Highly Evolved Leela plays “Bluefish” – a Supercharged Three Times Faster Stockfish in Catalan

Highly Evolved Leela plays “Bluefish” – a Supercharged Three Times Faster Stockfish in Catalan


Hi all. I have a rather fascinating game
to show you Leela – a version of Leela was
playing against BlueFish Dev So let me introduce BlueFish dev. This
has been used throughout TCEC as the kind of point of reference for the games in
progress. It’s a kind of supercharged version of Stockfish they say one
account says roughly three times the nodes per second as ordinary Stockfish
so it wouldn’t be allowed for the TCEC Hardware regulations basically. And also
not only that the real BlueFish has a
seven-man endgame tablebase rather than a six-man table base. If you know about
tablebases they take a huge amount of data. And one extra man is gigantic
increase in the amount of storage needed but so this this version actually of
bluefish was six man table base now the version of Leela is T 40
T six five three two and on the information about it
it says whimsically it’s from the parallel universe where Leela T 40
mutated and evolved a win draw loss head and we halved the learning rate (LR) and provided it a partial 7 man table base rescoring. Network
equivalent is 42128 by training stats forked from 42,000 and
still pure Leela and only trained using Leela self play games. So yes we have very
interesting versions of these arch-rivals in this game. Do they carry
the same genetic strengths and weaknesses? Let’s have a look d4 from
Leela we have the Nf6 c4 e6 g3 so Catalan territory
we have d5 Bishop g2 Bishop e7 Nf3 both sides castle and now D takes c4 so
this does seem a great stockfish choice which it has
suffered a number of losses in a particular variation of the Catalan
opening where it does play a kind of what humans might regard a slightly
double-edged move b5. It exposes that diagonal potentially. It does cling on to
the pawn for a bit. Ut’s kind of underminable with a4 at convenience and
let’s see so what happens? a4 – it is actually undermined immediately b4 so
the point is to try and perhaps get some counter play if white took immediately
maybe with Bishop a6 at some point but we have actually Nfd2 so it’s
interesting it’s like this version of stockfish generally is expecting a kind
of materialistic dialogue but a materialistic conversation and instead
Leela does the opposite of materialism just trying to celebrate the
other aspects of the pain that black is causing itself in this position. So just
hitting that rook and trying to really celebrate maybe the c4 square and this
pain could be the root of further issues which could be concrete later. So
that is very interesting here. For example the c4 square. If a bishop landed on c4 later
then that celebration of that pain that black is self-inflicting here by this
variation for example could be undermining the pawn chain later with f4 f5 you could imagine a nice Bishop perched on c4 as an example so the dialogue of both
is very very different so this is kind of unexpected perhaps Nfd2. And very
unexpected now after c6 blunting the bishop stopping the bishop taking the
rook is to just give up the d4 pawn but this is a very very interesting
subtle way perhaps of exploiting the pain and the suffering that black is
causing itself in this opening by taking this d4 pawn
which stockfish – the ordinary starfish has lost in a number of games now
including the against the Houdini chess engine which it got knocked out in a key
game in the catalan in the last TCEC cup. Houdini wasn’t been that much updated and it managed to beat some the ordinary stock
fish from this position so does bluefish dev which can calculate three times as
more powerfully – three times as as quick have something up its sleeve here not
known about. We have rook d1 the Queen goes to c5 so this looks very familiar
with the Queen bouncing via c5 to h5 to hit h2 to make Knight g4 a concrete
threat. It temporarily pins of course that knight. The bishop kicks the Queen
goes to h5 and now we have Nbd2. So the ordinary Stockfish would have
played Knight g4 here and the first surprise of the game Knight d5 so not
Knight g4. So we’ve seen games where Stockfish was interested in taking out the
dark square Bishop. Is that the case here? Nb3 and in fact not very long to
still do the same kind of thing Knight takes e3 is played here so why is this
because it did seem to blunt the diagonal – it seemed kind of nicely
perched on d5 wasn’t this going to be like the kind of refutation of whites
play in a way with the deeper version of stock fish if we look at Knight d7 Nd4
maybe a good try for example like this Bf3 kicking in the Queen and
Knight f5 and that knight shows itself as a kind of tactical liability – an excuse
to play moves like this where yeah d5 and a8 is loose and for example this
position where White could get a significant advantage. Structural damage concretely
