Post AS8JQxtNaXWDG8HY12 by bornach@masto.ai
(DIR) More posts by bornach@masto.ai
(DIR) Post #AS7TGvHxDc34fifPuq by simon@fedi.simonwillison.net
2023-01-29T00:30:05Z
1 likes, 1 repeats
One of the things I'm finding so interesting about large language models like GPT-3 and ChatGPT is that they're pretty much the world's most impressive party trickAll they do is predict the next word based on previous context. It turns out when you scale the model above a certain size it can give the false impression of "intelligence", but that's a total fraudIt's all smoke and mirrors! The intriguing challenge is finding useful tasks you can apply them to in spite of the many, many footguns
(DIR) Post #AS7X7eoJVvQcmCdkkS by dvd@dair-community.social
2023-01-29T01:53:51Z
0 likes, 0 repeats
@simon a finite improbability machine
(DIR) Post #AS7XYzfk3gi9ISfLDE by sofia@chaos.social
2023-01-29T01:58:43Z
0 likes, 0 repeats
@simon "there is a mechanism so it's fake"
(DIR) Post #AS7YC4nIwtSsTZq5tg by mikemol@pony.social
2023-01-29T02:04:11Z
0 likes, 0 repeats
@simon I think you're fundamentally right, but I also think it's little more than that; it's a lot more complicated than a simple Markov bot. It's not like GPT-3 is a 4000-wide expansion of the Markov bot that hung out in your IRC channel 15 years ago. That's more or less representable by a single node with 4k inputs and 4k outputs.This stuff does a lot of explicit language conceptual and semantic modeling, and is more like thousands (millions? billions?) of Markov models chained together.That just means it's going to be better at semantically and syntactically correct output, though, not that it's AGI.You're definitely right that now we have to figure out how to apply it safely; it's like dropping 2022 C++ in the laps of 1955 software engineers without an existing body of work studying that level of flexibility and evolving best practices.Gonna be an interesting and painful next couple decades.
(DIR) Post #AS7ZmDKemZ9LKqEdYO by jamescham@mastodon.social
2023-01-29T02:23:25Z
0 likes, 0 repeats
@simon That's right! It is interesting to watch folks building complicated applications with layers and layers of prompts to deal with some of these problems.
(DIR) Post #AS7aEB4InCNcQFv3nk by williamgunn@mastodon.social
2023-01-29T02:28:42Z
0 likes, 0 repeats
@simon This technology has finally drove the stake into the Q&A website business, unfortunately for me and the company that was until very recently my employer. There's still a niche for a place that focuses on quality and personal experience & there's a lot of refugees from the platforms that are being overrun with generated content.
(DIR) Post #AS7clMuhyNbmcmO0vo by codinghorror@infosec.exchange
2023-01-29T02:57:10Z
0 likes, 0 repeats
@simon yes, “google autocomplete for everything else”
(DIR) Post #AS7eiN0RN9kbx0lWPA by chriswerry@mastodon.nz
2023-01-29T03:18:43Z
0 likes, 0 repeats
@simon it will write regex for you if that’s something you need
(DIR) Post #AS7iTvH650pxhpvHhA by evannakita@mastodon.online
2023-01-29T04:00:26Z
0 likes, 0 repeats
@simon It’s effectively just a polished version of the suggested words that show up at the top of phone keyboards.
(DIR) Post #AS7ijo2yihoTvHtPwO by simon@fedi.simonwillison.net
2023-01-29T04:04:01Z
0 likes, 0 repeats
@evannakita yeah that's my absolutely favourite way to explain how it works to people - it's the iPhone suggested keyboard just with an unimaginably larger training set!
(DIR) Post #AS7q2WWYnEzSl3bMrg by trindflo@hachyderm.io
2023-01-29T05:25:40Z
0 likes, 0 repeats
@simon This has been the case practically since programming started being performed on microcomputers. It is a pipe dream that comes up from time to time that you can get computers to program themselves and not need to deal with those horrible programmers. Sort of like cold fusion and other perpetual motion machines, it is a pipe dream where you get something for nothing.
(DIR) Post #AS7rf9czxEFEzjb6ae by ph1@hachyderm.io
2023-01-29T05:43:44Z
0 likes, 0 repeats
@simon do you know of a good explainer of how these LLMs actually work, for someone who’s an experienced software engineer but doesn’t have any data science/ML background (i.e. me). Most content is either very surface level or uses a lot of words that I don’t understand!
