Posts by chengyjonathan@sigmoid.social
(DIR) Post #ASS2TuA960HtVwzl0i by chengyjonathan@sigmoid.social
2023-02-07T23:19:04Z
0 likes, 1 repeats
I’ve said it once, but I’ll say it again, Dreambooth raises an interesting distinction between training an “object” and a “style.”This distinction drives the English phd parts of my brain into overdrive.“Idk, are these training images more about the objects in them or the way they’re portrayed? You tell me!”
(DIR) Post #ASZpCQB7i77baNM7Wa by chengyjonathan@sigmoid.social
2023-02-11T17:24:06Z
0 likes, 0 repeats
https://www.newyorker.com/tech/annals-of-technology/chatgpt-is-a-blurry-jpeg-of-the-webTed Chiang is a good writer, but the momentum of this piece comes from a strawman, setting enthusiasm for llms as predominantly flowing from fantasies of anthropomorphized AI. From there, it’s *easy* to dismiss llms as “lossy memorizers.” Experts are still learning what/how exactly llms learn, so it’s premature to try to package llms into a metaphor — one of the most “lossy” forms of figurative speech. 😂
(DIR) Post #ASZpCQkvYzlHNQ6iKe by chengyjonathan@sigmoid.social
2023-02-11T17:25:34Z
0 likes, 0 repeats
“Why use a metaphor when you have the original research?”“Oh idk, maybe because we sometimes want to see if, with a figurative transformation function, we can succinctly condense how we describe a thing?”
(DIR) Post #ASZpCRLRNEy7CfBsFE by chengyjonathan@sigmoid.social
2023-02-11T17:28:16Z
0 likes, 0 repeats
This piece feels like it came from a higher up pressuring “we need a story about chatgpt, have it on my table tomorrow!” *1920s American Accent*Everyone at the office is like, “are any of us really capable of writing this thing?” Many shrugs.
(DIR) Post #ASZpJNwWQ882gBqYk4 by chengyjonathan@sigmoid.social
2023-02-11T17:31:26Z
0 likes, 0 repeats
@TedUnderwood to complete the circle. The rebuttal would piece would need to be a shorter critique of chatgpt written by chatgpt. 🤣
(DIR) Post #AStacxYCdvzWSxXuYy by chengyjonathan@sigmoid.social
2023-02-21T06:21:07Z
0 likes, 0 repeats
@TedUnderwood yeah, and I think we’re going to start to see labor unions for various crafts coordinate their efforts in the next year or so. Either to simply resist adoption or generate larger regulations.
(DIR) Post #ATaIPWByHiVkhP7dE8 by chengyjonathan@sigmoid.social
2023-03-13T20:49:34Z
0 likes, 0 repeats
@TedUnderwood how do you mean by brainstorm visually? As in, brainstorm how words might guide reconstruct an image?
(DIR) Post #ATaIyxbSMZJdqewVTE by chengyjonathan@sigmoid.social
2023-03-13T20:56:00Z
0 likes, 0 repeats
@TedUnderwood ahhh, I’m finding that the text equivalent of that kind of brainstorming has been triplet building. Where I have an idea of a document, refining that idea by deciding what entities and relations I care about, transposing those entities and relations for different use cases/ styles.Granted, this has implicitly been the case for me since Bamman brought triplets into modeling fiction.
(DIR) Post #ATdy26ckrT8YPR5XVI by chengyjonathan@sigmoid.social
2023-03-15T15:20:06Z
0 likes, 0 repeats
I’m thinking about a paper from 2022 about how RL found faster matrix multiplication algorithms than the best human solutions 😂 I suspect RL will, surprisingly, have a comeback to bridge that exact gap. *takes off tin foil hat*
(DIR) Post #ATkMm4qNNOJe3afm4m by chengyjonathan@sigmoid.social
2023-03-18T17:25:35Z
0 likes, 0 repeats
@TedUnderwood @brohrer @codinghorror ahhh, for some reason I thought the lyrics had changed. Something about Wilford Brimley appearing had me focused on that part instead 😂
(DIR) Post #ATmSyE2W05KC5qRkZ6 by chengyjonathan@sigmoid.social
2023-03-19T16:13:36Z
0 likes, 0 repeats
@zoeleblanc @TedUnderwood out of curiosity, what’s stopping universities from setting up the infra to run Llama or Alpaca models — with an api to control access / allow researchers to hit for free?Granted it’s not gpt4, but for research purposes, there are some pretty viable alternatives that do text and code completion: https://github.com/tloen/alpaca-lora
(DIR) Post #ATmVI44Iyxfyrts6U4 by chengyjonathan@sigmoid.social
2023-03-19T18:10:26Z
0 likes, 0 repeats
@TedUnderwood @zoeleblanc yeah I agree wholeheartedly! I think universities have/should be thinking about the infra needed to host these alpaca style models for open access.Moreover, the research to build *better* alpaca style models for wider university research is worth its own organization. Now how to organize such a thing… Get the NEH on the phone! 😂
(DIR) Post #AUQ04DdHogLHYsPxrc by chengyjonathan@sigmoid.social
2023-04-07T19:29:01Z
0 likes, 0 repeats
@TedUnderwood Also, time to buy up all domains with words that are harmonious with GPT and AI