Post ATc7WeaGluojZcbhPk by TedUnderwood@sigmoid.social
 (DIR) More posts by TedUnderwood@sigmoid.social
 (DIR) Post #ATc3TXzSQn8cRDeTei by TedUnderwood@sigmoid.social
       2023-03-14T17:11:39Z
       
       0 likes, 0 repeats
       
       GPT-4, y'all. #gpt4 https://openai.com/product/gpt-4
       
 (DIR) Post #ATc3vfMmbGSLQIyDVA by vh@sigmoid.social
       2023-03-14T17:16:44Z
       
       0 likes, 1 repeats
       
       @TedUnderwood Pricing:gpt-4 with an 8K context window (about 13 pages of text) will cost **$0.03** per 1K prompt tokens, and **$0.06** per 1K completion tokens.gpt-4-32k with a 32K context window (about 52 pages of text) will cost **$0.06** per 1K prompt tokens, and **$0.12** per 1K completion tokens.
       
 (DIR) Post #ATc4KOzmbVXBtC0Vcm by TedUnderwood@sigmoid.social
       2023-03-14T17:21:12Z
       
       0 likes, 0 repeats
       
       @vh yeah, not cheap!
       
 (DIR) Post #ATc4LuXKoudE0BQ7Ki by hackerfriendly@mas.to
       2023-03-14T17:21:27Z
       
       0 likes, 0 repeats
       
       @TedUnderwood They really shouldn't have outsourced the copy writing to ChatGPT.
       
 (DIR) Post #ATc5QFBw4GO3As2bVQ by TedUnderwood@sigmoid.social
       2023-03-14T17:33:27Z
       
       0 likes, 0 repeats
       
       It's accompanied by a document called a "technical report." This resembles a scientific "paper" — in the sense that it could be printed out *on paper.* Otherwise no resemblance. #LLM #GPT4
       
 (DIR) Post #ATc658RA5Loq6Bf3gW by djf@hachyderm.io
       2023-03-14T17:36:51Z
       
       0 likes, 0 repeats
       
       @vh @TedUnderwood Khan academy is using it? I’m not ready for this new world…
       
 (DIR) Post #ATc658yq48l1mdPxB2 by TedUnderwood@sigmoid.social
       2023-03-14T17:40:51Z
       
       0 likes, 0 repeats
       
       @djf @vh Universities aren't ready for it either.
       
 (DIR) Post #ATc6bmMBCUSXAqtevA by pfessenbecker@troet.cafe
       2023-03-14T17:46:41Z
       
       0 likes, 0 repeats
       
       @TedUnderwood  Looks like the hallucination problem is still unsolved, but it is cool —
       
 (DIR) Post #ATc7KLhf90MzXbCKVU by pfessenbecker@troet.cafe
       2023-03-14T17:49:35Z
       
       0 likes, 0 repeats
       
       @TedUnderwood although hang on, maybe this is still the normal ChatGPT.
       
 (DIR) Post #ATc7KMBnKyTN33IOTQ by TedUnderwood@sigmoid.social
       2023-03-14T17:54:48Z
       
       0 likes, 0 repeats
       
       @pfessenbecker yeah, I think that's just ChatGPT, though we can get access to GPT-4. I'll try.
       
 (DIR) Post #ATc7We68ZwiM4AVdRo by jburnford@hcommons.social
       2023-03-14T17:54:56Z
       
       0 likes, 0 repeats
       
       @pfessenbecker @TedUnderwood we currently need to pay for plus to test the new version.
       
 (DIR) Post #ATc7WeaGluojZcbhPk by TedUnderwood@sigmoid.social
       2023-03-14T17:57:01Z
       
       0 likes, 0 repeats
       
       @jburnford @pfessenbecker ah — this might be a reason to pay for plus
       
 (DIR) Post #ATc8sepadLZgfHpTLU by TedUnderwood@sigmoid.social
       2023-03-14T18:12:13Z
       
       0 likes, 0 repeats
       
       @jburnford @pfessenbecker Well, I've just subscribed to Plus. The GPT-4 option is not appearing yet; I'll let you know when I figure out how to get through to it.
       
 (DIR) Post #ATcAuQvZUtnK20bMIq by dotstdy@mastodon.social
       2023-03-14T18:20:26Z
       
       0 likes, 0 repeats
       
       @pfessenbecker @TedUnderwood this "problem" is fundamental though, it's not solvable. There's no version of this software where it generates correct answers.
       
 (DIR) Post #ATcAuRWRHpHjsLqnlg by pfessenbecker@troet.cafe
       2023-03-14T18:24:19Z
       
       0 likes, 0 repeats
       
       @dotstdy @TedUnderwood  In general I'm inclined to agree with you, but honestly at this point I think that's an empirical question. It's not going to get to correct answers the way we do, but I doubt this means it can't ever get there some other way.
       
 (DIR) Post #ATcAuScVCljHHRh9rE by dotstdy@mastodon.social
       2023-03-14T18:30:05Z
       
       0 likes, 0 repeats
       
       @pfessenbecker @TedUnderwood I'm not talking about a comparison to human intelligence, just that it's a category error to treat a language model (generate some statistically sound responses to a prompt) as some kind of souped up expert system (generate *the* correct answer according to a model, for a given prompt). It's just fundamentally incapable of being that. Plus garbage in, garbage out. It's trained on random examples of human language, not facts.
       
 (DIR) Post #ATcAuTIKhFBpNBGZ3g by TedUnderwood@sigmoid.social
       2023-03-14T18:34:55Z
       
       0 likes, 0 repeats
       
       @dotstdy @pfessenbecker It does pretty well on AP tests for a system that is fundamentally incapable of generating correct answers. :)
       
 (DIR) Post #ATcBmwClQQI0fpdIiu by dotstdy@mastodon.social
       2023-03-14T18:44:47Z
       
       0 likes, 0 repeats
       
       @TedUnderwood @pfessenbecker you misunderstand what's happening when it gives correct answers to things...