now for black. White is doing very well there so we have Knight takes e3 Nxe3.
And now a6 you might wonder why – move a6 – as an example a disastrous
example maybe which I haven’t shared with you about this position why a move like a6 would be needed if we look at rook d
eight this is a funny one white could take and then take here and if Knight
takes this is actually a double attack on a8 and e8 and the back row is actually weak. In this example this is checkmate so a move like a6 does
seem a bit weird. Despite all this calculation power black is reduced
to this position with just the Queen really out on h5. Most of the pieces
visually on the back row so it’s interesting this dialogue on the chess
board with this calculating monster and Leela – a great version of Leela still
results in in this position where you know Leela is really after the longer
term celebration of blacks materialism yeah the materialism has created some
symptoms I mean in particularly you know White’s pieces qualitatively are
controlling nice key squares c4 in particular really jumps out as a major
square to use as I mentioned before for example if a bishop was there later that
could be used for undermining the pawn chain. A lot of the pawns on light
squares here but furthermore the symptoms of black don’t end with the c4
square. The pieces over here are kind of having a certain threshold of
trapability I would say with these pieces not just being hemmed in or
contained but trapability if black’s not careful as some of the variations we’ll
see kind of witness. The other aspect is this d-file locking down and containing black’s pieces by doubling on the D file. The rook qualitatively will often be better than the black rooks. Now this is not a
hundred percent guaranteed win. There was a game at a faster time limit where stockfish
did actually draw with this position with black. Some thing
like this recently but it has lost a few as well. It’s not a guaranteed win though.
Let’s see what happened Knight c4 we have Ra7 so black is trying to crawl
out. Rac1 c5 so this does kind of contain – self contain the bishop on e7
Nba5 Queen g5 was played so very interesting move very very interesting
tactical move. It looks like it’s been tactically generated anyway. If Bishop d7
b3 is nice to play just keeping up this grip on the light squares saying to
black “you know what are you doing” and White could build up. This is an example say
Black was desperate enough to play this to challenge the bishop so giving white
the light square bishop this should be a great way to make sure the Knight is not
going anywhere. And as an example just build up like this is a great
example of light square domination basically. I would call light square
domination with an infiltration point on d6 here if the bishop ever neglects d6
this kind of position is just going to be very nice for white. so okay Queen g5
we have h4 Queen f6 and now a really interesting move Qe4. In some
weird and wonderful variations this Queen e4 does actually support the
move Queen a8 when the rooks not protecting it. So to sort of harass these
pieces this is the self-inflicted pain that black did in in
the opening. Because of the kind of materialistic
agenda that seems to you know be in the long engineered code one of the you know
the fundamental assumptions about evaluating positions “materialism”. But it
can be a source of pain and it’s as if Leela is ready
to celebrate it the kind of trappability of these pieces. If we look at rook c7 and
already even here this is a legal move Queen a8. It doesn’t actually do anything
much but under certain circumstances Qa8
could be painful for black. If King h8 just as an example I wanted
to see – I wanted to ask the question what would white do if black did nothing here.
And here’s an example. I’m just putting the King back and forwards. What would white do?
White could do something like this – use that c4 square
concretely as a pivot to attack the black Queen. The black queen is prone to
attack in this position and is ahead of the pawns. It’s not within the pawn structure
and you can see a big tempo gainer Rf4 here would have to be
factored in. And so this is a kind of disaster scenario if black did nothing.
That’s a total disaster. Just just for the record – how is Black suffering concretely if you wanted to ask that. You
could ask it just by some King moves token King moves but anyway rook c7
Rd3 free so this is dual purpose it seems offering you know it’s the
possibility of Rf3 which has to be factored in now. At all times. But the
real point is doubling the rooks so qualitatively the rooks are really much
stronger than the black rooks. We see g5 here. It’s interesting here
that this does neglect a8. Now this is just for amusement – Queen a8
just to prove it it is it is a possible move it could be played. For example like
this would end up with a small edge for White if blacks not careful. Black
