Post AWBvsM9t2aREqPQXPk by allafarce@hachyderm.io
(DIR) More posts by allafarce@hachyderm.io
(DIR) Post #AWBrK59lLcYES1loyO by simon@fedi.simonwillison.net
2023-05-30T19:27:42Z
0 likes, 0 repeats
I posted a mockup of a design change for ChatGPT that I think could help address the risk of people being lead astray by its incredible ability to invent faleshoods: ChatGPT should include inline tipshttps://simonwillison.net/2023/May/30/chatgpt-inline-tips/
(DIR) Post #AWBtM9bJzgj4ismC3s by abbynormative@mstdn.social
2023-05-30T19:50:08Z
0 likes, 0 repeats
@simon I like this idea, but also know that there are people who will still overlook inline warnings like this. This should be implemented, but before anyone can use ChatGPT, they should also have to go through a training on acceptable uses and how *their* data might be used in the model.
(DIR) Post #AWBvI8fRir9IZsv1ma by oscherler@tooting.ch
2023-05-30T20:11:43Z
0 likes, 0 repeats
@simon I was hoping for a 500 Server Error page. That would do it.
(DIR) Post #AWBvsM9t2aREqPQXPk by allafarce@hachyderm.io
2023-05-30T19:54:55Z
0 likes, 0 repeats
@dbreunig @simon My experience with ChatGPT so far is that they seem to be focusing on the cues being in the model response, not out of band, no? (Not saying I don't like the idea in general, just saying it seems to me they are more interested in directly caveating by model tweak than interface, which is especially important given this affordance doesn't really work in an API by default)
(DIR) Post #AWBvsMwo6hZPI8JbfM by simon@fedi.simonwillison.net
2023-05-30T20:18:30Z
0 likes, 0 repeats
@allafarce @dbreunig that's the approach that u don't think works well enoughI want those things to be visually distinct and shown outside of the main conversation
(DIR) Post #AWBw7NKMyb0WoWUYVc by dbreunig@note.computer
2023-05-30T20:21:02Z
0 likes, 0 repeats
@simon @allafarce Same. I don't think it's in their current priority set.They want to both drive massive adoption while stoking fears to drive protective regulation. There are a steady stream of examples of Altman talking out of both sides of his mouth.
(DIR) Post #AWBwIdWv4nmmFZjQye by edross@mas.to
2023-05-30T20:21:34Z
0 likes, 0 repeats
@simonMaybe it just takes lots of use. I've played with it for long enough to know it is a giant BS generator - but once you know that, you can have fun with it. Last night I had it start generating case law examples for a trial taking place in a D&D scenario I had it writing about
(DIR) Post #AWBwTgCjTPeXjKYU76 by runewake2@hachyderm.io
2023-05-30T20:23:03Z
0 likes, 0 repeats
@simon This doesn't really solve the problem, especially since OpenAI is advocating for more uses via their API. The output of these models has to be addressed and that's not something a web UX change can fix. The technology platform itself is unfit for the applications it is being used for.
(DIR) Post #AWBwfb0ub8IZ6fhymu by simon@fedi.simonwillison.net
2023-05-30T20:24:38Z
0 likes, 0 repeats
@runewake2 I'm not trying to solve the general hallucination problem here - I think just addressing the social problem of ChatGPT users being hoodwinked would be a big win
(DIR) Post #AWC9f0qCO0pIK4kWwa by runewake2@hachyderm.io
2023-05-30T22:53:01Z
0 likes, 0 repeats
@simon I think for that de-personifying output would be more effective. Using it pronouns and "this generative AI model" instead of a branded humanlike personification that it does now.To be clearer: I'd like the output of the model to be modified to clarify the AIs intentions and capabilities and very explicitly refer to itself as a computer algorithm and machine.
(DIR) Post #AWCqW3bd260wJxuVxg by eloquence@social.coop
2023-05-31T06:53:16Z
0 likes, 0 repeats
@simon There's also the #onboarding UX - e.g., guide the user through a tutorial before using ChatGPT, and keep that tutorial accessible later as well. For all the effort that OpenAI is putting towards "existential risk" mitigation and alignment, it's extraordinary to me that they're cooperating on major integrations (e.g., Bing) without solving for these fundamental #UX questions.
(DIR) Post #AWCrm7JZ6zPOPWgQlc by Transportist@mastodon.social
2023-05-31T07:06:50Z
0 likes, 0 repeats
@simon @eloquence cc @harrymccracken
(DIR) Post #AWCsAYAOvZGSdp30Jk by oliver_schafeld@mastodon.online
2023-05-31T07:11:43Z
0 likes, 0 repeats
We expect footnotes in scientific literature. I guess most of us would like blog posts to contain links to sources confirming key statements or expanding on the topic too.It's been fine to replace footnotes with hyperlinks. It's a very bad idea to sacrifice hyperlinks to "just being told by the app". But I'm afraid that's just how lazy/naive far too many people are.🤔 …if they had to include links to paintings, movies, etc that "inspired" these AI image generators — DMCA takedown tsunami?
(DIR) Post #AWDFtO0XqK4T6vmmZs by kellogh@hachyderm.io
2023-05-31T11:37:41Z
0 likes, 0 repeats
@simon someone in HN suggested that the logits from the model could be used as a confidence. Is this remotely true? (My hunch is yes, but not very well)
(DIR) Post #AWDKMPZTVc62ArDh0i by codybrom@mstdn.social
2023-05-31T12:27:35Z
0 likes, 0 repeats
@simon @eloquence In a recent post, I tried to replicate something I’d tried before and there seems to be some recent effort on their part to prevent major hallucinations like this. A similar prompt earlier this month hallucinated an entire episode of This American Life with a fake title, fake segment, fake people and fake Spotify link.But I agree, even if it’s saying the right answer now it could and should do that with better formatting that highlights that fact.