1
00:00:00,539 --> 00:00:03,539
foreign

2
00:00:08,340 --> 00:00:11,660
welcome back everybody

3
00:00:14,400 --> 00:00:19,560
our second speaker in this block this

4
00:00:16,680 --> 00:00:24,300
morning is the inevitable uh inimitable

5
00:00:19,560 --> 00:00:25,080
uh Chris neugebauer indomitable I'm not

6
00:00:24,300 --> 00:00:26,939
sure

7
00:00:25,080 --> 00:00:29,340
unavoidable yes

8
00:00:26,939 --> 00:00:32,700
um he is here today to speak to us about

9
00:00:29,340 --> 00:00:35,160
uh taking a swim in your kitchen sink

10
00:00:32,700 --> 00:00:36,300
while swearing no I'll leave it up to

11
00:00:35,160 --> 00:00:37,739
him to explain what's going on everyone

12
00:00:36,300 --> 00:00:39,300
please welcome Chris nog about at the

13
00:00:37,739 --> 00:00:40,739
stage hello

14
00:00:39,300 --> 00:00:42,660
all right

15
00:00:40,739 --> 00:00:44,520
uh I've never been described as

16
00:00:42,660 --> 00:00:47,520
inevitable before

17
00:00:44,520 --> 00:00:49,200
that is quite a long word given the the

18
00:00:47,520 --> 00:00:50,879
topic of today's talk today we're going

19
00:00:49,200 --> 00:00:52,559
to be talking about five letter words

20
00:00:50,879 --> 00:00:55,140
and I don't know how many letters are

21
00:00:52,559 --> 00:00:57,780
inevitable at the moment

22
00:00:55,140 --> 00:00:59,460
um yeah so uh before I begin uh Wordle

23
00:00:57,780 --> 00:01:01,620
is owned by the New York Times I'm not

24
00:00:59,460 --> 00:01:03,359
from the New York Times uh I will be

25
00:01:01,620 --> 00:01:05,299
using their trademark quite a bit during

26
00:01:03,359 --> 00:01:09,360
this talk it's uh

27
00:01:05,299 --> 00:01:12,320
unavoidable given the subject matter

28
00:01:09,360 --> 00:01:14,460
um so today we are going to talk about

29
00:01:12,320 --> 00:01:16,560
uh a few things we're going to talk

30
00:01:14,460 --> 00:01:19,200
about uh Wordle itself we're going to

31
00:01:16,560 --> 00:01:21,780
talk a bit about concurrency in Python

32
00:01:19,200 --> 00:01:23,580
and where things are at with that and

33
00:01:21,780 --> 00:01:27,600
also some things that are related to

34
00:01:23,580 --> 00:01:29,820
either Wordle and or concurrency

35
00:01:27,600 --> 00:01:31,619
um so I play wordal most days uh

36
00:01:29,820 --> 00:01:33,780
everyone else play Wordle

37
00:01:31,619 --> 00:01:34,979
wait who here is just here to watch me

38
00:01:33,780 --> 00:01:37,380
speak

39
00:01:34,979 --> 00:01:39,479
who's here for concurrency stuff

40
00:01:37,380 --> 00:01:41,159
okay right so all the people who are who

41
00:01:39,479 --> 00:01:43,439
are here for either of the two things in

42
00:01:41,159 --> 00:01:45,659
the middle will be slightly disappointed

43
00:01:43,439 --> 00:01:47,040
um so mediocre talk about Wordle a Medio

44
00:01:45,659 --> 00:01:48,540
could talk about concurrency hopefully

45
00:01:47,040 --> 00:01:50,939
two things together will be interesting

46
00:01:48,540 --> 00:01:53,759
I play wordal most days I play in hard

47
00:01:50,939 --> 00:01:55,860
mode I'm comfortably average at it

48
00:01:53,759 --> 00:01:58,140
um I usually take three or four guesses

49
00:01:55,860 --> 00:01:59,640
uh here's a picture from me playing on

50
00:01:58,140 --> 00:02:00,960
Wednesday where I managed to get it in

51
00:01:59,640 --> 00:02:03,420
two so I'm just making myself look

52
00:02:00,960 --> 00:02:06,360
better than I am my starting word these

53
00:02:03,420 --> 00:02:08,819
days is tears which I know for a fact is

54
00:02:06,360 --> 00:02:10,500
not a possible answer in Wordle as

55
00:02:08,819 --> 00:02:12,120
you'll see later it's what my solver

56
00:02:10,500 --> 00:02:14,760
thinks is the best possible starting

57
00:02:12,120 --> 00:02:16,379
word and who am I to argue with code

58
00:02:14,760 --> 00:02:21,000
that I've written

59
00:02:16,379 --> 00:02:22,440
uh my old starting word was irate uh if

60
00:02:21,000 --> 00:02:24,900
