🗿
this really blew up, thanks everyone. these are silly examples but i hope you took away that the "instant answers" that google and bing provide shouldn't be trusted. there's always the possibility of them misunderstanding your question or grabbing irrelevant sources to mislead you. "AI-powered" research is inferior to human research.
don't rely on paraphrasing; scroll right past and find a high quality source on the safety of drinking battery acid instead.
5
0
0
someone reposted this to the bird site where it has gotten almost 90k views so far, and sadly, it appears microsoft patched these results already.
never forget bing encouraged people to drive drunk 🫡
1
0
0
gravitos 
gravitos@wetdry.world
@keat i was surprised to see a fellow wet dry worldian in a random tech meme channel on telegram
1
0
0
Pär Björklund
Paxxi@hachyderm.io@keat and this is what Satya calls responsible AI, try to imagine what irresponsible AI would look like 🤣
0
0
0
Estarriol, lucozade dragon
Thebratdragon@mastodon.scot@keat add in the paid ads and bumped results from the battery acid selling to drinkers companies on top of ai's liking it and....
0
0
0
Tom Bellin 
tob@hachyderm.io
@keat I'd be fascinated to hear what the exact process is behind "patching" these kinds of answers.
I suspect that the answer is totally mundane. They just have a huge dictionary of questions with specific pre-selected (by humans) answers and the "AI" is not engaged to generate the answer at all.
0
0
0
@gravitos i'm glad they credited me unlike the repost i saw on twitter... not like i did anything that warrants credit though lmao
thanks for letting me know, never had anything go viral like this before and feels good knowing it made a bunch of people smile
0
0
0
Dave Rahardja
drahardja@sfba.social@tob @keat My guess is that they have an LLM for the LLM. The second (probably smaller) LLM is trained to watch for known-bad output from the first one. If found, it replaces the output with some benign phrase. In fact, the output of the second LLM could be not words, but just tokens representing OK or NOT OK.
0
0
0
bartholin (forbidden vacation arc)
bartholin@fops.cloud
1
0
0
Third spruce tree on the left
tezoatlipoca@mas.to@JamesHMcLaren @bartholin @keat yeah, that sure is An Intelligence right there.
0
0
1
coral castle
welcome to coral castle, a chill community sailing on the fediverse. we strive to be a safe, left-leaning, queer-friendly, communitarian space for all kinds of folk. before joining, make sure to read our community guidelines.
community guidelines
- you must be over 14 years old to join. coral castle's servers are hosted in chile, and under article 19,628 of the laws of chile, we are unauthorized to collect data from people under 14.
- do not spam or raid. the fediverse is not your space to intoxicate.
- nsfw/nsfl content must be marked as sensitive. coral castle is built to be used as a casual instance, and as such should be accessible anywhere.
- hate speech, paraphilia, and harassment will get you banned. having basic human decency is key anywhere. whether it is in real life or on the internet, don't be a jerk.
- keep things legal. this includes pirated content, since we want to keep a good relationship across fediverse servers. posting CSAM is strictly forbidden and will result in legal action.
moderation policies
the sanctions for local users will be relative to the gravity of their infringements, varying from simply deleting the offending material, to warning the user through a three-strike system, to immediately banning the offending user.
if you're an instance administrator and you think your instance was sanctioned by mistake, feel free to appeal by emailing me at lux@nixgoat.me
special thanks
to our donators, for helping keep this instance up:
also, to the following organizations: