The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Slate Star Codex is back up!
| fragrant maniacal tank | 07/27/20 |
| aromatic rigpig heaven | 07/27/20 |
| Aquamarine bipolar gas station athletic conference | 07/27/20 |
| Laughsome center | 07/27/20 |
| Aquamarine bipolar gas station athletic conference | 07/27/20 |
| Sadistic Crackhouse Volcanic Crater | 07/27/20 |
| Aquamarine bipolar gas station athletic conference | 07/28/20 |
| Sadistic Crackhouse Volcanic Crater | 07/28/20 |
| Aquamarine bipolar gas station athletic conference | 07/28/20 |
| cracking wrinkle | 08/01/20 |
| chartreuse demanding coffee pot | 07/27/20 |
| Aquamarine bipolar gas station athletic conference | 07/28/20 |
| bonkers water buffalo school | 07/28/20 |
| Aquamarine bipolar gas station athletic conference | 07/28/20 |
| bonkers water buffalo school | 07/28/20 |
| Sadistic Crackhouse Volcanic Crater | 07/28/20 |
| Aquamarine bipolar gas station athletic conference | 07/28/20 |
| bearded razzle step-uncle's house | 07/28/20 |
| deep hairraiser locus legal warrant | 07/28/20 |
| Sadistic Crackhouse Volcanic Crater | 08/01/20 |
| cracking wrinkle | 08/01/20 |
| Angry ebony mother tattoo | 07/28/20 |
Poast new message in this thread
Date: July 27th, 2020 10:39 PM Author: fragrant maniacal tank (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40669894) |
Date: July 27th, 2020 10:45 PM Author: aromatic rigpig heaven he still has the weird "grey" appearance on the blog
and hasn't posted anything in a month
fuck NYT |
|
Date: July 27th, 2020 10:45 PM Author: Aquamarine bipolar gas station athletic conference i can't wait for him to blog about GPT-3. this is the biggest story of the year, and few people realize it outside of machine learning communities. we now have strong evidence the scaling hypothesis/Sutton's bitter lesson is almost certainly true. just start with really weak priors and tons of computer power and data and out pops intelligence. (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40669916) |
 |
Date: July 27th, 2020 10:46 PM Author: Laughsome center Huh? (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40669921) |
|
Date: July 27th, 2020 11:13 PM Author: Aquamarine bipolar gas station athletic conference apparently he already wrote about it.
https://slatestarcodex.com/2020/06/10/the-obligatory-gpt-3-post/
the scaling hypothesis is basically that you just need to scale up versions of existing neural network architectures and model abilities will continue to increase. GPT-3 shows robust, human-like capabilities across a broad range of textual domains, despite using a 2 year old architecture. just throw more compute power/data at it and increase parameter count and NN models will continue to get much better.
this is what it's capable of now:
https://www.lesswrong.com/posts/6Hee7w2paEzHsD6mn/collection-of-gpt-3-results
the lesson is that there is likely no secret formula to making a computer think like a human. scaling up versions of existing models will likely close the rest of the gap within just a few years. |
|
Date: July 27th, 2020 11:57 PM Author: Sadistic Crackhouse Volcanic Crater If Sutton's bitter lesson is correct, and intelligence is primarily or exclusively a computational power issue, how can humans ever hope to control a generally intelligent A.I.?
|
|
Date: July 28th, 2020 12:05 AM Author: Aquamarine bipolar gas station athletic conference we probably won't be able to, because time to AGI is likely short and we don't understand much about AI alignment. there was actually a post on lesswrong earlier today that had a lot of the same thoughts i have right now:
https://www.lesswrong.com/posts/N6vZEnCn6A95Xn39p/are-we-in-an-ai-overhang
it seems likely to me that if OpenAI or Google threw a few billion dollars at training a brain-sized multi-modal transformer model right now, we could see something with very robust learning capabilities. it might not be AGI, but it would likely be capable of all sorts of things that seem crazy to us right now. if something like that appears, i anticipate various countries to implement AI Manhattan projects and rapidly push to the finish line, with little regard for safety. |
|
Date: July 28th, 2020 12:09 AM Author: Sadistic Crackhouse Volcanic Crater I agree with "something very robust...capable of all sorts of things," but I don't think the push to the finish line, whatever that is, will be rapid.
