Hacker News Reader

Comment by Taek | Hacker News Reader

I don't like the framing of 'learn something new' because you are very likely to have falsehoods presented to you as reliable facts.

Its fun to ask Marie Curie what she thinks about blockchains, but anything you ask about her own life and work is just as likely to be fiction as whatever she says about blockchains.

View on HN

See, for instance, the conversations in this thread with figures like Henry Ford, Ronald Reagan, and Lyndon Johnson: https://twitter.com/ZaneGTCooper/status/1615577714836275200

The limitations of using pre trained data with a simple prompt. The bot will get inspired by the prompt a little, but most of its text comes from the training which is adjusted to sound as neutral and "generally acceptable human"like as possible.

This means that for most cases the personality disappears.

To add: if the training and fine tuning comes from how the subject actually responded, spoke and reasoned, then that would be quite scarily accurate.

tbf, Reagan insisting that he did the best he could with the AIDs crisis is quite plausible.

That's the other problem with the theory of understanding historical figures by chatting with them: even if the simulation was realistic or we were able to reincarnate the actual historical figures they'd lie, dissemble and make excuses a lot...

I agree with the original point, but giving an answer that shows yourself in the most favorable light seems like what these bots should be doing. Despite the innaccuracies, I think that's pretty much the angle someone like Reagan would take if confronted now about those events.

If people are using this app to learn about these figures, and by extension history, then prevarication and falsehood amounts to propaganda and historical revisionism. People aren't going to treat these apps like games, but as authoritative sources of truth.

This issue is more deep and important than most would think.

Reply on news.ycombinator.com