can do better this was just really for amusement to see if Queen a8 was a playable move under certain circumstances. But black could equalize
with this kind of thing. A bit more accurately. Slight edge for white still yeah but
anyway Rd3 so the Queen does have that possibility of going to
the outer reaches of the position here touching these guys as I say.
They are kind of vulnerable they’re not just contained. It’s interesting
sometimes we don’t really think about the trappability aspects of contained pieces. We just think they are contained but they’re also kind of
tactically vulnerable in other respects as well. So we have g5 and this does seem prone to pawn structure fragmentation by moving pawns which
can’t go ever backwards. This is a very committal decision structurally so in
human chess of course we think very carefully about how we place our pawns.
Anatoly Karpov once said you know the opening where your pawn structure is decided you
have to like make your bed and sleep in it. After you have to be very careful
what you do with your pawns and the openings often determine the pawn structures. Hence Karpov’s quotation. If King h8 Rcd1 just as an
example again What if black didn’t do anything then this kind
of thing and you can see c7 a big target as well as b8 for example like this and
then b8 dropping. Yeah I mean look the b8 knight is if you look at this position
is amusingly is a tactical liability. It’s not just that they are contained. They are under observation potentially. They are under scrutiny. Their very existence is threatened
in some variations. And here if Bishop b7 this is an interesting one where rook
takes f7 is possible unveiling the attack on b7 so getting a big advantage. And let’s
see there was something else rook d7 here is an amusing example where
actually the move Queen a8 which Qe4 affords again is harassing
these pieces. So these contained pieces really do have other aspects of
tactical liability to be kind of celebrated by white. So for example like
this where Queen b7 and there’s a ridiculous idea from a clinical
perspective here white could just take on d7 but there’s actually just to
emphasize this other aspect of liability white could actually as well as rook
takes d7 play rook d6 rather amusingly and kind of checkmate the knight on b8
and this would actually be the top most clinical way of exploiting White’s
position that like being checkmated here so containment has led to trappability
and actually winning material. Bang! The knight would go off the board and white
could actually come back to sort out any counterplay issues with a big advantage
so yeah it shows that the materialism of black in the opening. Some of these
variations shows that it’s not just that the pieces are contained they’re actually
tactically vulnerable in their own right so anyway g5 it’s kind of an active
looking move and in fact instead of hxg5 which one might expect to try and
celebrate the h-file for example but it would immediately hit the rook.
There are immediate tactical issues on hxg5 but it was all rejected with h5
actually. So that’s interesting. Instead of okay so h5 just have a quick look at
hxg5. So Qxg5 hits the rook on c1. Say the rook moves it does control the d-file.
White would still be okay but maybe this is black’s way of trying to generate some
counter play. White still technically it seems has a small edge here so this
is very interesting. It’s more containment style you could say. h5 and
it’s on light squares it’s probing light squares and it’s got the
potential for the dreaded “thorn pawn” as well. h6 as well – the dreaded “thorn pawn”
seems to be a very recurrent theme which has aspects of containment of
course as well. And in fact here immediately we see in this
position h6 so if Queen takes h6 there’s actually
Qe5 targeting the rook and if Bishop d8 then bang rook takes d8 and
Queen takes absolutely winning. All three of these guys are now forked with the
Queen just doing her own thing over here does not help. An amusing
position indeed. So we have actually if rook a seven then Queen takes b8
again. The poor knight can be victimized in some of the variations of this game
at least. So Queen f6 but is this a tactical
liability later this h6 pawn. That’s the thing is it a bit double-edged? b3
is played Bishop d7. Here if rook d7 then rook takes Knight takes Knight c6 and
look at black’s position visually. It’s really suffering.
If King h8 Knight takes Queen c6 and Knight d6 and it’s the c8 Bishop which is now the victim of this. So as I say yeah it’s maybe containment – we can think of it may be about containment for its own sake or containment with a view to sometimes
winning the contained pieces at leisure as this shows as well. The knight
winning the bishop here is quite nasty so black has to be extremely careful in
this position so Bishop d7. We have Queen e3. g4 is played. On King h8 as an
example Knight d6 is a very good point to use in this particular position so
