We’re looking at how we can use local, on-device AI models – i.e., more private – to enhance your browsing experience further. One feature we’re starting with next quarter is AI-generated alt-text for images inserted into PDFs, which makes it more accessible to visually impaired users and people with learning disabilities. The alt text is then processed on your device and saved locally instead of cloud services, ensuring that enhancements like these are done with your privacy in mind.
IMO if everything’s going to have AI ham fisted into it, this is probably the least shitty way to do so. With Firefox being open source, the code can also be audited to ensure they’re actually keeping their word about it being local-only.
Nope, they can use your NPU, GPU or CPU whatever you have… the performance will vary quite a bit though. Also, the larger the model the more memory it needs to run well.
With it being local it’s probably a small and limited model. I took a couple courses on machine learning years ago (before it got rebranded as “AI”), and you’d be surprised at how well a basic image recognition model can run on the lowest-spec macbook from 2012.
Tbh the inversion of typical intuition that is LLMs taking orders of magnitudes more memory than computer vision can mess people unfamiliar up on estimates of the hardware required
IMO if everything’s going to have AI ham fisted into it, this is probably the least shitty way to do so. With Firefox being open source, the code can also be audited to ensure they’re actually keeping their word about it being local-only.
Don’t you need specific CPUs for these AI features? If so, how is this going to work on the machines that don’t support it?
Nope, they can use your NPU, GPU or CPU whatever you have… the performance will vary quite a bit though. Also, the larger the model the more memory it needs to run well.
With it being local it’s probably a small and limited model. I took a couple courses on machine learning years ago (before it got rebranded as “AI”), and you’d be surprised at how well a basic image recognition model can run on the lowest-spec macbook from 2012.
Tbh the inversion of typical intuition that is LLMs taking orders of magnitudes more memory than computer vision can mess people unfamiliar up on estimates of the hardware required
Yeah that’s very true.
You only need lots of precessing power to train the models. Using the models can be done on regular hardware.
Running AI models isn’t that resource intensive. Training the models is the difficult part.
The feature will obviously just be disabled on machines that don’t support it.