(;GM[1]FF[4]CA[UTF-8]AP[CGoban:3]ST[2]
RU[Chinese]SZ[19]KM[7.50]
PW[AlphaGo]PB[AlphaGo]C[Here is my commentary on the 38th AlphaGo self-play game. This is a combination of my attempts to explain what is going on, me wondering what the hell is going on and commentary on my thoughts and feelings on the game. One thing I've focused on is trying to work out some rationale for the order of the moves: AlphaGo seems to have a tendency to flit around the board, making exchanges at what may appear random times. Is there some meaning behind these moves: in that they are probes and depending on how one is answered play will resume at another place in a different way, or are they just Monte-Carlo style random exchanges which could have been made another time? [spoiler: I think some of both\]. I chose this game in particular because Kim Seongjin 7d and Stanislaw Frejlak 6d played a similar opening in their game at the Vienna tournament, and I though Stanislaw quickly fell behind following Seongjin's AlphaGo-esque early 3-3 invasion. My suspicion was he blocked on the wrong side, and indeed in this game AlphaGo blocked on the other side.]RE[W+Resign]
(;B[pd]
;W[dp]
;B[pq]
;W[dc]
;B[de]C[In the Master games AlphaGo displayed a marked preference for low over high approach to 3-4 which is different to pros these days (a long time ago i.e. 19th century Japan low approach was the only move, but then in the 20th century high became more popular and in modern pro games is played maybe a bit more than the low). Master on the other hand was maybe only a fifth or so high (one pattern is it would do it if the opponent had thickness or a moyo behind it). I'm not familiar with these 50 games yet, but AlphaGo now seems to like the high approach again, and more so than low. It's also noticeable that if the opponent attaches under and it hanes then it now seems to like the solid rather than hanging connection, a change from before. Commentary from one of the 5 WeiqiTV self-play games suggests this may be because it thinks solid connection is more sente and it's too easy for the corner player to tenuki after hanging connection and the follow up is not so severe.

I'm still not clear if these self-play games were with the same version as the Master games in new year 2016-17 and the Ke Jie match: it seems quite different to me (playing high approach, even earlier 3-3, more flitting tenukis and exchanges, very complicated fighting). Part of this could be with the same version against humans it gets an early lead and so plays simply, whereas against itself the game remains close so it has to play all-out so we see more crazy style. However, things like the high approach and early 3-3 are in positions so early the humans can't have been losing yet so it couldn't yet switch so safe and simple mode.]
;W[ce]
;B[cf]
;W[cd]
(;B[cq]C[These early 3-3 invasions are the hallmark of new AlphaGo. The standard human thinking was it was bad to invade so early because white's thickness would be better than black's territory]
;W[dq]
;B[cp]
(;W[cn]C[This jump is a human move, but was less common than the very natural hane at the head of 2. A common reason to play this move was to get sente]
(;B[co]
;W[do]
;B[bn]C[AlphaGo's joseki for 3-3 invasion. Note that white ends up with a wall of 3 rather than 2 stones so it's a bit stronger and the e5 peep isn't there like if you hane at head of 2]
(;W[df]C[White leaves the wall and cuts here where black didn't defend.]
;B[hp]C[As we so often see in these games: following a 3-3 invasion of a 4-4 stone the resulting 4-4 wall group gets attacked.]
(;W[qo]C[White doesn't answer: any answer would likely be a good exchange for black as it stopped white making a decent sized extension on the lower side.]
(;B[dn]C[This is the agressive followup when white didn't spend a move on the wall.]
;W[cm]
;B[bm]
(;W[op]C[This tenuki surprised me. Why is it played now? AlphaGo does tend to like this press a lot.]
(;B[qp]C[This answer was even more surprising. My understandind is it seeke quick life in the corner and avoids crawling towards h4 which could help white in the fight in that area (or become overconcentrated).]
(;W[cl]C[The bottom left fight resumes after the p4 r4 exchange. Would white have continued differently if r4 was at p3? I don't know.]
;B[en]
;W[cr]
;B[br]
(;W[bq]C[A tesuji/probe]
(;B[aq]C[This atari surprised me initially: ]
(;W[gp]C[White wants to get out or settle on the side]
(;B[gq]
;W[bs]C[But why this atari now rather than last move? White would have connected before, so if white didn't play this before (because lose ko threat) then how did g4 g3 exchange change that? Was there an answer other than g3 in which white wouldn't play this atari?]
;B[ar]
;W[go]
(;B[fp]TR[dn][co][cp][cq][dr]C[Much more powerful shape attack than my instinct, If you imagine a black stone at d2 exchanged for e2, you can see this is the vital shape point of the 3 white stones and mouth shape from d6 and d3.]
;W[fq]
;B[fr]
(;W[eq]
(;B[er]C[Surprised me initially]
;W[dr]
;B[hq]
(;W[fn]
(;B[dl]
;W[eo]C[This way white gets a real eye and takes liberties so f7 also sente, but loses the 3.]
(;B[ck]C[Fairly big and makes d6 safe, but corner was already alive and white got d14 already so left not too big potential for black.]
;W[in]C[Makes sense: use f6 to jump further and next could k4 to attack black]
(;B[kp]C[Makes sense too]
(;W[pp]
;B[qq]
(;W[ro]
(;B[nr]
(;W[pk]C[I expected a further extension like q11, but I suppose that's too far: we often see AlphaGo showing us that a wall we thought was thickness actually can be weak]
;B[pn]C[Similar to the peep against wall from 3-3 invasion under 4-4.]
