|
Shame Boy posted:when i first got my new laptop that had one of those i was like "oh can i run dall-e on that, it'd be fun to run the bigger/better model locally just to play around with" and the answer at the time was "lol no" Apple released an update to tensorflow-metal that targeted Stable Diffusion performance, and that's still using the GPU proper, not the Neural Engine. So the proprietary stuff might simply be useless for running anything non-proprietary Internet Janitor posted:i wonder what's going to happen to the lag time between "this cutting-edge llm runs on insanely expensive servers and must take place offsite" and "this llm has been crunched down to run on a current-gen apple silicon mac" over the next few years I'd say Alpaca-30B (the LLaMa leak finetuned on OpenAI API responses ) on M1 is comparable with text-curie-001 (instruct-style prompts totally work how you'd expect with GPT-3) and it's faster locally than online GPT-4 inference, so right now that's... 14 months? Assuming we're counting from the moment Curie got the instruct layer update
|
# ? Mar 31, 2023 12:49 |
|
|
# ? Apr 27, 2024 14:22 |
|
yeah, nobody has proven to have an actual durable edge
|
# ? Mar 31, 2023 12:52 |
|
Speaking of use cases, I watched my manager request a summary of the bullshit and interminably long "sparks of AGI" (PR) paper from chatgpt. The openAI version that has no internet access and had to go off the provided title alone. And received a short bit of totally different bullshit text with no relation to the original. Regardless of that fact, ChatGPT's response was accepted by him and the rest of the team as a fair summary. I'm still mulling over the proper lessons to learn having seen that.
|
# ? Mar 31, 2023 13:30 |
|
accuracy and veracity are so unimportant as to be completely irrelevant, as long as the output looks like what you are expecting so yeah, expect that as a rule people will accept whatever gpt spits out as truth or at least certainly close enough that they'll use it to form opinions and make decisions anyway
|
# ? Mar 31, 2023 13:38 |
|
that’ll facilitate new levels of telephone game previously not thought possible
|
# ? Mar 31, 2023 14:37 |
|
jemand posted:Speaking of use cases, I watched my manager request a summary of the bullshit and interminably long "sparks of AGI" (PR) paper from chatgpt. The openAI version that has no internet access and had to go off the provided title alone. And received a short bit of totally different bullshit text with no relation to the original. Regardless of that fact, ChatGPT's response was accepted by him and the rest of the team as a fair summary. Corporations are organisms, and just like in mammals, the signal to the muscles was sent before the situation was consciously assessed. The rest is rationalization
|
# ? Mar 31, 2023 14:43 |
|
izagoof posted:that’ll facilitate new levels of telephone game previously not thought possible i wonder what does happen when you have chatgpt continuously rephrase a text sample.
|
# ? Mar 31, 2023 14:44 |
|
idk someone's found at least one use for chatgpt it seems
|
# ? Mar 31, 2023 15:22 |
|
lol
|
# ? Mar 31, 2023 15:23 |
|
I think there’s interesting business applications of llms in the observability (or really, any data processing) field around natural language querying of near-real time ops data (“show me the top ten slowest sql queries in us-east-1”, “what’s the most popular item being sold in Jakarta”, etc) but there’s pretty significant costs in training and upkeep that can’t really be shared (I can’t train my model off someone else’s telemetry) and are constant (if I’m not constantly updating the model with changes to the system then it becomes less useful) there’s also some negative externalities? like, does the model ossify the system state or whatever
|
# ? Mar 31, 2023 15:24 |
|
jemand posted:Speaking of use cases, I watched my manager request a summary of the bullshit and interminably long "sparks of AGI" (PR) paper from chatgpt. The openAI version that has no internet access and had to go off the provided title alone. And received a short bit of totally different bullshit text with no relation to the original. Regardless of that fact, ChatGPT's response was accepted by him and the rest of the team as a fair summary. man, the new version of betteridge's law of headlines is gonna suck
|
# ? Mar 31, 2023 15:28 |
|
Powerful Two-Hander posted:idk someone's found at least one use for chatgpt it seems
|
# ? Mar 31, 2023 15:34 |
|
Chalks posted:It'll be interesting to see what the outcome is but I wouldn't be surprised if a set of neutral network weights qualifies as fair use, given that Google image search was fair use when its product actually contains the images in question. didn't google get sued over the button that let you directly see the iamge file
|
# ? Mar 31, 2023 15:36 |
|
Shame Boy posted:he's an "AI researcher" in that he writes lots and lots of words about how scary it would be if we made the robot devil, and how you should give him and his organization money so we can make robot god before that happens to fight the robot devil he is so loving stupid
|
# ? Mar 31, 2023 15:37 |
|
Improbable Lobster posted:he is so loving stupid very rude to say this about a person who has the superpower to literally rewire his brain, ONCE, in order to perfectly execute a single apparently-impossible task. no he's never used this superpower, but it definitely exists. no you can't see it either
|
# ? Mar 31, 2023 15:41 |
|
Grace Baiting posted:very rude to say this about a person who has the superpower to literally rewire his brain, ONCE, in order to perfectly execute a single apparently-impossible task. no he's never used this superpower, but it definitely exists. no you can't see it either do what now?
