Telegrams: AI Developer
Added 2020-05-16 20:06:03 +0000 UTCTomorrow I should be chatting with an AI Developer still at University for the next Broken Silicon Episode.
We will discuss loosely what he does (Reinforcement Learning, Upside Down RL, and more) so that hopefully people get a real idea of what these researchers are actually doing with Volta, Vega, Ampere, and soon RDNA 2. Eventually we will touch on gaming AI & ML as well.
Put your questions/comments/concerns BELOW! Be thoughtful, be concise, use good grammar, and we may just get to your posts.
Comments
Great episode. Will note, people like me who think AI is a likely risk do not think they will spontaneously become evil, nor do we think they will be programmed to be evil. We think the space of possible goals is very large, that specifying human values is an unsolved problem, and lacking a solution to the value specification problem we will be left with an ecosystem of agents with a wide variety of utility functions. If you know anything about natural selection, you know this is unlikely to be a safe equilibrium. Stuart Russel, who literally wrote the book on AI, has a great book called Human Compatible which makes the case for taking AI risk seriously. If you do not feel like buying the book, Scott Alexander has a great review of it here: https://slatestarcodex.com/2020/01/30/book-review-human-compatible/
Edward Huff
2020-05-20 14:53:42 +0000 UTCI often hear the perception that AI's in video games have been stagnant and have not improved since the first F.E.A.R with the exception of Alien Isolation. Why do you think that is? Is it because game developers are emphasizing too much on graphics or is there something more? And what could developers do to change this perception and develop better AI for all games?
yoda king
2020-05-18 08:17:18 +0000 UTCListening to mainstream media, one might get the idea AI is the solution to everything, from self-driving cars to AI-accelerated assistants like Siri. To which extent do you think AI will really revolutionize our daily lives during the next decade? Are there unexpected areas in which you see it as a gamechanger?
2020-05-17 20:52:38 +0000 UTCSince the "Johannes Gutenberg" (Introducer of the printing press in europe ;-) ) university in Mainz is using the MOGON II supercomputer to find a cure for Covid-19 and it took them 2 months (more than 30 billion individual calculations) to narrow it down to 300 candidates (and afterwards some more time to narrow it down even further to 10 candidates) - what impact does Nividias GA-100 announcement have for guys like you working in this field? FYI: MOGON II uses Xeon e5-2630v4 and Xeon gold-6130 cpu's as far as I know.
Benni_Berlin
2020-05-17 17:43:11 +0000 UTCIf you were to compare the most powerful supercomputer to an animal’s brain, what kind of animal would it be comparable to? For example, are we closer in processing power to an insect’s brain or, say, a squirrel’s brain?
Jake_ Dude_23
2020-05-17 14:13:45 +0000 UTCWhere does the fundamental problem lay, in allowing AI to be trained/train itself, in a manner simialir to the human mind? What is that quantum leap and how are researchers and developers being taught to approach this conundrum?
Dominik Koc
2020-05-17 00:05:57 +0000 UTCWhat will it take for Machine Learning to shift away from a reliance on CUDA? Is there any worthwhile pipeline for non-CUDA ML dev?
Trogdor
2020-05-16 22:59:37 +0000 UTCHow big a disparity is there between the common person/the media's perception of AI for commercial use and the reality of current development? For example, there've been articles about how AI for information processing/generation tasks like news reporting could be created within 5-10 years, but can that kind of AI be realized in that time frame?
Cleansweep
2020-05-16 22:32:23 +0000 UTCHow many people approach you with Terminator/iRobot doomsday scenarios when discussing AI? How often are you approached with such scenarios?
The Immortal Cameraman
2020-05-16 22:08:51 +0000 UTCWhat’s the most complex AI someone with a 3950x and 2080ti (or Radeon VII) could run?
hhectorlector
2020-05-16 21:13:12 +0000 UTCWhat is your opinion on the synapse vs parameter question? Artificial spiking neural networks perform really poorly compared standard stuff, which may imply 1 paramater > 1 synapse. Though software is an enormous aspect of the problem, in terms of minimal viable compute for human-level agents once we have enough memory/bandwidth to handle models with as many paramaters as human's have synapses, around 100 trillion, will this be insufficient, good enough, more than enough hardware?
Edward Huff
2020-05-16 20:54:43 +0000 UTCHow do you feel about 'AI' being an overused buzzword in some situations?
2020-05-16 20:46:14 +0000 UTCWhere is the bottleneck in AI training? Is it the actual compute power or is it something else entirely that we (the general public) do not really understand?
extramossy
2020-05-16 20:36:32 +0000 UTCWhere does one even start wrangling Robobo's artificial brain into thinking the thoughts you want it to think these days? Lots of stuff about The Deeds, not much on Where Do I Sign Up ;)
Bollocks
2020-05-16 20:12:35 +0000 UTC