(;W[po]C[White concedes. Now p4 would look better at o4.]
(;B[lo]C[Surprised me, seems slow, what is it defending against? Also has follow-up at o6. Still not clear what white would to if black doesn't play this.]
(;W[qc]C[Typically AlphaGo 3-3 invasion, hardly surprising anymore!]
(;B[pc]C[Which side to block?]
;W[qd]
(;B[qf]
(;W[qe]
;B[pe]
;W[rf]C[AlphaGo joseki]
(;B[qb]C[Often it tenukis instead as black now, but where to?]
;W[mp]C[Why now? And not before 3-3?]
(;B[lq]C[Solid answer, but can black tenuki? That this was played in middle of top right corner sequence suggest there is a link, or is it just Monte-Carlo random exchanges? Not sure, I don't understand this timing.]
;W[rb]C[Ok, this makes sense]
;B[rg]
(;W[pf]C[I used to find this move (and similar one in push only twice joseki, see 2nd var on move 59 then 3rd on 61) weird as you can save the one stone in sente, but it's to make bad aji on outside with ladder and liberty troubles after q18.]
;B[qg]
;W[pb]
(;B[re]
(;W[qa]
;B[of]
(;W[no]C[Big surprise. Threatens a ko, and helps reduce q6 aji, but why not q13? Black doesn't have ladder or net.]
;B[jm]C[This makes sense, makes the q13 ladder work whilst pressuring white]
;W[gl]C[If white thought k7 was sente maybe that explains why no q13]
;B[hl]
(;W[hm]C[Puzzling. AlphaGo does seem to place a high value on making a group have 2 eyes locally, but this seems submissive to me as I think a good exchange for black (outside for inside) so I'd want to come out instead.]
;B[ho]C[Is it sente?]
(;W[hn]C[Well I suppose it was sente then. Not totally clear to me it was or why it needed to be played now (white wasn't going to reverse sente at h5 anytime soon was he!).]
;B[rd]
;W[rc]C[A little nice for black to get that territory gain in sente, but white didn't want to make black take s14 off the board as then q13 is net rather than ladder.]
;B[cg]C[A big move (and my idea of white using sente to play here for 80 was probably a decent place to use sente)]
;W[ef]
(;B[eh]C[Hmmm, is this biggest place?]
;W[mr]C[So white makes use of ignored o5 to start the ko]
;B[mq]
;W[or]
;B[oq]
;W[nq]
;B[np]
;W[bp]
;B[bo]
;W[nq]
;B[fo]
(;W[gn]
;B[np]
;W[bk]
(;B[cj]C[If black keeps on answering this threat is a bit lossy, as helped black fix c11/d11 area weakness.]
;W[nq]
;B[ns]C[1 local threat]
;W[os]
;B[np]
;W[el]
(;B[nq]
;W[bl]C[Nice way to win ko bigly. Although black winning the ko does mean white lost a few points around p2 and has a nice peep at p6, I can't help but feel this is a bad result for black. Would it really have been worse for black to play e12 at o3 to defend agaisnt the ko?]
;B[bj]
;W[am]
;B[ap]
(;W[nk]C[Defends the right group (which would suffer if black played p6 which is a peep now that a black stone at o4 from winning ko). Also ladder breaker on q13.]
;B[qj]C[This makes net on q13 work again whilst threatening to go into white eyespace and making some for black. Locally speaking a bad shape, but once white has o9 black gives up on doing meaner things to that white group. And gets sente.]
(;W[qk]
;B[fc]C[Surprised me, though not sure what else I would do]
(;W[gd]C[More active way to make e14 safe and press black low.]
;B[gc]
(;W[id]
;B[fd]
(;W[rj]C[Surprising, but it's sente and quite large]
(;B[mg]C[Active way to defend against q13 running out. White's getting a lot of value out of that stone. This kind of bullying is why we often think it's honte to capture the ladder stone quickly instead of looser moves that allow it be netted.]
(;W[di]C[A shape weakness of black]
(;B[an]C[So black lives and creates emergency connection at a9]
;W[dh]C[Makes sense]
;B[dg]
;W[eg]
;B[ei]
(;W[ej]
(;B[gh]
(;W[gg]C[White prevents e15.]
;B[gf]C[Sharp move: getting this in before d10 means it's harder for white to capture f17 stones and although white can then capture g12 its served its purpose of preventing f11 ladder so black can d10 cut.]
;W[fg]
;B[dj]
;W[ic]C[So white tries to kill the top. The following fighting is rather above my pay grade, so I just sit back and watch]
;B[hd]
;W[he]
;B[ib]C[Fancy tesuji, like the one in the taisha that Shusaku famously didn't find in the ear-reddening game]
(;W[jb]
;B[hb]
;W[kc]
;B[hf]
;W[hc]
;B[ee]
;W[fh]
(;B[db]C[What?!]