dislocating that c5 pawn. It’s no longer supported and if c5 is won here then White’s
getting a big advantage with material back so g4 is played. Now Rcd1 is
the preferred move here. We have possibilities of Queen g5 now from black
by playing g4. That g4 square vacated so maybe trying to win h6 after. So white
plays this move to double the rooks okay and we have Bishop e8. On another example
King h8 what does white do? Knight e5 Knight takes g4 – that pawn is also a
liability so big advantage for white so the point is after Be8 but instead of Knight e5 here white
played the move king h2. This does get really involved. I am only showing a glimpse of some of the variations and some of them maybe the philosophies behind that. The philosophy we can get for our own games really. If Instead of Kh2 which looks a bit strange, I thought initially the idea was rook h1 and King
g1 and I thought maybe there’s some issue with bishop e4. Before we look at Bishop e4
directly maybe you might think Kg2 and Rh1 is tempting and maybe
something like Rh4 or Rh5 but here Queen g5. This position f5 tempo gainer
and supports g4 but even so, White would still end up with a small
edge in this kind of scenario. Black’s got that bishop pair but white has a lot of
control in this position. So ok King h2 Queen g5 for offering the exchange of
Queens if King h8 here Knight e5 and this
scenario Knights x g4 is possible showing that c7 as a liability. That waiting move
King h8 is not so neutral because of the check and taking here. If we try
something else like Knight d7 then knight c6 and then taking on e7 and then
Rd6 Once white gets this a6 pawn – this
is a strong outside passed pawn with its own perk of this Catalan Bishop
supporting that pawn these scenarios are very favorable for White generally speaking in the Catalan opening. The outside passed a-pawn is especially underlined by the Catalan
Bishop. So big advantage for white there so anyway Queen g5 Queen f4 just offering
exchange of Queens so Queen takes f4 was played. If we look
at an alternative to Queen takes. rook c8 – the rook was attacked
this one is still good. Torture on the queenside
and you can see actually that the b8 knight is again a kind of victim here
basically or here rook takes that knight If it ever moves then bang Bishop takes
and there’s a big punishment with Knight takes f7 check here tactically. So these
variations show that the black pieces are being targeted behind the
scenes a lot. Queen takes G takes Nd7 knight c6 Bh4 – the King goes back
if f3 that interrupts the c6 Knight so that could be taken so the King just
goes back.Nf6 we have Ne5 Nh5 e3 g3 and now f3. It
seems as though F takes is an alternative for example Knight takes
Knight d6 White has a small edge there as well so f3 we have Bishop e7
Knight d7 yes and now actually black gives up the light square Bishop. Rook
takes Rfc8 Knight b6 so more simplification. But now with two pieces
attacked here. So the rook is attacked. The bishop is attacked. The rook goes
to protect the e7 bishop. We have Bishop f1 targeting the a6 pawn
Now Knight f6 hitting the rook. Rook b7 and now to respond to the threat of Bishop
takes a counter threat on e3 is played Nd5 and in fact instead of Knight takes
here we have just the knight coming back. White it seems has for the pawn at
this point succeeded in encouraging a weak pawn here but this pawn is weak
especially to bishop F8 so once these two weaknesses are kind of the
exchange of prisoners as Aron Nimzovich would say
when both sides exchange off weaknesses What is the scenery here – what
does it actually represent? Well White has this rook on the seventh which
nimzowitsch show is also a great fan of a potentially dangerous passed a pawn
if the a6 pawn is ever taken this is still a potentially dangerous passed pawn.
More subtly though that c4 square especially in relation to the rook on
seventh rank and this mass of pawns here can be used by the bishop to establish
itself on c4 for f5 push which is thematic against the French defense
structures. And we’re about to see that now. So first – the first episode from here
is the exchange of prisoners episode so Bishop f8 Kg2 Bishop takes King
takes so now episode 2. The use of the c4 square as a basis for undermining the
pawn chain here with f5 especially with the rook on the seventh rank these two
guys are the next targets. Bishop goes back. e4 for kicking in the knight back to e7 shielding also the seventh rank. If Knight c3 that doesn’t shield the 7th rank or f7 and this scenario if black defends f7
it’s a major cost. A big outside passed pause pawn here. White has got a big
advantage so shielding that 7th rank. Knight d6 rook d7 this is pure torture
now this position. After rook d7 Black is waiting for things like Bishop
c4 and f5. The Knights kind of virtually pinned to f7 but f7 is the one to
dismantle black’s pawn chain in the center and expose King safety issues so
