Post ASTKLZMsWDHb2N5N3Y by simon@fedi.simonwillison.net
(DIR) More posts by simon@fedi.simonwillison.net
(DIR) Post #ASRivjQuaEufGJpRVw by simon@fedi.simonwillison.net
2023-02-07T19:40:31Z
0 likes, 0 repeats
This screenshot of the new language-model powered Bing from the Verge live blog at https://www.theverge.com/2023/2/7/23588249/microsoft-event-ai-live-blog-openai-chatgpt-bing-announcements-news is really interesting from an AI interface design perspective"According to my web search, some of the most influential Mexican artists and their top paintings are:" - then each thing it lists appears to include inline citation links
(DIR) Post #ASRj8GxNljZAi1zqV6 by simon@fedi.simonwillison.net
2023-02-07T19:42:41Z
0 likes, 0 repeats
"According to my web search..." is a clever piece of micro-copy: I think it successfully communicates that the AI is summarizing results that you can click through to yourself, which feels safer than regurgitating facts without any hint as to where the information came fromI won't be at all surprised if people quickly find examples of searches that produce additional (misleading or hallucinated) output that's not present in any of the cited results though
(DIR) Post #ASRjKusBBhF9yvalEm by ditherycarbon@sigmoid.social
2023-02-07T19:43:07Z
0 likes, 0 repeats
@simon pretty similar to how you.com works
(DIR) Post #ASRjXybWcrD3Jq1tjc by Colarusso@mastodon.social
2023-02-07T19:46:34Z
0 likes, 0 repeats
@simon hum, that looks familiar. ;) See https://suffolklitlab.org/bs/
(DIR) Post #ASRk6ltVsAFY2G2Djs by stammy@stammy.design
2023-02-07T19:53:14Z
0 likes, 0 repeats
@simon I think Neeva and Perplexity have been trying to do this as well with citation links
(DIR) Post #ASRkacQRdrpVM7fdui by sayrer@mastodon.social
2023-02-07T19:58:43Z
0 likes, 0 repeats
@simon oh wow, I immediately thought of https://www.cs.utexas.edu/users/EWD/ewd08xx/EWD854.PDFI just re-read the entire thing. I'm not even sure I agree, but he has a point. You can also so see his penmanship decline over the length of the letter, if you read the whole thing. I think he just getting mad and tired.People copy/paste the anthropomorphization quotes, but I find what follows to be more profound.
(DIR) Post #ASRlfmJgOYb3QyX2Tw by simon@fedi.simonwillison.net
2023-02-07T20:10:04Z
0 likes, 0 repeats
It looks to me like the Bing implementation is the same pattern that a bunch of other people are trying at the moment: you run a traditional search, then dump the search result summaries into a prompt and get the language model to generate text using those summariesI tried something similar against my blog a few weeks ago! https://simonwillison.net/2023/Jan/13/semantic-search-answers/
(DIR) Post #ASRmETljLY7Q2rCJKi by richvn@mastodon.social
2023-02-07T20:17:14Z
0 likes, 0 repeats
@simon Yes exactly. This is how Elicit works and perplexity.ai too.
(DIR) Post #ASSxGM3cHY5hMScX1k by halcek@merveilles.town
2023-02-08T09:55:29Z
0 likes, 0 repeats
@simon But were the results actually relevant in terms of comparing the artists? All I can see from that screenshot is the artists and their works being listed out, so the relevance of the results to the search query still seems kind of low on Bing anyway.
(DIR) Post #ASTKLZMsWDHb2N5N3Y by simon@fedi.simonwillison.net
2023-02-08T14:14:17Z
0 likes, 0 repeats
@halcek yeah it's actually a pretty bad example of a query, because it's not a question that's very well suited to how LLM-assisted search actually worksGives the impression that the AI is capable of evaluating all artists and picking the most influential, when really it's running a search for "influential Mexican artists" and summarizing whatever comes back in the first few results