The 7 stages of AI

[sjastro]

Here it what the former world no. 2 junior chess player and programmer of AlphaZero chess has to say on the subject of AI.

​

 

[AV1611VET]

 

[durangodawood]

....
The rest is up to AI where it trains itself through self-playing of chess games.
Initially it plays like a novice and builds up skill levels based on the number of self-played games.
This is the “knowing” or skill level part.
....

You seem to be unconcerned that "knowing" traditionally brings with the flavor of "awareness". Dictionary senses of "knowing" support this, as does most peoples intuitive sense.

Thats why, to me, all the chess engine action, even the learning part, seems like a lot of very elaborate doing. Its all strictly physical process playing out and none of it takes place in any ones conscious awareness.

Do you think I'm chasing ghosts, as it were, in looking for a conscious awareness to be the seat of real "knowing"?

 

[sjastro]

You seem to be unconcerned that "knowing" traditionally brings with the flavor of "awareness". Dictionary senses of "knowing" support this, as does most peoples intuitive sense.

Thats why, to me, all the chess engine action, even the learning part, seems like a lot of very elaborate doing. Its all strictly physical process playing out and none of it takes place in any ones conscious awareness.

Do you think I'm chasing ghosts, as it were, in looking for a conscious awareness to be the seat of real "knowing"?

As discussed earlier in this thread “awareness” is a topic which tends to confuse the issue in AI as we humans have problems in defining what exactly “awareness” is, which hasn’t been helped with recent claims in the field that AI has already become sentient.
I don’t believe AI has become sentient but there is the property of creativity which involves “knowing” and transcends the application of knowledge or “doing”.

Up to around 2010 computer chess engines were based on the gradual improvement of conventional computer chess programming dating back to the 1960s running on increasingly powerful hardware.
The early versions of these programs were brute force where the program would attempt to find all possible moves down to a particular search depth, to be gradually replaced with a selective depth search based on human knowledge programmed into the engine.

At best these engines were better versions of human players.
With the advent of reinforced learning, AI chess players are not only better versions of human players but play with a distinctive non-human style.
Humans did not create this non-human style as by definition reinforced learning does not involve humans.
The creativity comes from AI itself where it is “knowing” without necessarily being sentient.

 

[durangodawood]

As discussed earlier in this thread “awareness” is a topic which tends to confuse the issue in AI as we humans have problems in defining what exactly “awareness” is, which hasn’t been helped with recent claims in the field that AI has already become sentient.
I don’t believe AI has become sentient but there is the property of creativity which involves “knowing” and transcends the application of knowledge or “doing”.

Up to around 2010 computer chess engines were based on the gradual improvement of conventional computer chess programming dating back to the 1960s running on increasingly powerful hardware.
The early versions of these programs were brute force where the program would attempt to find all possible moves down to a particular search depth, to be gradually replaced with a selective depth search based on human knowledge programmed into the engine.

At best these engines were better versions of human players.
With the advent of reinforced learning, AI chess players are not only better versions of human players but play with a distinctive non-human style.
Humans did not create this non-human style as by definition reinforced learning does not involve humans.
The creativity comes from AI itself where it is “knowing” without necessarily being sentient.

I find it a challenge to get past our normal understanding of knowing which is to "be aware of..." I do agree tho that awareness in this classic sense is not necessarily relevant to the concerns we currently discuss about "them".

But... if machines did acquire classical-type awareness, that would imply sentience, I think, and open up whole other moral and historical vistas. We'd share the world with other vastly powerful "persons". Its hard to grasp what that would even mean.

 

[Halbhh]

Do you remember M chess - it was a tough cookie to beat on highest grades.
Fritz at the time wasnt bad but far easier to beat.

It was the first time cheap commercial programmes got over 200 ECf grade -say 2300 elo.

I remember the days when a guy called Brian levy began analysing the scope of the problem Pre PC
He Wrote some articles in Baruch woods “ chess magazine” At a time there were not even cheap calculators!
How much memory , computer power it would take , in the days of cards and paper tape.

