Post AXFq4HSwBDQOBJcNfM by UlrichJunker@fediscience.org
 (DIR) More posts by UlrichJunker@fediscience.org
 (DIR) Post #AXDSpecz5y1KPaSDnk by TedUnderwood@sigmoid.social
       2023-06-30T11:56:21Z
       
       0 likes, 0 repeats
       
       “A grad student who fell asleep in 1982 and woke up in 2022 might see large language models as a triumph for cultural theory.” My contribution to the debate this week in the CI blog. #LLM #Foucault  https://critinq.wordpress.com/2023/06/29/the-empirical-triumph-of-theory/
       
 (DIR) Post #AXF05gMtsnL05Z8pQO by markrstoll@masto.ai
       2023-07-01T05:43:43Z
       
       0 likes, 0 repeats
       
       @TedUnderwood I've had some students like that.
       
 (DIR) Post #AXFBI3CptiRDtCXBDM by UlrichJunker@fediscience.org
       2023-07-01T07:49:13Z
       
       0 likes, 0 repeats
       
       @TedUnderwood you are referring to #RLHF (reinforcement learning by human feedback) as a way of correcting transformer output by human authors. But this technique also covers learning preferences from humans and this aspect hasn’t found much attention in the debate of #LLMs, but may rather be determining for ChatGPT’s success. What is your opinion about this? https://proceedings.neurips.cc//paper_files/paper/2022/hash/b1efde53be364a73914f58805a001731-Abstract-Conference.html
       
 (DIR) Post #AXFaQc2gl5lp6Ry1dw by TedUnderwood@sigmoid.social
       2023-07-01T12:30:54Z
       
       0 likes, 0 repeats
       
       @UlrichJunker In that piece I think I’m talking about instruction tuning rather than RLHF as such? It was an earlier advance, although the purposes are similar. But I agree with you that this whole topic is under-discussed. One way to put it is that the models responded to / addressed the Stochastic Parrots critique that they weren’t grounded in a communicative situation. But it served no one’s polemical purpose to take note of that.
       
 (DIR) Post #AXFq4HSwBDQOBJcNfM by UlrichJunker@fediscience.org
       2023-07-01T15:26:09Z
       
       0 likes, 0 repeats
       
       @TedUnderwood yes that is a good point. Adding preferences to a system may make a system’s behavior more satisfactory to a user, but does not change the nature of this behavior. And the debate is about the latter.
       
 (DIR) Post #AXZmgWq4YmZpVDxZjs by cr@sigmoid.social
       2023-07-11T06:22:28Z
       
       0 likes, 0 repeats
       
       @TedUnderwood Finally went around to reading this, good stuff!