;W[fi]
;B[cb]
;W[ci]C[Black gets some points here, but losing c14 means he doesn't kill the corner so feels good exchange for white to me, though I didn't count anything]
;B[dm]
;W[em]
;B[dk]
;W[bi]
;B[ak]
;W[bc]
;B[bb]
;W[nh]C[Now it's kind of endgame]
;B[pi]
;W[mn]
;B[nc]
;W[if]
;B[ge]
;W[hg]
;B[hd]
;W[ie]
;B[dd]
;W[ng]
;B[nf]
;W[be]
;B[cc]
;W[ad]
;B[ri]
;W[pj]
;B[qi]
;W[kn]
;B[lg]
;W[oc]
;B[mi]
;W[oi]
;B[md]
;W[lj]
;B[li]
;W[kj]
;B[ob]
;W[io]
;B[jh]
;W[nb]
;B[od]
;W[jq]
;B[jr]
;W[ii]
;B[ab]
;W[jp]
;B[kr]
;W[af]
;B[oa]
;W[pa]
;B[ln]
;W[lm]
;B[mm]
;W[mo]
;B[nj]
;W[mj]
;B[ji]
;W[ni]
;B[jj]
;W[jk]
;B[ko]
;W[jn]
;B[ik]
;W[jl]
;B[rn]
;W[rp]
;B[kl]
;W[km]
;B[sc]
;W[sb]
;B[rm]
;W[qm]
;B[qn]
;W[on]
;B[om]
;W[nm]
;B[rk]
;W[rl]
;B[sj]
;W[sl]
;B[kk]
;W[im]
;B[ig]
;W[mb]
;B[mk]
;W[oj]
;B[ml]
;W[nl]
;B[jd]
;W[le]
;B[ja]
;W[jg])
(;B[fi]C[Why not answer?]
;W[gi]
;B[fj]
;W[gj]
(;B[ek]
;W[fk]
;B[ej]
(;W[hh]
;B[hg]C[Isn't black ok?])
(;W[hg]
(;B[if]
;W[ge]
;B[fe]
;W[ff]
;B[ie]
;W[hd]C[Is black in trouble here?]
(;B[hh]C[If can't capture white then dead inside?])
(;B[ec]
;W[dd]
;B[db]
;W[cb]
(;B[fb]
;W[da]
;B[ea]C[ko])
(;B[da]
;W[ia]
;B[ga]
;W[fb]
;B[eb]
;W[fa]C[dead])))
(;B[ge]
;W[if])))
(;B[fk])))
(;W[hb]
;B[hc]
;W[jb]
;B[hf]))
(;W[dj]
;B[ee]C[This is why g12 is sente.]
(;W[gf]
;B[ff]
;W[fg]
;B[gg]
;W[fe]
;B[fh]
;W[ff]
;B[hf]
;W[ge]
;B[hd]C[black threatens this.])
(;W[gg]
;B[gf]
;W[fg]
;B[hg])
(;W[ff]
;B[gf])))
(;B[dj]
;W[fi]C[can't cut yet because of ladder]))
(;W[dj]C[What if here?]
(;B[dk]C[here?]
;W[ej]
;B[dm]C[I think this is smaller than centre]
(;W[fi])
(;W[bh]))
(;B[gh]
(;W[dk]
(;B[ch])
(;B[ee]C[would capture these 3]))
(;W[gg]C[so if white defends, what now?]
;B[dk]
;W[ej]
(;B[hg]C[keep on fighting in centre?])
(;B[dm]C[or take the cash?]
;W[gi]
;B[ib])))))
(;B[dg]C[What if black connects? Can lower left die?]
(;W[an]C[This doesn't work...]
;B[ak]
;W[dm]
;B[ao]
;W[dk]
;B[al]
;W[bh]
;B[ch]
;W[ci]
;B[bi]
;W[am]
;B[an]C[Black is safe])
(;W[dk]C[but white here is sente to kill]
(;B[an])
(;B[dj]
;W[dm]
;B[an]
;W[ei]))
(;W[dm]
;B[an]
(;W[dk]C[Like this black is cut and c13 seems in some trouble])
(;W[ge]
;B[ak]))))
(;W[pg]
;B[ph]
;W[og]
;B[oh]
;W[ng]
;B[nf]
;W[nh]
;B[ni]C[This ladder works]))
(;B[ge]C[if tenuki]
;W[pg]
;B[oh]
;W[ph]
;B[pi]
;W[qh]
;B[rh]
;W[qi]C[white can escape]))
(;W[ge]))
(;W[hd]C[other idea]))
(;W[ed]C[This would be passive]
;B[ic]C[I suppose black might extend on the top. Although it's a wide area it seems hard for either to make much there.]))
(;W[pg]
;B[oh]
;W[ph]
;B[pi]
;W[qh]
;B[rh]))
(;W[an]C[Might look like white can kill]
;B[ak]
(;W[dm]
;B[ao]C[white short of libs to connect]
(;W[al]
;B[em])
(;W[dk]
;B[al]
;W[am]
;B[aj]C[though not actually eye on edge]))
(;W[al]
;B[dm])))
(;B[bl]
;W[nq]C[Black has no threats, and the ko is so big]
;B[ee]C[Probably this is biggest]
(;W[ms]
;B[ff]
;W[fd]C[W should be good. H7 answer to live locally instead of running out does make some more sense now.])
(;W[fe]
;B[np]
;W[em]
;B[nq]
;W[dm]C[Not so big for white])))
(;B[bl]
;W[nq]
;B[ns]
;W[os]
;B[np]
;W[cj]
;B[nq]
;W[dk]C[This would be even worse for black than game]))
(;W[ep]C[Is it worse? Normally we like to capture stones early, and white will have to eventually capture the stones anyway]))
(;B[nq]C[If defend ko]
(;W[ci]C[white here?])
(;W[dh]C[or here?])
(;W[cj]C[or something fancy?])))
(;W[cg]
;B[gn]
;W[fo]
;B[gm]C[impossible?]
(;W[em]
(;B[dm]
;W[hn])
(;B[fl]
;W[dm]
;B[gk]C[trade good for black]))
(;W[fm]
;B[hn]
;W[dm]C[Maybe good for black?])))