|
# ? Mar 31, 2023 15:45 |
|
kitten emergency posted:I think there’s interesting business applications of llms in the observability (or really, any data processing) field around natural language querying of near-real time ops data (“show me the top ten slowest sql queries in us-east-1”, “what’s the most popular item being sold in Jakarta”, etc) but there’s pretty significant costs in training and upkeep that can’t really be shared (I can’t train my model off someone else’s telemetry) and are constant (if I’m not constantly updating the model with changes to the system then it becomes less useful) you appear to have invented a very computationally expensive sql front end. i expect accenture will be banging down your door any day now
|
# ? Mar 31, 2023 15:45 |
|
Grace Baiting posted:very rude to say this about a person who has the superpower to literally rewire his brain, ONCE, in order to perfectly execute a single apparently-impossible task. no he's never used this superpower, but it definitely exists. no you can't see it either there are many yud facts but this is definitely my favorite one, it’s such a bizarre thing to claim and really paints a picture
|
# ? Mar 31, 2023 15:56 |
|
Improbable Lobster posted:didn't google get sued over the button that let you directly see the iamge file They were sued for image search as a whole and it was ruled fair use, so a product that takes vast quantities of copyright images and produces X can be fair use depending on what X is. If you produce a search algorithm that includes thumbnails of the works, apparently that's fine (although mirroring the full image wasn't allowed, you're right) Is a set of neutral network weights that cannot (?) produce a copy of the input works more likely to violate copyright? I'm not sure how you argue that. If I tell it to produce an image of Micky Mouse then that image violates copyright, but that's the same as if I made it in Photoshop or with a paint brush. The work violates, but the tool I happened to use? It'll be interesting to see what happens. Chalks fucked around with this message at 16:05 on Mar 31, 2023 |
# ? Mar 31, 2023 16:02 |
|
izagoof posted:there are many yud facts but this is definitely my favorite one, it’s such a bizarre thing to claim and really paints a picture yeah lol
|
# ? Mar 31, 2023 16:03 |
|
Shame Boy posted:how scary it would be if we made the robot devil lol you reminded me about the Unlimited Steam twitch bot that does the whole Simpsons 'Steamed Hams' skit over and over with all AI-generated content; because of the script lines 'delightfully devilish Seymour' and 'ye gods!' AI Skinner will more often than not beseech the Gods or the Devil at length for aid in fixing his kitchen mishaps.
|
# ? Mar 31, 2023 16:05 |
|
Chalks posted:if I made it in Photoshop or with a paint brush. i would disagree vehemently, though i'd also add that copyright law needs to be loosened/eliminated rather than strengthened
|
# ? Mar 31, 2023 16:05 |
|
izagoof posted:there are many yud facts but this is definitely my favorite one, it’s such a bizarre thing to claim and really paints a picture the brain-rewiring superpower is quite good, but i think my favorite is that yud believes that if you're smart enough you can advance the state of scientific knowledge from first principles by just thinking really hard for a guy who regularly cosplays as a scientist, his perspective on scientific inquiry is at best 2 millennia out of date
|
# ? Mar 31, 2023 16:07 |
|
Improbable Lobster posted:i would disagree vehemently, though i'd also add that copyright law needs to be loosened/eliminated rather than strengthened I assume you're not disagreeing that an image of Micky Mouse that I make in Photoshop would violate copyright because it certainly would. That's all I'm saying there, a tool's ability to make copyright violating content does not, historically, make it a copyright violation.
|
# ? Mar 31, 2023 16:08 |
|
jemand posted:Speaking of use cases, I watched my manager request a summary of the bullshit and interminably long "sparks of AGI" (PR) paper from chatgpt. The openAI version that has no internet access and had to go off the provided title alone. And received a short bit of totally different bullshit text with no relation to the original. Regardless of that fact, ChatGPT's response was accepted by him and the rest of the team as a fair summary. sounds like the computer was just doing the "fake it until you make it" dance the rest of us do at work.
