Post AU8wvmPWIfReaJqtu4 by nabsiddiqui@h-net.social
(DIR) More posts by nabsiddiqui@h-net.social
(DIR) Post #AU5i72YG8WcpV3eBfM by TedUnderwood@sigmoid.social
2023-03-29T00:33:36Z
0 likes, 0 repeats
Hey, @benmschmidt 's outfit NomicAI just released GPT4All, an open desktop-scale language model. I claim this as #DH. https://twitter.com/nomic_ai/status/1640834838578995202
(DIR) Post #AU5iJB1KTWhhEz21Ng by eliocamp@mastodon.social
2023-03-29T00:35:48Z
0 likes, 0 repeats
@TedUnderwood So is this a model trained on the output of another model?
(DIR) Post #AU5iMG9lprwe2TRxYW by TedUnderwood@sigmoid.social
2023-03-29T00:36:22Z
0 likes, 0 repeats
@eliocamp Indeed. That's how they're doing the Alpacas too. The technical term is RLAIF.
(DIR) Post #AU5ieNAeUmuv0PV1LE by afamiglietti79@mastodon.social
2023-03-29T00:39:38Z
0 likes, 0 repeats
@TedUnderwood @benmschmidt you can't just unilaterally declare something DH... (checks notes)... It appears you can! Carry on.
(DIR) Post #AU5isE4f75WoqaIrlQ by eliocamp@mastodon.social
2023-03-29T00:42:09Z
0 likes, 0 repeats
@TedUnderwood Huh.. Is this better than using actual human text? It seems that this can only lead to poorer outputs.
(DIR) Post #AU5j8OrYlk0qwOt45o by TedUnderwood@sigmoid.social
2023-03-29T00:45:04Z
0 likes, 0 repeats
@eliocamp It's not trained *only* on the model output; the base model here is LLaMA, and the model output from GPT 3.5 is used for "instruction tuning" to get the model to behave like a chatbot rather than a generic text-continuer. (E.g. if I say "1. List three shades of red. 2. List three shades of green" the model should do that and not say "3. List three shades of blue.") "Behaving like a chatbot" is an area where chatbots are kind of the experts.
(DIR) Post #AU5jD1KVgx6fP0dDvM by eliocamp@mastodon.social
2023-03-29T00:45:54Z
0 likes, 0 repeats
@TedUnderwood Ah, ok. So this is basically a way of emulating human-lead reinforcement learning but without using actual humans.
(DIR) Post #AU5jukxvtWTBOKvUFk by scott_bot@hcommons.social
2023-03-29T00:46:41Z
0 likes, 0 repeats
@afamiglietti79 @TedUnderwood @benmschmidt feels like that time the internet dubbed my text message article (https://www.vice.com/en/article/kzdn8n/the-route-of-a-text-message-a-love-story) DH, and I was like, what?
(DIR) Post #AU5julfXHPLdZZKJDU by TedUnderwood@sigmoid.social
2023-03-29T00:53:48Z
0 likes, 0 repeats
@scott_bot @afamiglietti79 @benmschmidt Anything you or Ben do is definitionally DH. If you rise to be Secretary of the Treasury, that will be DH.
(DIR) Post #AU5kDxE2XMQmutc9ZY by afamiglietti79@mastodon.social
2023-03-29T00:57:16Z
0 likes, 0 repeats
@TedUnderwood @scott_bot @benmschmidt (me, robbing a bank in 2042) "This money's got Scott's signature on it! It's DH! I claim it for my department!"
(DIR) Post #AU5vTtDsHllQdEL1yS by sramsay@hcommons.social
2023-03-29T03:03:24Z
0 likes, 0 repeats
@TedUnderwood @scott_bot @afamiglietti79 @benmschmidt I've been saying for years that DH is a big vault.
(DIR) Post #AU8wvmPWIfReaJqtu4 by nabsiddiqui@h-net.social
2023-03-30T13:44:16Z
0 likes, 0 repeats
@electricarchaeo @TedUnderwood @benmschmidt @simon Awesome stuff! How does the output of this compare to something like GPT4?
(DIR) Post #AU8wvmswXGys3ZcOlU by simon@fedi.simonwillison.net
2023-03-30T13:58:42Z
1 likes, 0 repeats
@nabsiddiqui @electricarchaeo @TedUnderwood @benmschmidt it doesn't compare well, sadly - GLT-4 is leagues ahead of any of the openly licensed models, at least for the moment - plus it appears way too large to ever run on a laptopI'm not looking for a GPT-4 I can run locally though: I want something much smaller that's powerful enough to run the ReAct pattern so I can plug in additional tools for looking up extra information, running calculations etc https://til.simonwillison.net/llms/python-react-pattern
(DIR) Post #AU94gSjTynEwqOOeZc by benmschmidt@vis.social
2023-03-30T15:27:24Z
0 likes, 0 repeats
@simon @nabsiddiqui @electricarchaeo @TedUnderwood Yeah, I've never seen any of these that are nearly as good as GPT3.5, let alone GPT4. IMO personal device LMs will involve fine-tuning on larger general-purpose transformers down to quantized weights for a specific purpose. In GPT4All we oversampled coding prompts specifically, e.g.. This strategy--take a big model, squash it into something that does a specific job--requires attention to training data above all, which is one Nomic interest here.