(;W[gk]C[My instinct was to come out. Why not?]
(;B[gm]
;W[fm]
;B[hn]
;W[ho]
;B[io]C[Can cut off 1 stone]
;W[dm]C[But f7 was sente to resurrect these stones, good for white])
(;B[hm]
;W[el]
;B[dm]
;W[hk]C[lose 1 stone but come out, seems good for white to me])
(;B[hk]
;W[gj]C[push from behind?])))
(;W[pg]C[Why doesn't white run out? The ladder doesn't work for black. I don't have a satisfactory answer]
(;B[ph]
;W[og])
(;B[oh]C[Because white didn't make s16 t14 exchange this doesn't capture]
;W[ph]
;B[pi]
;W[qh]
;B[qi]
;W[rh]
;B[ri]
;W[sg])
(;B[jm]C[Does black have a good ladder maker?]
;W[ph]
;B[hl]C[Can white live/escape?]
;W[fl]
;B[fk]
;W[em]
;B[dm]
;W[gl]
;B[gk]
;W[hm]
;B[im]
;W[ho]C[Alive, so I think this should be good for white as black lost so much on top right]
;B[qi]
;W[pi]
;B[qj]
;W[pj]C[I suppose can live like this]
;B[nd])))
(;W[ob]C[If here then black gets a nice squeeze]
;B[rc]
;W[qa]
;B[rd]
;W[qb]C[Poking out at p18 is nice, but s17 etc in sente is nice too, probably good for black as white low and not alive yet.]
(;B[of]C[Black doesn't have this ladder])
(;B[nd]))
(;W[rd]C[Interesting that white doesn't play this atari]
(;B[sf]C[Would black answer? If so white can now extend]
;W[ob]
;B[of]
;W[pg]
;B[oh]C[Because of s16 exchange black has s14 liberty so this net/loose ladder works. I think this is more of a plus for black than white's plus of getting p18, so good for black]
;W[ph]
;B[pi]
;W[qh]
;B[qi]
;W[rh]
;B[ri])
(;B[ob]C[The other choice would be to now atari here. Is s15 for s16 a good or bad exchange for white or black? If white gets to capture at t15 it's a little negative for black (white has more eyeshape and better yose at t13).]
(;W[qa]C[But I think white would answer here as p17 cut and reducing liberties ot q17 stones is important]
(;B[nc])
(;B[of]
;W[pg]
;B[oh]
;W[ph]
;B[pi]
;W[qh]
;B[qi]
;W[rh]C[Because of s15 for s16 exchange black's 3 stones aren't dead yet..]
;B[sf]
;W[ri]
;B[rj]
;W[qj]C[but black collapses])
(;B[nd]))
(;W[se]))))
(;B[ob]C[This is an instinct, but then...]
;W[qa]
(;B[nc]C[you kinda need to defend the cut])
(;B[of]C[Is p18 for r19 really a good exchange now? Maybe not.]
;W[oc]C[there is a ladder-dependent cut here])))
(;W[qg]C[Can't do this with r18 exchange]
;B[pf]
;W[rh]
;B[re]
;W[sg]
;B[rc]C[resign]))
(;B[rg]
(;W[mr]C[Is white follow up here?]
;B[nq]
;W[mq]
;B[np]
(;W[no]C[Can't block here can we?]
;B[mo]
;W[ns]
;B[or]C[corner is not alive but n2 should die first]
;W[jq]
;B[kq]
;W[kr]
(;B[lr]
;W[lq]
;B[jr]
;W[ls]
;B[jp])
(;B[jp]
;W[jr]
;B[rp]))
(;W[mo]
;B[no]
;W[mn]C[urgh what is happening?]))
(;W[lq]C[or is it here? seems rather soft])))
(;B[jc]
;W[qg]
;B[pf]
;W[qb]C[I think white is good to defend corner and grow right side.])
(;B[rg]C[If this one...]
;W[qg]
;B[pf]
;W[rh]
;B[re]
;W[sg]
;B[qb]
;W[rd]
;B[rc]
;W[se]C[Seems good for white: like the outside ponnuki after double hane variation, but white also has some of the corner, though black wall a little bigger.]))
(;W[rf]
;B[rg]
;W[re]
;B[qg]
;W[pb]
;B[ob]
;W[qb]
;B[pe]C[Probably white doesn't like being sealed in, q9 getting weaker with s13 turn present.]))
(;B[qe]C[AlphaGo normally does the jump, but what happens if here?]
;W[re]
(;B[qf]C[If the simple human joseki]
;W[rf]
;B[qg]C[I expect white would take sente now]
(;W[jd]C[here?])
(;W[pb]
;B[ob]
;W[qb]
;B[nc]C[Don't play this, makes black stronger.])
(;W[kd]))
(;B[rf]
;W[qf]
;B[pe]
;W[rg]
;B[rd]
;W[sf]
;B[rc]
;W[pf]C[Probably white spends a move here, and right side is quite nice.])
(;B[qb]C[This (instead of extend or double hane) is a not-so-common human joseki (about 100 of 1000 games in waltheri). Compare with game move 68.]
;W[rb]
;B[rf]
;W[pe]
;B[qf]
;W[pb]
;B[ob]
;W[qa]
;B[nc]
;W[oe]
;B[nd])))
(;B[qd]C[I think this is also possible, but it blocks top side because that is more open area?]))
