happens when it goes wrong? You won't know why. If you have a deterministic workflow, you can identify which step failed and fix it. We're doing the opposite of what everyone else is doing."

A significant factor in Kiraa AI's early success has been the series of YouTube videos Brandt has been publishing over the last few months. His initial audience of around 500 subscribers has grown to over 45,000. Among that audience were some of Apple's AI team who reached out to Brandt.

While Apple does not directly support or fund Kiraa AI, they have met with Brandt and his team suggesting they are on the right track. With about 1.5 billon Apple device that can run AI at the edge, it raises the question as to whether the current mania around building new data centres is completely justified.

Thus far, Kiraa AI has been bootstrapped by Brandt and his two co-founders putting up significant amounts of personal cash as well as some other small investors. Interestingly, the company has taken a very egalitarian stance on salaries with all employees, from the CEO to developers, paid equally. But the company is now looking at crowdfunding as it grows.

"Looking ahead, the big target date for us is 30 June 2029. By then, we want a thousand customers. After that, we'll assess what our next steps and goals are," Brandt said.

Kiraa AI's approach leverages Apple's unified‑memory architecture to deliver high‑performance AI workloads on commodity hardware. This enables organisations to retain control of their data by keeping it within their own environment which reduces costs associated with running AI workloads on cloud servers.