(DIR) Post #AS7sJ7xZZl5RP8HWgy by simon@fedi.simonwillison.net
2023-01-29T05:51:15Z
0 likes, 0 repeats
@ph1 this is a couple of years old now but I found it useful https://jalammar.github.io/how-gpt3-works-visualizations-animations/You might find something useful in the gpt3 tag on my blog https://simonwillison.net/tags/gpt3/
(DIR) Post #AS870BrGMB23Ixl54y by zhksh@sigmoid.social
2023-01-29T08:35:33Z
0 likes, 0 repeats
@simon its understandable but remains annoying that once people realize its "just next word prediction" they conclude thats all they need to know and call it a hoax. There is actually a huge body of research on what LMs actually learn, can learn in theory etc.
(DIR) Post #AS87uMIGkDZtt3Aob2 by edaross@neurodifferent.me
2023-01-29T08:45:52Z
0 likes, 0 repeats
@simonThey really should have a message like that much more prominently displayed when you start using it. I've enjoyed making things with it https://blog.edross.co.uk/archive/tagged/chatgpt but you have know what it is doing. It has also been useful for little JavaScript based tools at work for doing things like converting a web page exported from Excel so that it is interactive again. @ZhiZhu
(DIR) Post #AS8DiPPtuGcPp5QcCW by mikiobraun@hachyderm.io
2023-01-29T09:50:30Z
0 likes, 0 repeats
@simon that‘s the machine‘s version of „fake it till you make it.“
(DIR) Post #AS8FMi4c3pURDtFg9I by Octale@nerdculture.de
2023-01-29T10:09:16Z
0 likes, 0 repeats
@simon journalism, obviously. I still don’t understand why CNET shut their AI journalist down for getting basic facts wrong. Getting basic facts wrong is a journalistic tradition at least 25 years old.
(DIR) Post #AS8Fct5grlJi1gTCa0 by teixi@mastodon.social
2023-01-29T10:12:12Z
0 likes, 0 repeats
@simon The most difficult in order to make such kind of models usable in practice, is to filter the plausible yet totally surreal inferences from the reality grounded ones.ie: some of the most difficult ones are the plausible negation of things. Like it is/was not exactly in this way vs. a totally non-sense negation to disregard.
(DIR) Post #AS8HqkIeXzACixGjuC by faassen@fosstodon.org
2023-01-29T10:37:25Z
0 likes, 0 repeats
@simon People point this out a lot. It's true.It implies that our own human minds do something fundamentally different. I am not convinced that's true.Our minds evolved. That guarantees the use of tricks. Different tricks, yes: humans are interested in far more than predicting text; we are grounded in a very different experience. But tricks.I think some cognitive dissonance is at play in the widespread, strident insistence that these large language models are just a trick.
(DIR) Post #AS8I2dZTOFWxKXo4wa by ami@floss.social
2023-01-29T10:38:10Z
0 likes, 0 repeats
@simonHmm, I thought maybe we would be getting rid of it soon, but another thing that was "just the worlds most impressive party trick": electricity! we've since found other uses for that one at least, and it kinda stuck around.
(DIR) Post #AS8JQwjPu5xHewc4qe by dbreunig@note.computer
2023-01-29T01:21:34Z
0 likes, 0 repeats
@simon It does the trick performers (street performers, YouTube performers, magicians, etc) have done for ages: invest more time into a trick than any reasonable human would expect.
(DIR) Post #AS8JQxN7WTiLe5BmjY by simon@fedi.simonwillison.net
2023-01-29T01:31:02Z
0 likes, 0 repeats
@dbreunig I've seen wonderful material from magician Teller about that
(DIR) Post #AS8JQxtNaXWDG8HY12 by bornach@masto.ai
2023-01-29T10:55:12Z
0 likes, 0 repeats
@simon @dbreunig"Make the secret a lot more trouble than the trick seems worth"https://www.smithsonianmag.com/arts-culture/teller-reveals-his-secrets-100744801/
(DIR) Post #AS8KQsSrrsNWNcP5BA by sonjdol@ohai.social
2023-01-29T11:06:11Z
0 likes, 0 repeats
@simon it's an argument Simone Natalie makes in his book, too. It's all about deception https://global.oup.com/academic/product/deceitful-media-9780190080372?cc=de&lang=en&
(DIR) Post #AS8Ma0oYOQ6in61wvI by thiemo@phpc.social
2023-01-29T11:30:12Z
0 likes, 0 repeats
@simon https://phpc.social/@thiemo/109683178535094425
(DIR) Post #AS8MtW78FIKpMp1MES by bornach@masto.ai
2023-01-29T11:33:53Z
0 likes, 0 repeats
@simon @ph1I've found the articles explaining transformers on the Towards Data Science website quite accessiblehttps://towardsdatascience.com/transformers-141e32e69591
(DIR) Post #AS8R1NXAQhg4tECdeK by martinlindner_wb@colearn.social
2023-01-29T12:20:08Z
0 likes, 0 repeats
@simon (das ist sehr irreführend. mal abgesehen von den zahlreich fehlleistungen, die es da gibt: der unterschied zu dem, wie "kollektive humane intelligenz" texte in schriftsprache hervorbringt, ist nicht so einschneidend. in gewisser hindicht ist es quasi der diskurs, nicht die "autoren", der texte und aussagen erzeugt.)
