Post AU0UdI4xZTOPm8cw2C by TedUnderwood@sigmoid.social
 (DIR) More posts by TedUnderwood@sigmoid.social
 (DIR) Post #AU0UdI4xZTOPm8cw2C by TedUnderwood@sigmoid.social
       2023-03-26T12:09:00Z
       
       0 likes, 0 repeats
       
       It’s going to be hard for #AI to provide what people really need—which is, often, better judgment. E.g. what happens to editing software that advises its user “you should cut here and here because these ¶s are repetitive / predictable”?
       
 (DIR) Post #AU0WV8mYS7gQX8GA64 by budgibson@me.dm
       2023-03-26T12:29:57Z
       
       0 likes, 0 repeats
       
       @TedUnderwood People are also poor at tedium. That is potentially a clear win.
       
 (DIR) Post #AU0ZCkcFrYBmEXMfoW by budgibson@me.dm
       2023-03-26T13:00:14Z
       
       0 likes, 0 repeats
       
       @TedUnderwood LLMs are also really good at helping you fake it. The hot question being debated right now (and a bit the point of your post) is whether and when the faking is adequate. Faking is about adhering to a common pattern. Transcription and translation are two domains that seem to benefit tremendously from this property. They’re also tedious.
       
 (DIR) Post #AU0xX7lJufnGiBqv8S by pinecone@mastodon.social
       2023-03-26T17:32:51Z
       
       0 likes, 0 repeats
       
       @TedUnderwood Current AI, LLM models, don't use working memory to overview their action, so they don't and can't do judgement at all. That overview and working memory stuff is for the next generation of models.
       
 (DIR) Post #AU138kB2YABkL3dmOO by lobrien@sigmoid.social
       2023-03-26T18:35:40Z
       
       0 likes, 0 repeats
       
       @TedUnderwood I’m not sure that’s correct. I think an #LLM would be able to detect “rate of change” over a context window and suggest a more compact, denser text. (Of course, the LLM would have to be prompted/fine-tuned by someone with judgment, but surely that’s exactly the thunder we hear approaching.)
       
 (DIR) Post #AU1HjJDf8ip6ntCHFg by TedUnderwood@sigmoid.social
       2023-03-26T21:19:09Z
       
       0 likes, 0 repeats
       
       @lobrien If you follow through the thought experiment in the OP ('what happens to editing software ...') you'll see my concern is not that judgment is impossible to automate, but that automated judges won't be heeded.