|
# ? Mar 31, 2023 16:11 |
|
Cybernetic Vermin posted:i don't think asics and details of the raw compute matters that much for inference, in that i'd be kind of surprised if e.g. bing chat didn't do upwards of 1 terabyte worth of touches in very expensive memory per word of output. iirc one of the hyperscalers (goog I think) started running inference on spare general cpus but cut over to dedicated inference hardware over five years ago but the number of inferences you need to do that dedicating hardware makes economic sense is p high not sure if the math has changed any with all the avx512 inference instructions they’ve jammed in in the last few years or if the models are significantly larger than accelerator
|
# ? Mar 31, 2023 16:21 |
|
NoneMoreNegative posted:lol you reminded me about the Unlimited Steam twitch bot that does the whole Simpsons 'Steamed Hams' skit over and over with all AI-generated content; because of the script lines 'delightfully devilish Seymour' and 'ye gods!' AI Skinner will more often than not beseech the Gods or the Devil at length for aid in fixing his kitchen mishaps. its so funny, he just randomly gets extremely religious
|
# ? Mar 31, 2023 16:28 |
|
what is the brain rewiring thing
|
# ? Mar 31, 2023 16:30 |
|
haveblue posted:what is the brain rewiring thing wasn't able to find it, but I did find this passage, where he notes that he and Buffy carry the same heavy burden of being superheroes quote:That said...
|
# ? Mar 31, 2023 17:52 |
|
Powerful Two-Hander posted:idk someone's found at least one use for chatgpt it seems lmfao gently caress i really need to start using this thing more
|
# ? Mar 31, 2023 17:54 |
|
the big distinction is if it's a "tool that can be used to create copyright violations" or effectively a compression/representation algorithm for copyright violations such that you get something like stable diffusion plus a prompt counting as a copyright violation in (sort of) the same way that an integer that can be interpreted as a jpeg counts as a copyright violation. once you go the latter route, it's not insane to just count the system that allows copyright-violating prompts as the problem instead of requiring the prompt to count-as a violation to be clear this says nothing about the concept of copyright or how laws should regulate it; saying that copyright should be different or that the "real" concept of copyright is looser or whatever is orthogonal to the above
|
# ? Mar 31, 2023 18:06 |
|
izagoof posted:wasn't able to find it, but I did find this passage, where he notes that he and Buffy carry the same heavy burden of being superheroes lomarf
|
# ? Mar 31, 2023 18:08 |
|
izagoof posted:wasn't able to find it, but I did find this passage, where he notes that he and Buffy carry the same heavy burden of being superheroes all "smart" people are idiots, example number 6.02e23
|
# ? Mar 31, 2023 18:10 |
|
one of our "thought leader" types at work posted this "I got chatgpt to rewrite the agile manifesto with a modern twist!" thing on linkedin and I sent it to a friend who thought I'd gone mad and was about to say "I thought we were friends why did you make me read this?" Until he saw my "wtf is this poo poo" message at the bottom I hadn't even read it myself lol
|
# ? Mar 31, 2023 18:33 |
|
bad faith must be met in kind. your countermeasure should be less effort than the original bad faith post (much like my posts, yeah, etc etc)
|
# ? Mar 31, 2023 18:36 |
|
bob dobbs is dead posted:bad faith must be met in kind. your countermeasure should be less effort than the original bad faith post the power of the quote-and-'lol'
|
# ? Mar 31, 2023 18:36 |
|
Powerful Two-Hander posted:one of our "thought leader" types at work posted this "I got chatgpt to rewrite the agile manifesto with a modern twist!" thing on linkedin llms and ganns over individuals and interactions time to market over working software wild guesses over customer collaboration hyperbolic extrapolation over responding to change
|
# ? Mar 31, 2023 18:50 |
|
replacing my dvd encryption key T-shirt with a serialized GAN dump T-shirt good thing I’m really really fat
|
# ? Mar 31, 2023 18:54 |
|
izagoof posted:
god what a great line
|
# ? Mar 31, 2023 19:08 |
|
|
# ? Apr 27, 2024 14:22 |
|
like if i had to condense the psychological damage putting kids in a gifted program does into two sentences, i don't know if i could do much better
|
# ? Mar 31, 2023 19:08 |