Post AcTO7O4M0L44SOiaS8 by carlmjohnson@mastodon.social
(DIR) More posts by carlmjohnson@mastodon.social
(DIR) Post #AcRDqmCtcUrOXREbWC by clive@saturation.social
2023-12-03T16:34:09Z
0 likes, 0 repeats
Want an LLM to run locally on your laptop?Go follow @simon's instructions on download's Mozilla's "llamafile"I just did, and it's quite remarkableGranted, I'm running it on a speedy new M2 mac, but the performance is amazinghttps://simonwillison.net/2023/Nov/29/llamafile/
(DIR) Post #AcRDy3hey4RcpS62aG by chris_hayes@fosstodon.org
2023-12-03T16:35:23Z
0 likes, 0 repeats
@clive @simon interesting, I didn't realize Mozilla was involved with LLMs, LAION has been the main FOSS group in the LLM landscape
(DIR) Post #AcREZO1AOwzEY31IFk by clive@saturation.social
2023-12-03T16:42:06Z
0 likes, 0 repeats
@chris_hayes @simon Yeah, I didn't know either!
(DIR) Post #AcRG5HYQ4JK1iqLXk0 by kittylyst@mastodon.social
2023-12-03T16:59:05Z
0 likes, 0 repeats
@clive @simon "If you don't care whether the answers are actually accurate or not, I can get them as fast as you like"
(DIR) Post #AcRPEI9o5x7VmoCmcC by clive@saturation.social
2023-12-03T18:41:32Z
0 likes, 0 repeats
@simon @kittylyst š
(DIR) Post #AcRPRy33ejaXTRZMpM by oschene@mastodon.social
2023-12-03T18:42:35Z
0 likes, 0 repeats
@clive @simon Considerably less creepy to use than the online models. Might actually learn something, this way.
(DIR) Post #AcRPW82GSwFjbhBNPU by simon@fedi.simonwillison.net
2023-12-03T18:43:41Z
0 likes, 0 repeats
@kittylyst @clive I'm not looking for a model that knows accurate facts about the world, I want a model I can run on my laptop which can do the "calculator for words" (and apparently now images) thingAnd now I've got one! https://simonwillison.net/2023/Apr/2/calculator-for-words/
(DIR) Post #AcRPsaGaBUo9XCaSzA by pdcawley@mendeddrum.org
2023-12-03T18:47:19Z
0 likes, 0 repeats
@simon @kittylyst @clive whatās the ācalculator for wordsā thing though?
(DIR) Post #AcRQki5zIAPRSBROOu by mjgardner@social.sdf.org
2023-12-03T18:57:28Z
0 likes, 0 repeats
@simon @kittylyst @clive I always thought of calculators as giving the same answer to the same inputs tomorrow as they do today, and whose operations I can not only understand but reproduce given enough time and patience.But silly me, I guess theyāre just magic boxes where numbers go in and other numbers come out, and I have to coax them and hope I can rely on their results.#LLMs #ChatGPT
(DIR) Post #AcRR7iocmVQ2LZIAxE by uberprutser@social.vivaldi.net
2023-12-03T19:02:46Z
0 likes, 0 repeats
@clive @simon This is probably going to spit you in your face :)
(DIR) Post #AcRkbSZ5fBNXlNv5gu by simon@fedi.simonwillison.net
2023-12-03T22:40:17Z
0 likes, 0 repeats
@pdcawley see linked post - i want a small LLM that can manipulate text - summarize, extract facts etc https://simonwillison.net/2023/Apr/2/calculator-for-words/
(DIR) Post #AcRlCibmyc6CEgwrbc by simon@fedi.simonwillison.net
2023-12-03T22:47:02Z
0 likes, 0 repeats
@mjgardner @kittylyst @clive yeah, I talk about that flaw in the analogy in my post: https://simonwillison.net/2023/Apr/2/calculator-for-words/#but-not-repeatable
(DIR) Post #AcSeS21CaXN2G19LvM by kittylyst@mastodon.social
2023-12-04T09:05:32Z
0 likes, 0 repeats
@simon @mjgardner @clive There's also this aspect that I specifically enoucnter in my technical writing: https://mastodon.social/deck/@kittylyst/111517513515068361
(DIR) Post #AcSzuynviP6FiyaRFI by clive@saturation.social
2023-12-04T13:07:24Z
0 likes, 0 repeats
@simon @kittylyst @mjgardner Maybe better metaphors would be in music?I believe it was Ian Bogost who, early on, analogized ChatGPT to a synthesizer for language ā ie a tool that lets you muck around with the wordstuff of the worldās expression, tweaking knobs to see what outputs one getsOr a guitar pedal ā something that takes an incoming signal (your prompt, as it were) ā and morphs it, accentuating elements, suppressing others 1/4
(DIR) Post #AcT00WoAiVgydLLUNk by clive@saturation.social
2023-12-04T13:07:25Z
0 likes, 0 repeats