ok we have a5 at least the bishop not going to be winning
the a pawn – Bishop c4 but what a beautiful place for the bishop. So one of
the classic Fisher vs Spassky 1972 games had a bishop on c4. Really
nice but here you can see the power of f5 incoming f5 is hugely
powerful now in this position not only that what is also happened is that King
qualitatively is much stronger than the black King. It’s much more out and about
this King on g3 so blacks pieces are passive although it’s opposite color
bishops this seems very bad for black in fact we have more simplification knight
c8 the Knights come off and this gives up the a-pawn potentially so rook a7
Bishop d6 and that pawn is taken so an outside passed pawn. Now yeah
blacks in big trouble and there is still this idea of undermining the pawn chain
once white is ready by unpinning the f5 pawn. The f5 push is going to be extremely
dangerous thematically for black and we see that soon. e5 and now rook takes c5
actually so black volunteered the c5 pawn here. The bishop didn’t go to e7 because rook a7 and F5 is really good. Stuff like that is really good. If it went to b8
Ra8 is a nasty pen and this scenario for example that a pawn is winning here with
this kind of scenario. It’s going to be winning material yeah it’s
just hopeless for black so we have Bishop c7 and now takes c5 Ke7
a5 King d8 a6 Bishop b6 rook b5 Bishop a7 and now finally here we
have the move f5. So yeah this reflects on black’s earlier moves right in the opening. It seems that c4 square is actually useful for something
in a strategic sense in a long-term strategic sense. Perching the bishop on c4 is a great use of the c4 square sometimes as well just it’s undermining this pawn
chain now and with the King on g3 this is only like a temporary pawn
sacrifice. If white basically ends up getting two connected passed pawns
that’s extremely dangerous and in fact here rook b8 was played. If exf5 then
the King walks in with King F4 taking here. Two connected passed pawns
it’s a massive position for white so rook b8 pleading more exchanges
rook a8 is played. Ra5 exf5 bishop takes yeah this f5 looks
vulnerable now. Bishop b6 Ra2 Rc8 King comes to f4. h5 and we’ve got now two connected passed pawns here as well as a passed pawn here. Is black overloaded
despite the opposite colored bishops? We see Rc7 and this pawn is just
given up basically. Yeah with Rc7. That’s taken and it does seem as
though there’s a lot of good options for white in this position. A lot so yeah
this is rather desperate. Rc7. that’s taken so there’s two connected
passed pawns now here and one there so the opposite colored bishop seem very
unlikely to help. So a6 is exchanged for b4.
Two connected passed pawns here as well as the third pawn over here it’s all over
bar the shouting. The great Leela technique is about to be witnessed for
this end game phase. So what is the Leela technique here for this end game?
Well pushing pawns seems fair enough so far nothing too special there and yep
the pawns are ready to roll. Giving up b5 there are still two connected passed
pawns and now e7 is going to be the one supported to push these pawns through. This was desperate from Black Rc7. White’s still in time now like in a famous
Magnus Carlsen vs Sergey Karjakin game we once covered where similar kind of scenario
opposite colored bishops Magnus Carlsen eventually crashed through despite the
opposite color bishops just about. So King takes queening and the game at move 103 here carried on to move 212. It is a particular
version of Leela – maybe with greater Benny Hill aspects. So okay high level
shuffling and it continued it continued yes under promoting a knight
okay so Knight came over here and curious yes it was curious. This is just
for the record it’s not the official version of Leela honest so Leela really having fun here in
this winning position a kind of celebration of mockery having a winning
position giving up the Knight. Eventually the Queen and King kind of
cooperate when they’re bothered to checkmate the black King eventually okay
yeah so it seems even if stockfish is supercharged in a nutshell this Catalan
variation we can say it seems to be a bit dodgy in a nutshell. This Catalan
variation we can also say that containment of pieces also sometimes
that this has this aspect of winning those pieces that are contained I
believe from variations in the game. If you enjoyed this game video please click
on the top left box which should appear shortly become a member at www.chessworld.net to play against other youtubers. You can also test yourself on the variations
covered on the games from the improve menu – there’s a puzzle book
option which also has a link to the annotated game. Comments, questions
donations – see description. Likes shares subscribes with notification bell really
appreciated. Thanks very much. Challenge me for a game at www.chessworld.net 🙂