(;W[mq]
;B[nq]
;W[np]
;B[mp]C[allows this cut])
(;W[qf]C[I might want to approach here to make the right side into territory, but I suppose that's hard to do so AG just likes easy points in corner and also the resulting wall isn't alive.])
(;W[mp]C[See move 64]
;B[lq]C[Would black answer here now?]))
(;B[qf]C[Say black does some big opening move over here]
(;W[mq]C[Is this a problem?]
;B[nq]
(;W[np]
(;B[lq]C[If black has this to connect is it ok. Still rather low and n4/5 sente is nice for white to reduce q6 worries]
(;W[mp]
;B[mr])
(;W[mr]C[Can't]
;B[mp]
;W[mo]
;B[lp]C[w fail]))
(;B[mp]C[can black cut?]
;W[lp]
;B[mo]
(;W[kq]
;B[lo]
;W[lq]
;B[jp]C[W looks dead])
(;W[lo]C[How about this?]
;B[lq]
;W[kq]
;B[mr]
;W[lr]C[can white get this?]
(;B[mq]
;W[jp]
(;B[ko]
;W[jo]
;B[ln]
;W[jr]C[Maybe this is the plan, to kill h4. Problably good for white, though next...]
;B[pl]C[White also has troubles])
(;B[kr]
;W[ko]
;B[jr]C[Is this so bad for black? Getting forced low is sad, but n5 split is big and white probably needs to defend.]
;W[on]C[Black gets sente again]
(;B[jc])
(;B[oc])))
(;B[ko])
(;B[jq]))))
(;W[mo]
(;B[np]
;W[no]
;B[mp]
;W[lo]
;B[lp]C[Maybe this is white's plan?])
(;B[mp]C[better?]
;W[lo]
;B[lp]C[Not clear to me white gained, not really thick and black got 1 tenuki and territory]
;W[ko]))
(;W[mp]
;B[np]
;W[no]
;B[mo]
;W[lo]
;B[mn]
;W[lp]C[Is black worried about this fight? White seems to have troubles too.]
;B[nn]
;W[oo]
(;B[on]
;W[qm])
(;B[pl])))
(;W[mp]
;B[lq]C[I don't feel sad for black given I got 1 tenuki])
(;W[nq]
;B[mq]C[Black's ok.])
(;W[lp]C[Is it some crazy attachment?]
(;B[lq]
;W[mq]
;B[mp]
;W[lo]
(;B[kq]
;W[mr]
;B[nq]
;W[np]
;B[mo]
;W[mn]C[aha, this is starting to look like a plan for white.])
(;B[nq]
;W[kq]
;B[mr]
;W[lr]
;B[mq]
;W[jp]))
(;B[kq]
;W[mq]C[Now black can't cut]
;B[mr]
;W[ko]C[White can force black low and cover outside, but gote and not thick, is this good for white? I'm doubtful.])
(;B[ko]
;W[mq]C[Now can cover]
;B[mr]
;W[lr]
;B[kq]
;W[lq]
(;B[rp]C[If black needs to spend a move (does he?) then I can see splitting in sente is nice for white])
(;B[oc]
;W[sr]
;B[rr]
;W[sp]
;B[rq]
;W[sq]
(;B[pr]
;W[qs]
;B[rs]
;W[oq]
;B[or]
;W[ms]
;B[ns]
;W[ps])
(;B[qs]
;W[pr]
;B[or]
;W[ps]))))))
(;W[on]C[Can white not resist was my feeling?]
(;B[po]
;W[pm]
;B[qn]
;W[rn]
;B[qm]
;W[ql]
;B[rm]
;W[rl]
;B[om]
;W[sm]
;B[pl]
;W[oo]
;B[pm]
;W[nm]C[It's a ladder, but maybe black would tenuki and it's so bad aji]
;B[ol]
;W[nk]C[Or net. Hmmm...])
(;B[pm]
;W[om]
;B[po]
;W[oo]
;B[rm]
;W[pl]
;B[rn])
(;B[pl]C[Ah, maybe here]
;W[ol]
;B[om]
;W[nm]
;B[pm]
;W[ok]
;B[oo]
;W[no]
;B[po]
(;W[nn]
;B[np])
(;W[mp]
;B[nn]
;W[mn]
;B[rn]))))
(;W[pi]
(;B[pn]
;W[po]
(;B[pk]C[problems like this?])
(;B[nn])
(;B[pl]))
(;B[qg]C[Maybe black would answer here first, and then if white ignores do mean things]
(;W[cg]
;B[pn]
;W[po]
;B[pk]
;W[oh]
;B[nn]C[white suffering])
(;W[qh]
;B[pg]
;W[ok]C[Is there any good way to fix? Seems not.])
(;W[pk]C[ugh]))))
(;B[nq]
;W[oq]
;B[or]
;W[nr]
;B[mr]
;W[pr]
;B[ns]
;W[qr]C[b fails]))
(;W[pn]C[This would be thicker on outside, but s5 is huge for territory and also means black could now o3 jump]
;B[nq]
;W[oq]
;B[or]
;W[nr]
;B[mr]
;W[pr]
;B[ns]
;W[qr]
;B[rr]C[r2 doesnt' capture (well there is a ko)]
;W[rq]
;B[rp]
;W[sr]
(;B[sq]
;W[rs]C[famous ko shape])
(;B[rs]
;W[sp])))
(;W[oq]C[I had previously expected this one as the follow-up if white got sente, but now that black has L4 I can see building a wall in this direction is not so useful]
;B[ro]
;W[pp]
;B[qq]
;W[qn]
;B[rn]
;W[ql]
(;B[nc]C[Would black tenuki?]