(DIR) Post #AS8aU8Ip2Oy5dUHMSe by feynman@tooted.ca
2023-01-29T14:03:48Z
0 likes, 0 repeats
@simon I also don't like the bias they are doing after fine tuning. You can see some subjects are giving scripted responses on the public version.
(DIR) Post #AS8i77LbDqaje879cW by simon@fedi.simonwillison.net
2023-01-29T15:31:52Z
0 likes, 0 repeats
@zhksh yeah I've been following as much of that research as I can with huge interest!That's why I think it's the worlds most impressive party trick: it turns out next word prediction can do a mind boggling array of things that no one expectedIt's crucial though that people understand that it's not an "AI" in the same way that science fiction AIs are intelligent - telling people its a really powerful version of an iPhone keyboard I think helps them understand the many limitations more
(DIR) Post #AS8imNkX90SIODSBcm by simon@fedi.simonwillison.net
2023-01-29T15:36:49Z
0 likes, 0 repeats
@ami that's a useful comparison - just because something is a party trick doesn't mean it can't have a lasting impact on the world!
(DIR) Post #AS8l0S4i3inRlcjAUy by simon@fedi.simonwillison.net
2023-01-29T16:04:08Z
0 likes, 0 repeats
And in case this post wasn't clear: I'm all-in on large language models: they confidently pass my personal test for if a piece of technology is worth learning:"Does this let me build things that I could not have built without it?"What I find interesting is that - on the surface - they look like they solve a lot more problems than they actually do, partly thanks to the confidence with which they present themselvesFiguring out what they're genuinely good for is a very interesting challenge
(DIR) Post #AS8mA70PLiJBGtsy8W by wichitalineman@hachyderm.io
2023-01-29T16:14:38Z
0 likes, 0 repeats
@simon Ha! Super true. People incorrectly put faith in ChatGPT for the same reason they incorrectly put faith in me when I'm on pub quiz teams. We just can't help but speak with undue confidence. I'd imagine that ChatGPT would also be a frustrating quiz team mate.
(DIR) Post #AS8mNa4sZT1lCZOpX6 by danbruder@toot.cafe
2023-01-29T16:16:51Z
0 likes, 0 repeats
@simon that’s a great way to put it.
(DIR) Post #AS8nFshvWQHQcPYXiK by ncweaver@thecooltable.wtf
2023-01-29T16:27:04Z
0 likes, 1 repeats
@simon The way I look at it. Machine learning in general (including these large language models) are great when you have the following problem criteria#1: You need to build a pattern matcher#2: You don't know what to look for.#3: When the pattern matcher is finally built you don't care to know what it actually looks for#4: The results are allowed to be hilariously, insanely wrong some % of the timeAnd there are actually a lot of things that match that criteria
(DIR) Post #AS8nSuNec4ABZWHlIG by dbreunig@note.computer
2023-01-29T16:26:23Z
0 likes, 0 repeats
@simon The applicable problem space is even fuzzier because there’s some things they do well, but not consistently. Which to me means the problem space has to be limited to areas where the cost of a false-positive/negative is low.
(DIR) Post #AS8nSux6UGWHLSs4Y4 by simon@fedi.simonwillison.net
2023-01-29T16:29:22Z
0 likes, 0 repeats
@dbreunig right - there are so many potential applications where you might get good results 90% of the time and utter garbage 10% of the time, which for things like loan application evaluation should be completely unacceptable, but might be fine for things like bulk sentiment analysis to identify general trendsThere's a lot of depth just to learning how to identify places that the technology is a good fit
(DIR) Post #AS8neqOzEMvMhwbH6m by simon@fedi.simonwillison.net
2023-01-29T16:30:48Z
0 likes, 0 repeats
@ncweaver I really like that criteria listIt would be really useful if there was a solid, easy to understand list of use-cases and anti-use-cases to point people to
(DIR) Post #AS8oH9P71vCye5vPXM by pamelafox@fosstodon.org
2023-01-29T16:33:24Z
0 likes, 0 repeats
@ncweaver @simon im totally on board with ML if we acknowledge #4- also why I'm uncomfortabe with ML being used for 1:1 therapy situations at this point.
