3 ms·
It’s just that bots haven’t been good enough yet. With the new LLM tech they can pretty much pass every hurdle you’ll throw at them. Even if you require people
by EGreg 2y ago
It’s just that bots haven’t been good enough yet. With the new LLM tech they can pretty much pass every hurdle you’ll throw at them. Even if you require people to show up in person, they’ll do that but then run a bot the rest of the time in their account.
I am sure that LLMs and bots will be able to fool many people on HN and run “rings” around dang’s ring detection software, in about 5 years. It’s a gameable metric, after all.
They were already able to do it on 4chan in 2020 with just GPT3! And the most impactful thing is users started accusing each other of being bots! It literally enshittified the whole forum overnight:
https://finance.yahoo.com/news/breaches-every-principle-human-research-092305390.html https://finance.yahoo.com/news/breaches-every-principle-huma...
- Tiberium 2y agoAnd to be more exact, GPT-4chan is based on GPT-J (same architecture as GPT-3 whose weights were never released) which only had 6B params and that was back in 2021-2022.
- MichaelZuo 2y agoThere's a straightforward but costly way, tie it to something that costs money, over the long term. E.g. Utility bills, bank account statements, etc..., for x number of years. And manually confirm with the companies at random.
- EGreg 2y agoTie what? They’ll just check in and then run bots in their account. Line a chess bot for example
- MichaelZuo 2y agoHow can they fake the bank or utility companies internal records?