(;W[kn]C[If so would white play here?]
(;B[lq]C[Maybe black's happy to defend and make points?])
(;B[mq]))
(;W[kr]C[or maybe steal the base/territory?]
;B[kq]
;W[lr]
;B[kn]C[j6 not alive either though, don't know how to judge.]))
(;B[kn]C[or here?])
(;B[mq]C[or here?])))
(;B[oq]C[Is tenuki possible? Seems not]
;W[jp]
(;B[jr]C[Probably white is happy to press black low in sente])
(;B[np]
;W[jr]C[don't want to die, much bigger than p3/o4])
(;B[jq]
;W[kq]
;B[kp]C[How about this? It's a common shape]
;W[lp]
;B[ko]
;W[jo]
;B[lq]
;W[kr]
(;B[lr]C[Can't play this]
;W[jr]
;B[iq]
;W[mp]
(;B[ir]
;W[kn]C[White has ladder (or squeeze if not).])
(;B[jn]C[sometimes this kind of thing is a tesuji to break ladder, but]
;W[kn]
;B[lo]
;W[ln]C[still ladder])
(;B[kn]
;W[mr]
;B[ir]
;W[ls]SQ[hp][gq][hq][iq][jq][er][fr][ir]C[Black dead inside]))
(;B[jr]
;W[lr]
;B[mq]
;W[mp]C[Make 2 ladders]
(;B[mr]
;W[kn])
(;B[kn]
;W[nq])))))
(;B[in]
;W[el]C[can't tenuki]))
(;B[eo]C[Impossible]
;W[em]))
(;W[fo]C[This has false eye problem]
(;B[eo]C[now? loses lib]
;W[ep]
;B[el]C[Black aims h6 or c9])
(;B[dl]C[or would it be here?])))
(;B[fo]C[What if here?]
(;W[gr]
;B[hq]
;W[er]C[Now white getting b1 atari is useful to ensure white's eyes.]
;B[hr]
;W[fs]
(;B[hn]LB[gl:B][fm:A]C[Net rather than ladder I think though it does leave some nice sentes for white like A/B I think a ladder breaker on top right hurts more. Although white lived in sente I thought black was ok here, but AG didn't like it for black? Because white got sente? Edit: white wouldn't g2 for 40 but g6 so probably AG agrees with me this is good for black, but wouldn't happen.]
(;W[pp]C[Here now?])
(;W[cg]C[or here? big points and helping c8])
(;W[fm]
(;B[gn]
;W[cg]
;B[el]C[black has a liberty problem to play here]
;W[em]
;B[dm]
;W[dl]
;B[ek]
;W[fn])
(;B[fn]C[This keeps e8 alive but gives white more forcing moves]
(;W[gm]
;B[gn])
(;W[gl]C[or this]))))
(;B[gn]
;W[pf]
(;B[nd]
;W[ho]
;B[io]
;W[hn]
;B[gm]C[Ladder no longer works, but maybe black could keep fighting as white is rather bad shape for now, but getting cut is huge.])
(;B[im]
;W[nd]C[Humans think this double high approach is very painful for black, but given how AlphaGo likes 3-3 invasions maybe it wouldn't mind so much?]
;B[pc]
;W[nf]
;B[nb]C[I could imagine AlphaGo not minding this black territory and saying white outside not so thick/amazing])))
(;W[gn]C[Aha, can white fight?]
(;B[hr]C[Can't]
;W[fn]
;B[eo]
;W[em])
(;B[el]
;W[hq]
;B[gr]
;W[hr]C[Black collapses])
(;B[dl]
;W[hq]
;B[gr]
;W[hr]
;B[ck]C[Even though d8 would let black capture c8 3 stones, g3 dying is so much bigger: f5 cutting stone meaningless now.]))))
(;W[fo]
;B[eq]
(;W[eo]
;B[dr]
;W[fm]C[White could connect but black got a big territory and safety and sente so I think this should be good for black, but not so awful I'd resign as white.])
(;W[er]
;B[eo]C[white short of libs])
(;W[dr]
;B[eo]
;W[ep]
;B[fq]
;W[hq]C[At a minimum black can capture the 6 white stones.]
(;B[hr]
;W[iq]
;B[er])
(;B[iq]
;W[hr]
;B[ir]
;W[ho]
;B[ip]
;W[er]
(;B[hs]C[If here]
;W[fs]
;B[gr]
;W[fm]C[Might think white has this tesuji])
(;B[es]C[but]
;W[gr]
;B[fs]C["Golden chicken standing on one leg" tesuji, white can't approach]
;W[fm]C[White can get outside but too much sacrifice])))))
(;B[fq]C[My instinct was here, but then white can connect out]
;W[eo]))
(;B[ho])
(;B[hq])
(;B[go]))
(;W[fo]
;B[eo]
;W[fp]
;B[ep]
;W[eq]
;B[fq]C[oops.])
(;W[bs]C[Will black not connect now?]
(;B[dr]
;W[ar]
(;B[er]
;W[bo]C[fail])
(;B[bp]
;W[ap]
;B[bo]
;W[er]C[fail]))
(;B[ar]C[So black would connect now])))
(;B[dr]C[So the first question is what happens if black captures the 2nd line stone?]
(;W[cs]
(;B[ds]C[assume for now black keeps on going..]