@simon @kittylyst @mjgardner (Indeed, one of the things that make synthesizers and guitar pedals useful metaphors here is that, rather like LLMs, one needs to use them a lot to figure out how best to use them, and when not to use them ā¦When Iām using a pedal it changes *what* I choose to play and *how* I play it; I need to work with the affordances of that particular circuitry 2/4
(DIR) Post #AcT00xF0RMbcc66Hom by clive@saturation.social
2023-12-04T13:07:25Z
0 likes, 0 repeats
@simon @kittylyst @mjgardner For example, after playing with llamafile a bunch I am concluding, alas, that its ability to summarize is pretty weak and very low-resolution Iāll put in a long article as a prompt and then probe the model to see how it represents the information within ā¦ā¦ and it can only offer the most banal, obvious, 50,000-foot-in-the-air gist of what the article says When queried about specific elements in the piece ā¦3/4
(DIR) Post #AcT00zfJQWpw8GQK6C by clive@saturation.social
2023-12-04T13:07:25Z
0 likes, 0 repeats
@simon @kittylyst @mjgardner ā¦ it completely whiffs. Iām going to continue to experiment with it, but itās too bad ā good, nuanced and specific summarization (one that can rope in the language representations that of words) has always struck me as the part of LLMs Iām most interested inBut this one isnāt remotely there, near as I can tell 4/4
(DIR) Post #AcT22cXUa1fWtZfYnI by clive@saturation.social
2023-12-04T13:31:14Z
0 likes, 0 repeats
@kittylyst @simon @mjgardner I like that!
(DIR) Post #AcT28wMNndha6dUDtw by jhhl@toot.community
2023-12-04T13:32:25Z
0 likes, 0 repeats
@clive @simon It's pretty speedy and of course, being all-in-one means I don't have to chase down dependencies to get it to work. LLMs have to digest a vast number of sentences in order to get an idea of what grammar is, and as a side effect, palimpsests of those training sentences remain in the data. I'd like to familiarize it with my own writing, I guess I have to find a trainer that can build a compatible checkpoint.
(DIR) Post #AcT4f0uId48rNplw8G by markhurst@mastodon.social
2023-12-04T14:00:29Z
0 likes, 0 repeats
@clive @simon @kittylyst @mjgardner I see LLMs as spackle:https://creativegood.com/blog/23/ai-is-spackle.html
(DIR) Post #AcTIAaOG3IPPoWAbse by simon@fedi.simonwillison.net
2023-12-04T16:30:43Z
0 likes, 0 repeats
@kittylyst @mjgardner @clive that kind of thing is why I think generating text for publication is one of the least interesting applications of this technology (the term "generative AI" has started to feel pretty misleading to me)
(DIR) Post #AcTIHuK2LOn7etgeKe by simon@fedi.simonwillison.net
2023-12-04T16:32:09Z
0 likes, 0 repeats
@clive @kittylyst @mjgardner wow, that really is an excellent metaphor - I love that it hints at how hard they are to really use effectively, despite being easy to start getting them to spit out weird noises
(DIR) Post #AcTIpK6tUBQNC9TPVo by abucci@buc.ci
2023-12-04T14:59:41Z
0 likes, 0 repeats
@simon@simonwillison.net @mjgardner@social.sdf.org @kittylyst@mastodon.social @clive@saturation.social I don't know that this post really addresses the flaw in the analogy. Large language models are nothing like calculators, in my view.The important fact about calculators isn't that they give you repeatable answers. Though obviously that's an important feature in some cases, there are calculators that let you sample from a probability distribution. No, what's important about calculators is that if you "ask" a calculator a "question", you can rely on its output to be a "reasonable" answer that you'll understand. I've quoted those words because I'm being vague on purpose, but I think we could all fill them in with precise definitions for a calculator.If you "ask" a calculator "what's 2+2?", it's not going to spit out Wikipedia-style text about the history of arithmetic, or some poor random person's PII. It will give you the expected arithmetic result.To put it differently, a calculator has a well-defined, well-scoped set of use cases, a well-defined, well-scoped user interface, and a set of well-understood and expected behaviors that