(DIR) Post #AS8oHA5eTlEgm1pNqK by simon@fedi.simonwillison.net
2023-01-29T16:40:31Z
0 likes, 0 repeats
@pamelafox @ncweaver the 1-1 therapy thing is terrifying to me - imagine trying to get therapy from your iPhone keyboard!But that said, I do use ChatGPT as an alternative to rubber duck programming sometimes: if I'm stuck on something I'll kick off a conversation purely as a thinking aid, and it's often effectiveThat feels pretty different to me from the therapy thing, but not a million miles away from it
(DIR) Post #AS8oU4AFp2GO9y8M5I by simon@fedi.simonwillison.net
2023-01-29T16:42:45Z
0 likes, 0 repeats
@wichitalineman Hah, I really love that model of ChatGPT where it's a pub quiz team member who's often right, sometimes wrong but expresses complete confidence in their answer either way!
(DIR) Post #AS8oewcaCvP1x0ZUoa by mterhar@bfd.so
2023-01-29T16:44:40Z
0 likes, 0 repeats
@simon Dunning-Krueger as a service.
(DIR) Post #AS8oqwLH4Td4YmCB84 by dvogel@mastodon.social
2023-01-29T16:46:26Z
0 likes, 0 repeats
@simon you'd likely enjoy this episode of the Ezra Klein podcast. There's an excellent tangent re: "bullshit" as it pertains to ChatGPT et al.https://www.nytimes.com/2023/01/06/opinion/ezra-klein-podcast-gary-marcus.html
(DIR) Post #AS8rNah2mXk8sj4hQO by simon@fedi.simonwillison.net
2023-01-29T17:13:24Z
0 likes, 0 repeats
@dbreunig one area I'm particularly excited about it data extraction: given a huge jumble of badly OCRd documents, can a language model be used to ask questions of each scanned page to extract relevant facts from them?If you applied data entry people to a task like that you'd also get a portion of errors
(DIR) Post #AS8s58A6DpgR8QQVvc by dbreunig@note.computer
2023-01-29T17:22:11Z
0 likes, 0 repeats
@simon Also agree. Bureaucracy—where forms and formulas have tried to interface with humans as if they were APIs—and it’s artifacts is where the potential energy lies. Anywhere there is rote tasks and boredom is the X marking the spot.I call it the ‘Brazil Antidote’ model. Doesn’t get enough attention because it’s not sexy. I’m going to found the Boring AI Working Group.
(DIR) Post #AS8sSpoI6ZgfATWr1U by simon@fedi.simonwillison.net
2023-01-29T17:25:31Z
0 likes, 0 repeats
@dbreunig I find it pretty interesting that a while ago web scraping was rebranded "robotic process automation" and quietly became a multi-billion dollar market https://en.m.wikipedia.org/wiki/Robotic_process_automation
(DIR) Post #AS8vdjSmLKG8WzM3O4 by Sandra@idiomdrottning.org
2023-01-29T17:45:16.712515Z
0 likes, 0 repeats
@simon A missing 🧩 here is that making large language models is a huge destructive climate impact and we should quit it until we've got the climate sitch under control.Using them is the same as any other app, so end-users and API-users, don't feel guilt-tripped. Making them, on the other hand, wrecks the world 💔
(DIR) Post #AS8vdk1sEqKeHpm55c by simon@fedi.simonwillison.net
2023-01-29T18:03:16Z
0 likes, 0 repeats
@Sandra I've not found that argument very convincing yetSure, there's a HUGE energy cost in training a model... but that model can then be put to use for many years into the futuretext-davinci-003 was trained once, at great expense... but has since run inference millions (probably billions) of times for millions of peopleThis looks even better for openly released models like Stable Diffusion: trained once, then distributed to anyone who wants to use it
(DIR) Post #AS8vqPnWllsrIx6iqu by simon@fedi.simonwillison.net
2023-01-29T18:05:29Z
0 likes, 0 repeats
@Sandra I remember being amused when I saw one model that was trained by a university in France that boasted that 90% of the power used to train that model came from a nearby nuclear reactor!Wish I could find that reference now
(DIR) Post #AS8wJ1alqRl8d6maGG by apkoponen@mstdn.social
2023-01-29T18:06:56Z
0 likes, 0 repeats
@simon This is exactly what I was thinking today in the car when my wife asked: ”What’s on your mind?”
