The world of web browsers have not been spared by the trend of integrating LLM functionality. But there are fundamental issues with it and Vivaldi addresses them.
I’d love a browser-embedded LLM that had access to the DOM.
“Highlight all passages that talk about yadda yadda. Remove all other content. Convert the dates to the ISO standard. Put them on a number line chart, labeled by blah.”
You are falling into a common trap. LLMs do not have understanding - asking it to do things like convert dates and put them on a number line may yield correct results sometimes, but since the LLM does not understand what it’s doing, it may “hallucinate” dates that look correct, but don’t actually align with the source.
Thank you for calling that out. I’m well aware, but appreciate your cautioning.
I’ve seen hallucinations from LLMs at home and at work (where I’ve literally had them transcribe dates like this). They’re still absolutely worth it for their ability to handle unstructured data and the speed of iteration you get – whether they “understand” the task or not.
I know to check my (its) work when it matters, and I can add guard rails and selectively make parts of the process more robust later if need be.
I’d love a browser-embedded LLM that had access to the DOM.
“Highlight all passages that talk about yadda yadda. Remove all other content. Convert the dates to the ISO standard. Put them on a number line chart, labeled by blah.”
That’d be great UX.
You are falling into a common trap. LLMs do not have understanding - asking it to do things like convert dates and put them on a number line may yield correct results sometimes, but since the LLM does not understand what it’s doing, it may “hallucinate” dates that look correct, but don’t actually align with the source.
Thank you for calling that out. I’m well aware, but appreciate your cautioning.
I’ve seen hallucinations from LLMs at home and at work (where I’ve literally had them transcribe dates like this). They’re still absolutely worth it for their ability to handle unstructured data and the speed of iteration you get – whether they “understand” the task or not.
I know to check my (its) work when it matters, and I can add guard rails and selectively make parts of the process more robust later if need be.