For interested AI developers

https://www.general-ai-challenge.org/

Just thought I'd drop this here. The first challenge starts in February, and while they (naturally) don't have a clear-cut series of steps to achieve general AI, that's the overall goal of the company.

They also talked on stream today:
Part 1 (1 h 7 min in)
Part 2
Part 3
(stream was not very stable)

[Spoiler]

Just for a reality check..

While deep blue was credited as having beaten a world-class chess player, it required IBM to tune the program between every match. There was also considerable evidence that there was cheating involved (ie employing a real human pro to play on the computer's behalf, or something equivalent) and that the entire thing was nothing more than a farce to boost IBM's public image. The abilities of current state-of-the-art chess AIs also strongly support the cheating hypothesis, as they are not known for beating world class players.

RTS games are considerably more complicated than chess, and involve a significantly larger possibility space. I've never heard of an RTS AI that can compete with world class players even when using maphack, and the strongest ZK AI is probably only ~1500 elo equivalent (and does not cheat). Starcraft competition AIs have been known to use a variety of advanced AI techinques including artificial neural networks.

By comparison the human brain is composed of a number of neural networks each semi-specialized towards particular tasks. Some parts are more specialized than others, such that if you lose some parts of your brain the remainder may be able to compensate by re-adapting, while losing other parts may cause permanent disability. Brain functions rely both on electrical signals as well as chemical signals, some of which are local and others of which regulate brain functions in a more global way. The number of signaling chemicals used by the brain is rather large, and their effects are not particularly well understood. The method by which the various localized neural networks of the brain are syncronized and produce a coherent thinking whole is also not remotely well understood, and even the low-level interactions of neurons are still largely a mystery.

A single human brain is estimated to produce around 40 petaflops worth of computational work, which is similar to the most powerful computing clusters available today. Running a physical simulation of the human brain is even more demanding, requiring a 10 petaflop computer and 40 minutes just to simulate one second's worth of brain activity. Said simulation also required 1 petabyte of RAM even though the number of simulated neurons was only around 1.7% of the estimated number contained in a typical human brain (1.7 billion vs 80-100 billion).

Aquanim is correct. One of the top chess players, Magnus Carlsen, has a chess rating of about 2900. In comparison, one of the top chess engines, Stockfish, has a rating of about 3400. Between the two, it wouldn't really be close.
[Spoiler]

It's difficult to find sources for recent games between top-level players and AIs... because the AIs are so good that the humans no longer bother.

Apparently the computational power of a smartphone was sufficient for grandmaster-level play in 2009 though. (source)

CAI is general AI:

Cheats
Adapts, pretty shit but whatever
Can do whatever
Goes planes and does craptastic stuff


[Spoiler]

I clicked that link and i was disappoint that there is no actual challenge. They haven't even formulated their first milestone challenge. So the entire thing boils down to "we are going to do an AGI challenge, but first we have to figure out what the challenge will actually be".

Perhaps they should make a challenge to figure out the milestone challenge.

Well, that, and yet another ted talk.

quote: we are going to do an AGI challenge, but first we have to figure out what the challenge will actually be

"...and we want you to discuss with us what it should be and how you think the contest should be held".

I agree that it's a bit in the direction of "do our job for us", but on the other hand they have a 5 million prize pool lined up.

I guess waiting until the first challenge is out would've been preferable, but I also think some people around here might be interested in it beyond a programming exercise.

I don't see how it is an AI competition when there is no well defined way to compete. At some point it's just hiring people (albeit, in their spare time) and deciding not to pay most of them.

really? when was the last time somebody made a prediction of that kind, archivist?

quote: yesterday in the "ZK Developer Support of AIs" thread.

What was actually said:

quote: A pipe dream scenario would be an AI tourney being declared a month after Steam release, which in this optimistic scenario happens in early Spring 2017. The tourney actually happens in summer or autumn, giving entrants time to build their bots; generating new ai devs and projects and galvanizing the community. A pessimistic scenario is that such radical growth does not happen, and then the effort should be to maintain the current beneficial climate, keeping the currently active projects alive. Perhaps delaying engine upgrades for moderate grace periods pending on AI compat is one way to achieve that. Perhaps even that is not needed as long as no radical destructive actions are taken.

Like all self-righteous zealots, knorke considers it acceptable to tell lies that take 5 minutes to disprove.