35 thoughts on “Highly Evolved Leela plays “Bluefish” – a Supercharged Three Times Faster Stockfish in Catalan

  1. Q. Can Neural networks find new positional Opening Gambits previously uncharted ? 🤔😀😎🌍 Video: http://bit.ly/2VzGsHA

    Replayable game: http://bit.ly/2GFXr1l

    #chess #chessgame #ArtificialIntelligence #AI #MachineLearning #DeepLearning

  2. Loving the intro KC! Very good to know for people not involved so much in tcec what these engines are all about.

  3. Ya a shame there are not more bookless games. Leela with her wings on full display is extraordinary. This is why I very rarely tune into tcec and never donate.

  4. Bluefish was running on the 176-thread analysis box. The GUI said 170 threads but that might have been a bug. I believe Leela was running on the regular GPU hardware, so quite a mismatch in that sense, but fascinating that Stock(blue)fish still wasn't able to come up with any new ideas despite ~115-150 million nodes per second.

  5. Nice analysis!
    Afaik the reason for Leela trolling so heavily in this endgame was the missing DTZ tablebase data on the gpu server (only has wdl info) in conjunction with no adjudication used for the game. Usually Leela would use the tb to play the fastest dtz moves and end the game much earlier (as she does on CCC). So it's not a net issue but just a quirk of the settings/setup tcec used for this game.

  6. I find this whole Alpha/Leela thing confusing. So Alpha plays a crippled Stockfish without book and claims to be stronger? Shouldn't they use the same processors and each be configured to play their best?

  7. @Kc I have an important question. Was this leela in a much stronger computer or not? Becouse if it wasn't leela is already much stronger that stockfish.

  8. Wow. Obviously chess programs aren't perfected yet — not just with the underpromotion/endgame inefficiency — but Bluefish's putting a Pawn on his QR3 when it was already isolated was probably THE losing move, if you had to pick one.

  9. Interesting game, reminds me of hydra an old chess engine with huge computing power, I guess quality is always better than quantity.

  10. could you add the names of the players to the video, so that we can jump in at any point and know who is playing which side?

  11. I haven't checked the code of Leela, but I think it dances around when probability of winning is equal among options and it chooses the move randomly. It has no understanding of shortest path to winning. Basically in the endgame pretty much all moves have equal chances of winning (albeit in different move counts) except those that cause stalemate.

  12. at 27:17 after black plays Nc8, why doesn't white go for Nf7, which seems obvious for Be6. Am I missing an obvious counter?

  13. Still find it bizarre that a computer that can beat a computer that can beat Kasparov can't see how to promote a second queen and win quickly.

Leave a Reply

Your email address will not be published. Required fields are marked *