The whole demo about this was so stupid. The user said “the text on this website is to small, I cannot read it”, and the ai points out how to click through the windows system setting to find the UI scale option which was at 150%. Then it suggest to set it to the current setting of 150% because there was a label “recommended” on it.
It’s a fail on so many levels. Why could it not just set it without telling the user to “click here” through 5 different dialogs and windows? Why set it to the same setting it already has? Why not set the browsers scale if the user said “this website is to small”?
They just want to stuff ai into everything without knowing what problem they need to solve.
As to why not just do it, problem is that the LLM will generate something to do even if it doesn’t know the correct answer. You don’t want agentic ai to go to town because it will screw up and be even harder or impossible to undo whatever it generated to do.
This specific demo worked, but it’s a crapshoot as to whether a scenario will work as an llm “failure” still generates output, and nothing knows that it is “wrong”.
The whole demo about this was so stupid. The user said “the text on this website is to small, I cannot read it”, and the ai points out how to click through the windows system setting to find the UI scale option which was at 150%. Then it suggest to set it to the current setting of 150% because there was a label “recommended” on it.
It’s a fail on so many levels. Why could it not just set it without telling the user to “click here” through 5 different dialogs and windows? Why set it to the same setting it already has? Why not set the browsers scale if the user said “this website is to small”?
They just want to stuff ai into everything without knowing what problem they need to solve.
As to why not just do it, problem is that the LLM will generate something to do even if it doesn’t know the correct answer. You don’t want agentic ai to go to town because it will screw up and be even harder or impossible to undo whatever it generated to do.
This specific demo worked, but it’s a crapshoot as to whether a scenario will work as an llm “failure” still generates output, and nothing knows that it is “wrong”.