(DIR) Post #AS907E8nlKBntJKXOy by pamelafox@fosstodon.org
2023-01-29T18:29:39Z
0 likes, 0 repeats
@simon @ncweaver yeah to be fair, I'm totally gonna try chatgpt for therapy-light, like social situation advice, but I'm emotionally stable enough and am aware that chatgpt3 is just an LLM. my concern is for folks who may be close to harming themselves or others, and an overly humanized chatbot tells them something that guides them awry.
(DIR) Post #AS91lOETlGHaj8hVdA by vaurorapub@infosec.exchange
2023-01-29T19:08:00Z
0 likes, 0 repeats
@simon one of my favorite applications of chatbots is NPCs in video games. I hate when they repeat the exact same set of sentences
(DIR) Post #AS9A1FAs0SmGts5EQa by simon@fedi.simonwillison.net
2023-01-29T20:43:01Z
0 likes, 0 repeats
@wim_v12e @Sandra retraining entire models I've been exploring alternatives to re-training the entire model to bake in new facts through mixing in results from other systems directly into the prompt - it's a really promising avenue: https://simonwillison.net/2023/Jan/13/semantic-search-answers/The cost of using them is definitely enormous - I've seen reports of ChatGPT costing $3,000,000/day or more - but again, that's spread across many usersAt least it's not Bitcoin mining!
(DIR) Post #AS9AQovoCR6T2REb8y by simon@fedi.simonwillison.net
2023-01-29T20:44:19Z
0 likes, 0 repeats
@wim_v12e @Sandra I remain hopeful that some day in the future it will become possible to run a large language model on a personal device - for both energy and privacy reasonsI can run Stable Diffusion on my iPhone already, but that's a MUCH smaller model that the various LLMs
(DIR) Post #AS9CFqClKB6Li83Zsu by Sandra@idiomdrottning.org
2023-01-29T21:06:41.145675Z
0 likes, 0 repeats
@simon Talking about making the models, not running them
(DIR) Post #AS9CFqmDCNSRU4dt8i by simon@fedi.simonwillison.net
2023-01-29T21:09:29Z
0 likes, 0 repeats
@Sandra I was replying to @wim_v12e who said "And also, this totally ignores the cost of using these things at scale, which dwarfs the cost of training them."
(DIR) Post #AS9HixrqPheNe1ylQ8 by plznuzz@pawb.fun
2023-01-29T22:10:28Z
0 likes, 0 repeats
@simon @dbreunig The use case you describe is commonly done with BERT, which is another language model but I think it works in a different way. I would be slightly curious to see how ChatGPT compares, since BERT was more directly designed for that but ChatGPT is newer and larger.
(DIR) Post #AS9Hupi2ZT7tol3KK0 by simon@fedi.simonwillison.net
2023-01-29T22:11:57Z
0 likes, 0 repeats
@plznuzz @dbreunig I'd love to read more about using BERT for this kind of project - I wouldn't know where to even start with that right now
(DIR) Post #AS9IAErdKRFCB3syMS by teixi@mastodon.social
2023-01-29T22:15:41Z
0 likes, 0 repeats
@simon Humans readout is what gives rational interpretation of prompt responses that had been modeled for looking plausible.Useful when those given responses are factual and accurate. Otherwise misleading interpretation and thus requiring critical filtering.This is a critical failure of current models, so to be improved in future models.
(DIR) Post #ASAxtrpRb1IjKmhJLs by patriciajhawkins@mastodon.art
2023-01-30T17:37:06Z
0 likes, 0 repeats
@simon Yeah -- It's the world's biggest, best supplied predictive text game.
(DIR) Post #ASAyawFJ36SJUBKfLs by patriciajhawkins@mastodon.art
2023-01-30T17:45:17Z
0 likes, 0 repeats
@simon Research. It knows how to give properly referenced citations, so you ask it to write an article summarizing the research on the pros and cons of xyz topic, listing your concerns. And then you use that as a well-structured summary for digging into the articles -- and authors -- that it found. Maybe you should list it as a co-author? Or just mention using it as a research technique. Rather like using a reference librarian, though. Or an old-fashioned card-catalogue.