Hacker News Reader

Comment by Nextgrid | Hacker News Reader

LLMs don't need browser automation though. Multimodal models with vision input can operate a real computer with "real" user inputs over USB, where the computer itself returns a real, plausible browser fingerprint because it is a real browser being operated by something that behaves humanly.

View on HN

But will they behave like same user in past? I would guess there is lot of difference between how bot accesses page and real user has historically accessed them. Like opening multiple tabs at one time, possibly how long going through next set takes. How they navigate and so on.

There might be lot of modelling that could be done simply based on words used in searches and behaviour of opening pages. All trivially tracked to user's logged in session.

> But will they behave like same user in past? I would guess there is lot of difference between how bot accesses page and real user has historically accessed them. Like opening multiple tabs at one time, possibly how long going through next set takes. How they navigate and so on.

That would be making an assumption that a device and/or account maps 1:1 to a specific human. It does not. People share accounts, share devices, and ask others for one-off help ("hey can you finish buying this for me while I deal with $[whatever our kid just did]", this kind of stuff).

Sure, the cost of that goes way up though, especially if it has to emulate real world inputs like a mouse, type in a way that’s plausible, and browse a website in a way that’s not always the direct happy path.

Reply on news.ycombinator.com