How long will the next AI winter last? Any predictions?

I think this one may even be longer than the last one because this cycle was entirely due to the existence of GPUs and NVIDIA. There hasn't been a single fundamental innovation. Neural networks are ancient and are unlikely to be the holy grail of intelligence after multiple decades of failed attempts. It's like running an old game on an overclocked modern desktop and claiming you invented a new game. Since we are at the absolute peak of hype now (maybe a little past the peak already), this should be the start of the bear market for AGI. The real question is if Lisp will make a comeback in the aftermath.

>>1
I strongly disagree. Long short-term memory works amazingly well to solve increasingly complex problems we used to program for. Take a look at AIs playing games, GPT-2, Transformer writing anything from genius anti-semitism to homosexual jokes, and the incipient but game-changing business created by AIs able to create music.

Let me take the AI-composed music as an example. Here: https://www.youtube.com/watch?v=TqH6ionF94I after learning Chopin, and here: https://www.youtube.com/watch?v=2kuY3BrmTfQ after learning Vivaldi. Sure, it sucks. It can't compare to Chopin's genius, who was one of the very best composers ever, highly regarded and IMO still grossly underrated. But the key is not that it's amazing. It's that it's good enough. Good enough to be used by a low or mid-budget Netflix show. Good enough for most videogames. Far too good to be used by today's shitty TV programmes, let alone background music in company videos, facilities, speeches, and so on. It's good enough and it works.

I'm talking about businesses like this: https://www.aiva.ai/

If this is not revolutionary, tell me what is. I don't think winter is coming.

http://ai.neocities.org/AgiMind.html -- Artificial General Intelligence.

>>2
Each AI hype cycle is not without a few actually good results. The first one gave us LISP. The second one in the 80s gave us heuristic search algorithms (leading to Deep Blue), NLP (leading to Google), backpropagation and automatic differentiation. This one gave us reinforcement learning, the transformer, and actually good computer vision. It gets better each time.

Still a peak though.

What you see here with GPT-2 and deepfake GANs is the absolute most you can squeeze out of the hardware now. You have big corporations throwing millions of dollars at training neural nets. Even if you can somehow make a custom GPU/ASIC to do 100x what a NVIDIA card can do, you're just increasing the throughput which is already well achieved and surpassed by big corporations on existing hardware. There's no new progress to be made on hardware for deep learning that will make those big promises deliverable.

Either someone comes up with something better than neural networks now or it all goes bust. All those companies who rebranded themselves around AI will jump as fast off the sinking ship as they boarded it. Eventually "self-driving cars by $CURRENT_YEAR+2" will lose all credibility. The market will overcorrect. That's how you get winters.

The current AI is horribly slow, requiring lots of GPU time. It would be a breakthrough to do any of the neural net stuff on a PC.

Recently they had some success scanning animal brain structures. For example, the fruit fly brain was completely scanned:
https://www.hhmi.org/news/complete-fly-brain-imaged-at-nanoscale-resolution

So it is just a matter of analyzing and re-implementing them in computer.

Neural networks are already better than human brain at matching basic items. The remaining is integrating all that shit into single general intelligence, capable of acting in real world.

>>6
And yes, insect brain is actually harder to scan than mammal brain, because insects have smaller neuron size.

>>7
Can you please stop being fucking retarded for a second and think about the ramifications >>6 posted?
If fucking Ashburn Virginia scientists working for Chevy can get financed to research fruit fly brain neurology, think of what China with no ethics already does to in-vitro babies for their child soldiers.
All it fucking takes is artificial wombs
https://quarkmag.com/the-first-artificial-womb-and-what-it-means-for-humanity-f88041634f41
Plenty of egg and sperm cells,
http://www.ihr.com/infertility/provider/donoregg.html
and repetition.
In 20 years, mapping a full chipped human brain into adulthood is completely within our current capabilities, we already hacked rats.
http://www.nbcnews.com/id/3077264/ns/technology_and_science-science/t/scientists-wire-rats-remote-control/

If you want to keep away an AI winter, turn up the gas.

chiggers

Double Ones -- you check 'em, we spec 'em.

https://old.reddit.com/r/theology/comments/dkcadl/what_bible_and_bible_study_software_or_apps_do/f55ee1b

I make here (https://eduardocapaverde.wordpress.com/2019/10/21/ninguem-quer-falar-de-estrategia/ ) an appeal to nature: Knowledge in nature is always the encoding of an use, and as the uses compete for use and one gets less used, it goes away. The same pattern that makes organs in animals devolve causes human forgetfulness. It is a 'compact' solution, not an extensive one.

(Knowledge is never a goal in itself (in nature), but merely persists through its utility (to survival, for instance))

But let's not imitate nature completely; what does this tell us about our interests - encoding it for machine use? I would not like a complete automaton, black box that learns by itself; rather, I would like the programmatization of science for the clarification of its human concepts, as once it has been mathematized for the clarification of physics.

Is there knowledge without a subject? If not, we need to construct the subject, or at least imagine it. Since knowledge is not without interest, what are the interests of the subject? Does it at least have a body to protect? A history to honor? A small-minded goal?

>>13
You sound like a Russian asset.

>>14
I am not a russian asset. It worries me that you might think I'm one. I ask of you: what is it in my words that makes me sound like a russian asset (to you)?

http://ai.neocities.org/Dushka.html

>>16
You, on the other hand, are even named "Russian Asset". You are clearly a Russian Asset. The other guy seems to have confused me with you.

GPT-2 is crap