;W[bo]
;B[bs]
;W[cr]
;B[cs]
;W[bp]
;B[cr]
;W[bl]C[White captures 2 stones to settle c8 and develop left side]
(;B[eq]C[Black can firmly capture the 3 white stones here. How to judge this result? Given that black didn't choose this line, I think we can assume AlphaGo thought it was good for white. That seems plausible to me: in terms of territory in the corner both got about the same, but white got strong on the left side so it will be harder for black to settle c14 (e.g. c13 e14 c10, or maybe white can push at d13) and h4 is strangely placed. Most importantly though white got sente so can play the huge p3 or q4. So yes I like white here.]
(;W[oq]
;B[ro]
;W[pp]
;B[qq]
;W[qn]
;B[rn]
;W[ql]
;B[cg]
;W[ef]
;B[cj])
(;W[pp]
;B[qq]
;W[ro]
;B[nr]
;W[pk])
(;W[dg]C[Or here?]
(;B[pp]
;W[po]
;B[oo]C[But this corner seems more urgent.])
(;B[oq])))
(;B[]C[Can black tenuki?]
(;W[fq]
;B[eq]
;W[ep]
;B[fp]
;W[er]
;B[eo]
;W[eq]
;B[gq]C[White dies first])
(;W[fp])
(;W[eq]))
(;B[go]C[But maybe some looser move is also possible, trying to be more efficient and useful on the outside, but leaves some bad aji and white may be able to get a semedori (force black to spend move inside territory to capture dead stones).]
;W[eq]
;B[aq]
;W[am]
;B[fr])
(;B[gn]))
(;B[aq]
(;W[er]
;B[bs]
;W[ds]C[we also need to consider if black might switch back and say the b1 for c1 exchange is good for him.])
(;W[bs]C[can white get this sente?]
(;B[ar]
;W[er]C[b1 for a2 a territorial gain for white, but d2 not captured means some liberty problems for the 3 stones.])
(;B[er]
(;W[ar]
;B[bo]C[White 3 stones can't escape, and corner not alive yet]
(;W[bp]C[Taking liberty means white c8 is sente and can connect under in gote, but if black does take a liberty from d4 stones then a5 becomes his sente.])
(;W[ap]))
(;W[bp]
;B[bo]
;W[ar])))))
(;W[bo]C[If white can't sacrifice the extra stone...]
;B[cs]
;W[bp]
;B[cr]
;W[bl]
;B[aq]
;W[am]C[Black has more liberties and an eye at b1 so not as bad shape as the dango before, so maybe tenuki or more efficient answer is more possible now.]))
(;B[bp]C[If black declines the tombstone sacrifice I expected this atari to capture white, because covering the b5 cut is valuable and it also reduces the size of white's (non-lossy) ko threats here: b1 threatens to just take 1 stone so some of the group remains not dead, versus 3 stones in atari so whole group dying with game atari.]
;W[ar]C[But the reason is probably this move to create a ko]
(;B[bs]
;W[ap]
;B[aq]
;W[pp]
;B[qq]
;W[bq]
;B[bo]
(;W[cs])
(;W[er]))
(;B[aq]
;W[bs]
;B[as]
;W[pp]
;B[qq]
;W[ar]
;B[bq]C[Black can avoid ko but loses territory and eyeshape]
;W[er]C[if white gets this later and black doesn't answer in corner then problems... So this is very nice for white to get in semi-sente to build eyshape, especially if black can't atari at c1.]
(;B[fp]
;W[ao]
(;B[as]
;W[ro]C[assume white has ko threats]
;B[nr]
;W[ar]
;B[an]
;W[bl]C[corner died])
(;B[bl]
;W[bo]))
(;B[bl]
;W[ck]
;B[fp]
;W[ao]
;B[an]
;W[bk]
;B[al]))))
(;W[er]C[Another idea would be here]
(;B[cs]C[If black ataris then e2 helps]
(;W[fo]C[e.g. can get out here])
(;W[go]
;B[fp]C[Maybe not g5 though]))
(;B[fp]C[But would black c1 or here instead?]
;W[bs]C[It's nice to get this in sente to squeeze the corner, but white is in trouble on the outside too.]
;B[ar]
(;W[bl]C[Locally speaking this is sente, but...]
;B[fr]C[Does white die first?])
(;W[gq]
;B[hq]
;W[fo]C[Is there some tesuji?]
;B[eo]
;W[ep]
;B[fq]C[Hmm]
;W[fr]
(;B[gr]
;W[go]
;B[gp]
;W[fm])
(;B[gp]
;W[gr]
;B[eq]
;W[dr]
;B[hr]
;W[fn]
;B[em]C[Probably white dead])))))
(;W[oq]C[Locally speaking this should be good for white]
;B[ro]
;W[pp]
;B[qq]
;W[qn]
;B[rn]
(;W[ql]C[Black is small, but gets sente]
(;B[en]C[I could believe this is good for black.])
(;B[cl]))
(;W[cl]C[If white tenukis then r7 is very painful later and did white really gain much by blocking at p3?])))
(;B[oq]C[This is the normal move]
;W[np]
;B[mq]C[But black is low and gote: I think it considers those exchanges good for white.]
;W[cl]C[Probably still here])
(;B[cl]C[What if black here]
(;W[oq]C[Would white trade?]
;B[en]
;W[cr]
;B[br]
;W[gp]C[Something like this? I don't know if this is good for white or black.])
(;W[dm]C[or save the stones?]))
(;B[en]C[Or would it be here?]
