4 ms·
>Q: What happens when output from LLMs start feeding back into training new LLMs? Humans do this all the time. For example, fake news spreading and most people
by senttoschool 3y ago
>Q: What happens when output from LLMs start feeding back into training new LLMs?
Humans do this all the time. For example, fake news spreading and most people believing in it.
- staticman2 3y agoWhat's the point of constantly comparing this technology to humans? If I said my calculator gives false numbers you wouldn't reply "humans are sometimes bad at math".
- _ea1k 3y agoNo, but if my LLM were bad at math, I might say "humans are sometimes bad at math". One major missing element in current ML approaches is tool usage and how to decide when to use a tool vs not using one. It is an area that is currently being explored and expanded quickly.
- dekhn 3y agobecause the whole point of these systems is to take over menial tasks that humans are slow/bad/find boring. Same reason as calculators: they do something humans could do, but faster and more accurately, allowing us to focus our intelligence on higher order problems.
- raspyberr 3y agoBut some humans are skeptical and/or critical thinkers and will get this new information verified or disproven through other methods. Also, fake news spreading seems mainly an issue of media/social media. There is no information spreading mechanism set up yet for LLMs. Furthermore, to LLMs all information looks the same. To humans more emotionally charged information is more prominent.
- squarefoot 3y agoOne difference being that we have previous education, culture and critical thinking allowing us to not fall into bullshit even in case it is being spread by more sources than the correct information. I'm not sure an AI being trained on the load of garbage that comes from social media today could develop that level of resistance.
- senttoschool 3y agoYou overestimate the average human's ability to critically think for themselves.
- jqpabc123 3y agoTrue but it misses the point. LLMs will only help fake news spread faster because they're dumb --- they can't tell the difference between real and fake. They are indiscriminate multipliers of both real and fake data.