occur in response to manipulations of that interface.Language models have none of those qualities. They have an open-ended set of unspecified use cases. Sam Altman seems to believe you can do almost anything at all with them, for instance; Timnit Gebru has argued they're trying to "be God". The user interface is fairly well scoped, I guess, but it's extremely impoverished relative to comparably-complex tools. The outputs are not at all well understood, even by the people who make these models.I'd add, as an aside, that the underlying architecture and engineering principles of calculators are almost fully transparent. Undergraduate computer science and electronics engineering students, among others, are taught all the principles they need to know to build a calculator. By contrast, the underlying architecture of GPT has never been published, despite the "open" in OpenAI's name; nor, to my knowledge, has their training procedure. There are increasingly more open alternatives, which is nice to see, but no one really knows what's going on inside a deep neural network (and anyone who claims to is either naive or trying to mislead you). They are deep black boxes. To the extent these tools are based on deep neural networks and variations thereof, they are not transparent.To sum up, the analogy between large language models and calculators only appears to be valid in the most surface, superficial reading, and does not hold up to scrutiny. Which suits the topic I suppose, since the outputs of large language models are also the most surface, superficial expression of language and do not hold up to scrutiny either.
(DIR) Post #AcTIpLNEmu5l787ycq by simon@fedi.simonwillison.net
2023-12-04T16:35:58Z
0 likes, 0 repeats
@abucci @simon @mjgardner @kittylyst @clive I agree with everything you said thereI addressed the flaw by agreeing it was a flaw and punting with "All analogies are imperfect, but some are more imperfect that other"
(DIR) Post #AcTJoErDd874ZIBixU by simon@fedi.simonwillison.net
2023-12-04T16:48:42Z
0 likes, 0 repeats
@abucci @mjgardner @kittylyst @clive mind if I link to this from my blog post? It's a really great counter-argument
(DIR) Post #AcTO7O4M0L44SOiaS8 by carlmjohnson@mastodon.social
2023-12-04T17:38:05Z
0 likes, 0 repeats
@clive @simon my riff on Simonās calculator was/is ākaleidoscope for wordsā. It doubles as a 1984 reference
(DIR) Post #AcTgrxwojvi8EBYLuy by clive@saturation.social
2023-12-04T21:08:39Z
0 likes, 0 repeats
@simon @carlmjohnson I like that!
(DIR) Post #AcTh2JtpygnBCr62PA by clive@saturation.social
2023-12-04T21:10:30Z
0 likes, 0 repeats
@mjgardner @simon @kittylyst Yep yep it can take me months or even years to fully explore the expressive qualities that a particular guitar pedal adds to (and subtracts from!) my playing
(DIR) Post #AcThpdRY6TAs0fNcrw by brianstorms@mastodon.social
2023-12-04T21:19:23Z
0 likes, 0 repeats
@clive @mjgardner @simon @kittylyst My kingdom for an AI that can listen to any musical recording and, knowing what effects pedals, sound libraries, and AU plugins I have, tell me which one or which combo to use to replicate any sound. Or, just whips up an AU instrument on the fly that makes that exact sound.
(DIR) Post #AcUQGQaEQ7hxvY1kTA by clive@saturation.social
2023-12-05T05:37:24Z
0 likes, 0 repeats
@simon @mjgardner @kittylyst @markhurst I like that!
(DIR) Post #AcVgwsODDkuSkmd0GO by abucci@buc.ci
2023-12-05T16:12:55Z
0 likes, 0 repeats
@simon@simonwillison.net Absolutely, please feel free. I wrote up these ideas (with a shout out to datasette!) on my blog, which might work a bit better: https://bucci.onl/notes/Word-calculators-dont-add-up I put additional background in there to hopefully clarify and provide a more interesting/useful read. @mjgardner@social.sdf.org @kittylyst@mastodon.social @clive@saturation.social
(DIR) Post #AcVgwtNXXjxxozJyr2 by simon@fedi.simonwillison.net
2023-12-05T20:17:51Z
0 likes, 0 repeats
@abucci @mjgardner @kittylyst @clive Nice, linked to that from here: https://simonwillison.net/2023/Apr/2/calculator-for-words/#update-dec-5