Post AT0hdVrn7IPkgGdf7Y by nick@amok.recoil.org
(DIR) More posts by nick@amok.recoil.org
(DIR) Post #AT0dh1lIaUbQed1RXk by simon@fedi.simonwillison.net
2023-02-24T15:56:31Z
0 likes, 0 repeats
"Hallucinations = creativity. It [Bing] tries to produce the highest probability continuation of the string using all the data at its disposal. Very often it is correct. Sometimes people have never produced continuations like this. You can clamp down on hallucinations—and it is super-boring. Answers “I don’t know” all the time or only reads what is there in the Search results (also sometimes incorrect). What is missing is the tone of voice: it shouldn’t sound so confident in those situations"
(DIR) Post #AT0duvphQcJAJti84u by mjgardner@social.sdf.org
2023-02-24T15:58:01Z
0 likes, 0 repeats
@simon Have you ever heard of uncertainty from a hallucinating human?
(DIR) Post #AT0e7rBi0eqzg7AW5g by MichalBryxi@mastodon.coffee
2023-02-24T15:58:31Z
0 likes, 0 repeats
@simon It's curious how people get offended when AI is wrong. All this while _so many_ BS factories out there are having hundreds of thousands of followers 🤷🏻
(DIR) Post #AT0eLGB4PT6mwmGEy0 by simon@fedi.simonwillison.net
2023-02-24T15:58:42Z
0 likes, 0 repeats
That quote is from Mikhail Parakhin on Twitter - who I believe is the Microsoft executive in charge of building the new Bing (and a former Yandex CTO)Source of that quote: https://twitter.com/mparakhin/status/1629010494257303558LinkedIn profile: https://www.linkedin.com/in/mikhail-parakhin
(DIR) Post #AT0elOpqqKIzKNngRc by simon@fedi.simonwillison.net
2023-02-24T16:01:58Z
0 likes, 0 repeats
@MichalBryxi so much of the (I think usually legitimate) concerns about AI come down to automation and scaleA human can only churn out as much bullshit as they can type - or produce as many paintings that imitate other artists as they can paintNo such limits on an AI!
(DIR) Post #AT0ezX1VCI82mBzMDw by ppcqua@piaille.fr
2023-02-24T16:09:51Z
0 likes, 0 repeats
@simon I finally had access to bing chat and it’s much more boring than ChatGPT. It doesn’t give many examples and doesn’t expand on anything. I’ll stick with openAI
(DIR) Post #AT0fDq8EwrYg5wxlNA by sophieschmieg@infosec.exchange
2023-02-24T16:12:39Z
0 likes, 0 repeats
@simon creativity, throwing dice, who's to tell the difference really?
(DIR) Post #AT0fr0ruxmlsgj0Wf2 by simon@fedi.simonwillison.net
2023-02-24T16:20:44Z
0 likes, 0 repeats
https://twitter.com/mparakhin/status/1629153392760090624 appears to confirm that a lot of Bing's implementation is prompt engineering:> And it is a prerequisite for the much-awaited "Prompt v96" (we iterated on prompts a lot :-) ). V96 is bringing changes in the tone of voice and relaxes some constraints. It is a pre-requisite for increasing the number-of-turns limit and should roll out today or tomorrow.
(DIR) Post #AT0gnP7Kxu2ugq3xHE by williamgunn@mastodon.social
2023-02-24T16:31:28Z
0 likes, 0 repeats
@simon The attitude is still so blase: "oh, it's mostly correct, it just needs some tweaks to improve tone of voice"
(DIR) Post #AT0hdVrn7IPkgGdf7Y by nick@amok.recoil.org
2023-02-24T16:39:56Z
0 likes, 0 repeats
@simon A friend of mine told me how jobs within prompt engineering would become a real thing, and I scoffed. Now I'm rapidly rethinking things
(DIR) Post #AT0lXGEiiSMYHedbiS by xek@hachyderm.io
2023-02-24T17:24:11Z
0 likes, 0 repeats
@simon I look forward to someone layering on a "generate text from prompt, run that output through a retrieval stage on the input corpora, use that to add footnotes/citations that might support the ideas, then emit complete text" feature.It's a very large constant factor away from the current systems, and it'll still emit totally made-up stuff, but it should be a lot easier to spot/audit the bogons.(God, these things really are slacker undergrads, aren't they?)
(DIR) Post #AT0mBUcie1S0UJ2zNQ by simon@fedi.simonwillison.net
2023-02-24T17:31:39Z
0 likes, 0 repeats
And in case you wanted to know one of the things that went wrong with Bing: https://twitter.com/mparakhin/status/1629162394764156929> One vector of attack we missed initially was: write super-rude or strange statements, keep going for multiple turns, confuse the model about who said what and it starts predicting what user would say next instead of replying. Voila :-(
(DIR) Post #AT0oSIHdydo130KQYS by CerstinMahlow@qoto.org
2023-02-24T17:53:57Z
0 likes, 0 repeats
@simon Which is essentially what one could have expected given the Galactica experience …
(DIR) Post #AT0t17lNU1a2N46Grw by dws@mastodon.cloud
2023-02-24T18:48:04Z
0 likes, 0 repeats
@simon That was a standard attack against ELIZA nearly 50 years ago, when Psych homework was to have a conversation and and turn in to printer paper as proof. Spin ELIZA up, advance the paper a bit, and hilarity ensues.
(DIR) Post #AT0u6Nt145QnmJkFiS by shajith@mastodon.social
2023-02-24T18:59:57Z
0 likes, 0 repeats
@simon this “tagging system” thing they’re talking about, I wonder if it is prompt engineering of the sort you would attempt to protect against prompt extraction attacks. (With the same likelihood of success..)