The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Leopold Aschenbrenner essay on the AI arms race that will change world by 2027
Poast new message in this thread
Date: June 12th, 2024 3:47 PM Author: Glittery milk Really good read and a podcast he did about it. Dude is very smart.
tl;dr version: superintelligence is within reach and the capex to get there is being spent, the teams are being assembled, the power plants built, etc.; eventually US government will step in to oversee the project as a kind of Manhattan Project because a six-month head start is the equivalent of 20-30 years in arms development; things get really weird from a game theory perspective once superintelligence is achieved but before the robots take over. It's US v. China with the Middle East and Russia posing major risks because of their abundant natural energy resources.
Essay: https://situational-awareness.ai/ Four-hour podcast: https://www.dwarkeshpatel.com/p/leopold-aschenbrenner |
Date: June 12th, 2024 4:28 PM Author: Glittery milk (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47740975) |
|
Date: June 12th, 2024 4:50 PM Author: Chrome box office psychic pretty chill that the world will end in five years and I spend my precious remaining days poasting on a decrepit law school admissions forum (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47741056) |
 |
Date: June 12th, 2024 6:14 PM Author: Glittery milk It’s part of the data set (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47741277) |
|
Date: June 12th, 2024 7:22 PM Author: lilac stain parlor (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47741556) |
|
Date: June 12th, 2024 7:25 PM Author: Bonkers laser beams *superintelligent AI addressing the world* "I'm gay and retarded" |
|
Date: June 12th, 2024 7:26 PM Author: lilac stain parlor 180000000000000000000000000000 (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47741570) |
|
Date: June 14th, 2024 9:03 PM Author: Glittery milk (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47748063) |
|
Date: June 12th, 2024 7:22 PM Author: 180 indecent sex offender main people He just had another podcast with Chollet that was interesting and raised some concerns about this LLM maximalist point of view. LLMs still perform quite poorly on ARC. There are real concerns that transformer scaling will not address all the problems with out of distribution generalization and algorithmic reasoning. This buys a bit of time and probably pushes out timelines past 2027. Things will become progressively weirder though over the next decade and it’s hard to see how AGI is out more than 15 years. (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47741554) |
|
Date: June 12th, 2024 7:29 PM Author: Lavender balding office degenerate nothing being done in "ai" right now is a road to actual intelligence. what we have right now is essentially a sophisticated version of a database that can accept unstructured queries and that can interpolate between dataset entries. it would take a complete paradigm shift to get to something qualitatively more powerful than that. may or may not happen. (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47741579) |
|
Date: June 13th, 2024 11:34 AM Author: Glittery milk This is not right. The trend lines are clear. (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47743102) |
|
Date: June 13th, 2024 12:51 PM Author: Lavender balding office degenerate no (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47743420) |
|
Date: June 13th, 2024 1:23 PM Author: 180 indecent sex offender main people Most people seem like pattern matchers who aren’t able to robustly deal with novel situations. I would trust GPT-4 more to solve a random math word problem or answer RC questions from an LSAT than most people. The interpolation/extrapolation distinction becomes less and less meaningful in high dimensional spaces when you are training on everything. I agree LLMs don’t go all the way but they very likely go a lot further than they do now and there are other things in the AI space used for game playing that look like they can be combined with LLMs to bootstrap capabilities. (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47743513) |
|
Date: June 13th, 2024 2:01 PM Author: Lavender balding office degenerate llms far surpass humans at defined, stereotyped tasks that we have a lot of data for. they can't generalize to unseen environments (new probability distributions) and they lack an inherent teleology to prompt them interact with those environments. (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47743680) |
|
Date: June 13th, 2024 2:47 PM Author: 180 indecent sex offender main people I agree with this but most people suck at dealing with novel situations too and they can’t even do the routine cognitive tasks that GPT-4 excels at. I don’t think a median human could be made competitive with it at reading comprehension tasks. With 100x the training compute, there is no chance. There is really no principled reason why LLMs paired with model based RL couldn’t be made data efficient and adaptable. (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47743890) |
|
Date: June 13th, 2024 4:14 PM Author: Lavender balding office degenerate this is silly (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47744229) |
|
Date: June 13th, 2024 11:39 AM Author: Poppy stirring national I, for one, cant wait for AI to do all my TPS reports for me. Whatever will I do once those are automated? Never mind that basically everything we do in Compliance is bespoke transaction monitoring and the vanilla bulk transaction stuff was automated 24-25 years ago and hasnt had humans involved in a generation anyway.
Like the biggest thing is Compliance now is marijuana and the hemp exemption BS that plagues us. The fuck is AI going to do about any of that? |
|
Date: June 13th, 2024 11:43 AM Author: Glittery milk The interesting thing about his essay is the geopolitics part. The time when AGI is pursued by private companies is coming to a close. (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47743136) |
|
Date: June 16th, 2024 7:28 PM Author: Glittery milk (http://www.autoadmit.com/thread.php?thread_id=5539938&forum_id=2#47752081) |