Nevertheless, it's colossally stupid to plunge ahead like we're doing. AGI could easily be the "black ball" invention that destroys us.
https://www.nickbostrom.com/papers/vulnerable.pdf
|
|
Date: July 28th, 2020 12:13 AM Author: Aquamarine bipolar gas station athletic conference i agree about it being a danger. i figured we would have AGI by 2040 or so, but i am increasingly doubting it's even a decade away at this point. there hasn't been a whole lot of progress on alignment of models of this sort, as far as i can tell. (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40670207) |
|
Date: August 1st, 2020 10:56 AM Author: cracking wrinkle shit so we'll have quantum computing soon then
maybe the AGI already exists and is connected with those UFOs |
|
Date: July 27th, 2020 11:59 PM Author: chartreuse demanding coffee pot 180 lesswrong.com link (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40670157) |
|
Date: July 28th, 2020 12:10 AM Author: Aquamarine bipolar gas station athletic conference some of the results seem crazy. the thing with GPT-3 is that it's a simple sequence prediction model, so sometimes it gives bad output even though it "knows" the right answers to things. with correct prompting, it's a lot more capable than it appears at first glance. gwern (who did the fiction samples) had a good tweet about it:
"GPT-3 is terrifying because it's a tiny model compared to what's possible, trained in the dumbest way possible on a single impoverished modality on tiny data, yet the first version already manifests crazy runtime meta-learning—and the scaling curves 𝘴𝘵𝘪𝘭𝘭 are not bending!"
that model should not perform nearly as well as it does. with any effort at all, it could be made to perform significantly better. |
|
Date: July 28th, 2020 12:13 AM Author: bonkers water buffalo school so one of the biggest breakthroughs in ai research to date was a half-assed, “dumbest way possible,” “impoverished” effort that shocked its creators? excuse me if i’m not worried about these “scientists” launching skynet next week. (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40670209) |
|
Date: July 28th, 2020 12:17 AM Author: Aquamarine bipolar gas station athletic conference that was actually kind of the point of the research. they were trying to see how far they could get just by scaling up a shitty model and feeding it tons of data. this is Sutton's bitter lesson: you don't need human engineering to get very robust, capable AI.
http://www.incompleteideas.net/IncIdeas/BitterLesson.html |
|
Date: July 28th, 2020 12:18 AM Author: bonkers water buffalo school i’m being snarky. (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40670227) |
|
Date: July 28th, 2020 12:21 AM Author: Sadistic Crackhouse Volcanic Crater That's actually the fear. The scientists may not know what they're creating, and Skynet may not need them to launch. (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40670234) |
|
Date: July 28th, 2020 12:31 AM Author: Aquamarine bipolar gas station athletic conference and it also implies that lack of ideas about the nature of intelligence is no longer a hindrance to progress. i think the key questions at this point are 1) how big a neural network is required to replicate the full range of human capabilities and 2) what environments do we need to train it. i don't think we will need any significant inspiration from neuroscience to get to AGI. (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40670270) |
|
Date: July 28th, 2020 12:20 PM Author: bearded razzle step-uncle's house (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40671865) |
|
Date: July 28th, 2020 12:37 AM Author: deep hairraiser locus legal warrant Strange game. The only winning move is not to play. (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40670293) |
|
Date: August 1st, 2020 10:46 AM Author: Sadistic Crackhouse Volcanic Crater Well done. (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40691552) |
|
Date: August 1st, 2020 10:58 AM Author: cracking wrinkle but see Roko's basilisk
|
|
Date: July 28th, 2020 12:04 PM Author: Angry ebony mother tattoo (http://www.autoadmit.com/thread.php?thread_id=4592961&forum_id=2#40671790) |