In one way, this approach is a complete failure to understand what AI is all about. We have the power to distil information from all over the internet, taking in the viewpoints and experiences of millions of people creating billions and billions of data points. On the other hand, when you want to know what time the last bus home is on a Sunday, you don’t need billions of data points: you need one. 

If you were to create a LLM trained only on words that you’d written yourself – your diaries, your musings, your social media posts – you’d vastly cut down on the amount of data used, which should in turn lessen the processing power taken to churn through that data. Suddenly the environmental damage, the much-publicised water usage of AI data centres, shrinks to nothing. 

There would be no ethical concerns over copyright, as you’d own everything. And there would be no problem verifying the accuracy of the data, in part because you’d have no incentive to fill the LLM with nonsense, and also because you’d know your own limitations. There would be no point asking an LLM trained on my writings how to build a nuclear power station, for example, because I know that I don’t know. 

If this sounds fanciful, maybe it shouldn’t. Useful Sensors has a range of sensors with inbuilt AI running on the device, so that it can, for example, recognise when a user is looking directly at a camera, without sending any data back to its parent company. And in issue 158 we ran a project showcase on the Ubo Pod AI assistant, a personal, private LLM that processes your data locally. When we get AI right, odds on it’ll be thanks to small firms, motivated individuals, and Raspberry Pi.