I haven't personally used "computer use" for this, but I did have ChatGPT and Claude Code help me solve system issues via backend tools (scanning task list, memory hogs), and even a big one that made YouTube videos choppy. (Turns out it was a screen refresh setting issue, and ChatGPT troubleshot and told me how to fix it.)
Yes, a bit - it's a little jarring. I like the backend approach as well, but still haven't cleaned the hardware up yet! I'm gonna give that a shot once I'm caught up.
Yeah I've now had a few situations where having Computer Use walk me through interfaces was quite handy! What parts of the process did you need help with?
I didn’t use Computer Use 🙈 (never have; I have used browser control and even that limited. I do a lot of UX verification but I just use screenshots).
Rebuilding my box, I needed to disable bitlocker, make sure everything was backed up, install clean wiping c: but not d:, etc. just an interactive checklist of prepping and doing the upgrade. Once I was in it, I switched the chat over to my iPad … Its so much faster now, etc.
Ah, gotcha, thought you were addressing Andrew's (of GoatFury fame) question about using “computer use” for screen troubleshooting. Still cool how much help you can get from AI in areas you don't quite understand yet.
Yeah for sure, but keep in mind that local agents don't need computer use to do lots of damage to your files and emails due to prompt injection, bad instructions, etx.
In fact, that can do that quicker and in a more subtle way with background scripts and tools.
If anything, computer use lets you see what they're doing in real time and stop them before they do any damage (especially seeing how slow the process is)
The deleted-emails story is a real risk, but what actually decides how bad it gets isn't whether it's computer-use or a background script, it's whether the action can be undone. Delete-forever and send-to-everyone are the actions that cause real damage. If an agent's writing to a draft folder or a staging list first, a bad prompt just leaves you a mess to clean up.
Worth checking what the tool can actually do before you hand it something, not after.
Hey interesting! I usually find it cringy when AI labs demo their agents by showing the "grocery shopping" use case, but it's cool to hear that it's legitimately helpful to people. Do you find that it's reliable enough to actually do the job well without you having to review and change stuff?
Awesome! How long would you estimate it takes it to click through everything and fill a cart? I mean, the good thing is that you can just walk away even if it takes it a long time and only review at the end.
Building on what Allie found, the time it takes only works because a wrong or missing item in a cart is cheap and easy to catch at checkout. That's really about how safe the mistake is to catch, more than how slow the task is.
A lot of tasks in that same tedious bucket aren't like that, a wrong output is harder to spot, and that's usually where the review step ends up costing more time than just doing it by hand.
Ever use this feature to clean up your computer, like improve performance or do maintenance? "Make my computer load faster", that type of thing?
I haven't personally used "computer use" for this, but I did have ChatGPT and Claude Code help me solve system issues via backend tools (scanning task list, memory hogs), and even a big one that made YouTube videos choppy. (Turns out it was a screen refresh setting issue, and ChatGPT troubleshot and told me how to fix it.)
Have you tried "computer use" yourself?
Yes, a bit - it's a little jarring. I like the backend approach as well, but still haven't cleaned the hardware up yet! I'm gonna give that a shot once I'm caught up.
AI held my hand while I rebuilt my windows 11 box from the ground up and I’ve never been happier
Yeah I've now had a few situations where having Computer Use walk me through interfaces was quite handy! What parts of the process did you need help with?
I didn’t use Computer Use 🙈 (never have; I have used browser control and even that limited. I do a lot of UX verification but I just use screenshots).
Rebuilding my box, I needed to disable bitlocker, make sure everything was backed up, install clean wiping c: but not d:, etc. just an interactive checklist of prepping and doing the upgrade. Once I was in it, I switched the chat over to my iPad … Its so much faster now, etc.
Ah, gotcha, thought you were addressing Andrew's (of GoatFury fame) question about using “computer use” for screen troubleshooting. Still cool how much help you can get from AI in areas you don't quite understand yet.
I'm not sure I would want to do anything like that given the story about the developer who had ALL her emails deleted by a screwed up prompt.
Yeah for sure, but keep in mind that local agents don't need computer use to do lots of damage to your files and emails due to prompt injection, bad instructions, etx.
In fact, that can do that quicker and in a more subtle way with background scripts and tools.
If anything, computer use lets you see what they're doing in real time and stop them before they do any damage (especially seeing how slow the process is)
The deleted-emails story is a real risk, but what actually decides how bad it gets isn't whether it's computer-use or a background script, it's whether the action can be undone. Delete-forever and send-to-everyone are the actions that cause real damage. If an agent's writing to a draft folder or a staging list first, a bad prompt just leaves you a mess to clean up.
Worth checking what the tool can actually do before you hand it something, not after.
I wouldn't be doing this at all if I didn't have a 2nd laptop that's somewhat siloed. Still, I get the heebie-jeebies too every day.
I fully expected to see the AIs pull-up photos of THE 'Cat In The Hat'!
Ha, I think it consciously avoided that due to copyright fears and specifically search for royalty free images!
I found computer use to be most useful for ordering groceries online, which falls into your tedious task bucket.
Hey interesting! I usually find it cringy when AI labs demo their agents by showing the "grocery shopping" use case, but it's cool to hear that it's legitimately helpful to people. Do you find that it's reliable enough to actually do the job well without you having to review and change stuff?
I do review it and occasionally make a change or two but it still saves me time, as I typically have over 20 items in my cart.
Awesome! How long would you estimate it takes it to click through everything and fill a cart? I mean, the good thing is that you can just walk away even if it takes it a long time and only review at the end.
Yes I leave it to do other things. Last week it took 18 minutes to add all the items to the cart.
Building on what Allie found, the time it takes only works because a wrong or missing item in a cart is cheap and easy to catch at checkout. That's really about how safe the mistake is to catch, more than how slow the task is.
A lot of tasks in that same tedious bucket aren't like that, a wrong output is harder to spot, and that's usually where the review step ends up costing more time than just doing it by hand.