Its only in the late 80s pcs got beyond 1 m ram and 10-100m disk.
Its hard to remember now but in 1990 a single image wouldn’t fit in PC ram!
Array processors were only Tens of MOps, Although could be networked (, which was the bane of physical modelling - eg weather and signal processing too)
Since then both computer and memory power growth has been exponential growth
, and that rather than cleverness overwhelmed it.

Mchess itself fitted on multiple 1 m floppy disks.( who else remembers them?)
Those programmes were in essence a deep opening database , candidate moves, analysis 5-10 moves deep then mechanised positional analysis which is how players of the time did it too. (The cleverest of them like Fischer were better able to assess position , and could certainly analyse deeper Who remembers when he thrashed other grandmasters 6-0 in world champs games! my idol at the time!)

But Brute force has allowed the problem to be overwhelmed rather than solved,
With the ability to analyse 20-30 moves deep and big hash matrixes of 6 endgame pieces etc,

The point I am making…
the reliance of brute force has REDUCED the need for accurate positional assessment, because the search depth allows the positional advantage ( at say 5 moves ) to translate into material advantage at 10-15 moves or big positional advantage which is far easier to measure, for a man as well as machine. But a machine can think quicker and do more.

Humans have a harder job Because they can never analyse so many or so deep variations , they still have to be better at positional assessment.

in short human reasoning is more versatile than machine , but overwhelmed by mechanised power.

I’ve not followed it last couple of years but

The 2800 elo grade seems to represent pretty much human limits,

Great post. I didn't play M chess. I wasn't that good anyway compared to that 2000+ level. I recall that Fritz was very strong in the mid to late 90s and that's when I was trying to beat Fritz, and it took a lot of games to find a line that was even possible for me, a very long term plan and a lot of perfect play on a game that had a simplified situation due to the structure of the pawns, etc. to figure out a game where I could beat the Fritz version (which really was from a few years before) on my PC in the late 90s (which was at that time only an ok, mid budget PC of like mid to late 90s level). Fritz was very impressive to me back then. Yes, like 98 times out of a 100 in just varied games, Fritz would win due to me missing some obscure tactical play it found that often was only 3-6 moves deep.

 

[sjastro]

Great post. I didn't play M chess. I wasn't that good anyway compared to that 2000+ level. I recall that Fritz was very strong in the mid to late 90s and that's when I was trying to beat Fritz, and it took a lot of games to find a line that was even possible for me, a very long term plan and a lot of perfect play on a game that had a simplified situation due to the structure of the pawns, etc. to figure out a game where I could beat the Fritz version (which really was from a few years before) on my PC in the late 90s (which was at that time only an ok, mid budget PC of like mid to late 90s level). Fritz was very impressive to me back then. Yes, like 98 times out of a 100 in just varied games, Fritz would win due to me missing some obscure tactical play it found that often was only 3-6 moves deep.

By the late 1980s chess engines running on the very slow PCs of the time could beat 99% of all chess players.
Here is an interesting video on the improvement of chess engine performance since 1985 and how they are now way beyond the super grandmaster level of 2800 elo.

 

[Halbhh]

By the late 1980s chess engines running on the very slow PCs of the time could beat 99% of all chess players.
Here is an interesting video on the improvement of chess engine performance since 1985 and how they are now way beyond the super grandmaster level of 2800 elo.

Fun video! It was also fun to see Fritz going from like ~2300 ish to 3100ish(!) I bet mostly from just computer power as much or more than any refinements in its search/evaluation. I remember also later in the 2010s when Stockfish ruled the tank a while, but I've not paid as much attention in the last many years.

 

[sjastro]

Fun video! It was also fun to see Fritz going from like ~2300 ish to 3100ish(!) I bet mostly from just computer power as much or more than any refinements in its search/evaluation. I remember also later in the 2010s when Stockfish ruled the tank a while, but I've not paid as much attention in the last many years.

I recall reading somewhere that in the 1980s computer scientists using mainframe brute force chess computers found increasing the depth search by an extra ply increased ELO performance by around 250 points, so the progressive increase in ELO over the years has a lot to do with hardware.

With the advent of reinforcement learning AI chess engines however the depth search is not as deep as conventional programs since it takes longer to evaluate the position which is considerably more complex since AI applies its vast knowledge of chess obtained during the learning process.