;W[oq]
;B[fp]C[And now black would rather have a move here than c8 to pretty much finish off the 3 stones, and white can save the 2 stones but they aren't so improtant if the 3 are dead. Again hard to judge this trade.]))
(;W[cl]
;B[en]C[Why not continue the fight now?]))
(;B[cm]
;W[dn]
;B[cr]C[We also see this sequence, but with h4 already pincering white it seems slack.]))
(;W[cm]
;B[bm]
;W[cl]C[White could answer here to be strong, but the wall is inefficient]
(;B[ql]C[Would black play on the right side such as around here?])
(;B[kp]C[or maybe help h4? My feeling though is AlphaGo tends to allow white to invade the lower side because as h4 runs out there are some chances to attack the white wall a bit too (e.g. e6 peep).])))
(;W[cm]C[What would happen if answer]
;B[bm]
;W[cl]C[Now tenuki where?]
(;B[jp]C[My feeling is lower side biggest, but who knows what AlphaGo would do?])
(;B[gf])
(;B[pk])))
(;B[bn]C[This was a common answer]
;W[qo]))
(;W[co]C[The standard variation]
;B[bo]
(;W[cn]
;B[bn]
;W[cm]
(;B[dr]C[The usual human joseki is to continue here, we think of it as double sente. ]
;W[er]
;B[cr]
;W[fq]C[This makes white's wall strong and have eyeshape in exchange for some more secure territory and eyes in the corner for black.])
(;B[jp]C[Omitting the d2 etc exchange is AlphaGo's new strategy which we saw a few times in the Master games (but it doesn't much play this in the self-play). The idea (as much as we can anthropomorphise a computer program) seems to be black is already alive (though white can bully a bit) so it's better not to give the wall eyeshape and you can even attack/pressure it later.
Maybe AlphaGo would play an extension around here.]
(;W[qo]
;B[eo]C[AlphaGo often peeps here later]
;W[do])
(;W[dr]C[One reason we humans didn't like to tenuki was if this move is sente for white]
(;B[cr]C[It's probably a good exchange for white (but probably still not as good for white as if he gets the e2 f3 shape).])
(;B[bm]C[But probably AlphaGo would push here once]
;W[cl]
(;B[cr]C[And now maybe come back.])
(;B[bl]C[Or even crawl again?]
;W[ck]
;B[pk]
;W[br]C[Black isn't dead, but it's groveling on the 2nd line which humans are loathe to do. But black did get a lot of tenukis, and if these can mitigate the influence of the wall maybe it's not so bad.])))))
(;W[bn]
;B[cn]
;W[do]
;B[bm]
;W[bp]
;B[an]
;W[bq]C[Humans play this double hane variation as does AlphaGo several times in these self-play games.])))
(;B[dg]C[After the Master games but before these self-play Park Yeounghun played this (right side a little different)]
;W[fc]
;B[cq]))
(;B[qd]C[Kim Seongjin 7d (black) vs Stanislaw Frejlak 6d at the Vienna tournament]
;W[dc]
;B[pp]
;W[dp]
;B[de]
;W[ce]
;B[cf]
;W[cd]
;B[cq]C[Kim played this 3-3 invasion. The left side is identical to the AlphaGo game, but black's corners on the right are different. This changes things a bit (e.g. top side relatively bigger as place to approach or make shimari from 3-4, white has a normal aproach on lower side to 4-4 so we could say blocking to make wall facing lower side makes more sense here) but probably the left side same shape is most important.]
;W[cp]C[Stanislaw blocked this side. I think this could be a mistake.]
;B[dq]
;W[fq]
;B[eq]
;W[ep]
;B[fr]
(;W[df]C[Natural follow up when black didn't defend cut, but could other choices be better given lower left?]
;B[cg]
;W[ef]
;B[cj]
(;W[gr]
;B[gq]
;W[fp]
;B[hr]
;W[bq]
;B[gs]
;W[cr]
;B[dr]
;W[ds]
;B[er]
(;W[cl]C[White wants to play here in terms of making a checking extension to pressure the black group above, but it's too close to the wall below and makes black's 3-3 invasion look good. But if you don't play here black can extend again and the white wall lacks a base. I feel white is behind now. ]
;B[lc]
;W[jc]
;B[je]
;W[qn]
;B[qk]
;W[qq]
;B[qp]
;W[pq]
;B[op]
;W[oq]
;B[np]
;W[nq]
;B[rq]
;W[rr]
;B[le]C[Black's lead seems clear now]
;W[qf]
;B[pd]
;W[qi]
;B[gd]
;W[ok]
;B[pj]
;W[rl]
;B[rk]
;W[ql]
;B[pi]
;W[ph]
;B[oh]
;W[pg]
;B[om]C[shape attack]
;W[on]
;B[nn]
;W[nm]
;B[mm]
;W[nl]
;B[pn]
;W[pm]
;B[oo]
;W[ol]
;B[ri]
;W[qh]
;B[ni]
;W[mn]
;B[mq]
;W[lm]
;B[re])
(;W[oc]
;B[cm]C[Black strengthens his group and weakens white (but maybe would answer p17 first?)]))
(;W[bq]C[This is another choice]
;B[br]
;W[gr]
;B[fp]
;W[gq]
;B[bp]
;W[fo]C[But white doesn't have this ladder.]))
(;W[cj]C[But maybe something like this could be better to extend from the wall, now black doesn't want to connect as heavy and no extension]
(;B[df]
(;W[fc])
(;W[ed]))
(;B[nc]C[So probably tenuki for now]))
(;W[dj])))