you play Wordle you'll know that that

61
00:02:22,440 --> 00:02:26,879
was an answer within the last month uh

62
00:02:24,900 --> 00:02:30,739
which is really disappointing because I

63
00:02:26,879 --> 00:02:30,739
started using the word Tears by then

64
00:02:31,440 --> 00:02:35,819
um anyway there's no actual spoilers in

65
00:02:34,260 --> 00:02:37,980
this talk I wrote most of the slides

66
00:02:35,819 --> 00:02:39,780
earlier this week I might have looked at

67
00:02:37,980 --> 00:02:41,340
wordlebot around then I wrote all the

68
00:02:39,780 --> 00:02:42,599
slides before I knew what today's word

69
00:02:41,340 --> 00:02:44,640
was and I've already forgotten what

70
00:02:42,599 --> 00:02:46,140
today's word was and also I think it's

71
00:02:44,640 --> 00:02:48,480
important to mention this talk probably

72
00:02:46,140 --> 00:02:52,140
won't change how you approach Wordle as

73
00:02:48,480 --> 00:02:54,060
a human being which I believe most of

74
00:02:52,140 --> 00:02:56,459
you are Beno

75
00:02:54,060 --> 00:02:58,560
um yeah exactly

76
00:02:56,459 --> 00:03:01,500
um I'd hate to spoil your enjoyment of a

77
00:02:58,560 --> 00:03:03,420
thing that's generally quite a pleasant

78
00:03:01,500 --> 00:03:04,080
thing that some of us

79
00:03:03,420 --> 00:03:06,420
um

80
00:03:04,080 --> 00:03:09,239
derive some level of enjoyment from

81
00:03:06,420 --> 00:03:12,060
so uh I don't think Wordle are solving

82
00:03:09,239 --> 00:03:13,560
world as a human uh resembles solving

83
00:03:12,060 --> 00:03:16,500
word or like a computer would because

84
00:03:13,560 --> 00:03:19,040
we're really not good as human beings at

85
00:03:16,500 --> 00:03:21,659
reasoning about a large amount of data

86
00:03:19,040 --> 00:03:24,120
in our heads and what do I mean by a

87
00:03:21,659 --> 00:03:26,819
large amount of data well according to

88
00:03:24,120 --> 00:03:30,120
the uh yet another word list word list

89
00:03:26,819 --> 00:03:32,480
that uh that a number of word games use

90
00:03:30,120 --> 00:03:35,640
there's

91
00:03:32,480 --> 00:03:36,659
12036 ish English language five letter

92
00:03:35,640 --> 00:03:38,940
words

93
00:03:36,659 --> 00:03:41,519
the solver that I use uses this word

94
00:03:38,940 --> 00:03:45,060
list and even if you filter out obvious

95
00:03:41,519 --> 00:03:47,220
plurals which uh Wordle itself does for

96
00:03:45,060 --> 00:03:48,180
its answer list there's still in excess

97
00:03:47,220 --> 00:03:50,879
of

98
00:03:48,180 --> 00:03:52,500
8 000-ish words it's really hard to keep

99
00:03:50,879 --> 00:03:54,840
that many words in your brain and reason

100
00:03:52,500 --> 00:03:56,879
about that sort of problem space now

101
00:03:54,840 --> 00:03:59,459
even after you make a good first guess

102
00:03:56,879 --> 00:04:00,900
which most people will have a good first

103
00:03:59,459 --> 00:04:04,260
guess there's there's normally somewhere

104
00:04:00,900 --> 00:04:05,760
between 60 and 500 Words left in the

105
00:04:04,260 --> 00:04:07,799
problem space to make your last five

106
00:04:05,760 --> 00:04:09,900
guesses with and

107
00:04:07,799 --> 00:04:12,540
even with that in mind it's really

108
00:04:09,900 --> 00:04:14,280
difficult to memorize which 60 to 500

109
00:04:12,540 --> 00:04:16,560
words are going to be left in in the

110
00:04:14,280 --> 00:04:18,239
list because every word will split that

111
00:04:16,560 --> 00:04:20,340
that set of English language words

112
00:04:18,239 --> 00:04:23,880
differently and so the way that most

113
00:04:20,340 --> 00:04:25,860
people go about solving Wordle is to uh

114
00:04:23,880 --> 00:04:27,900
basically just just do what everyone

115
00:04:25,860 --> 00:04:28,740
else does

116
00:04:27,900 --> 00:04:31,500
um

117
00:04:28,740 --> 00:04:34,320
and just to give you an idea of how well

118
00:04:31,500 --> 00:04:37,440
we can see how everyone else plays at

119
00:04:34,320 --> 00:04:39,720
Wordle uh the New York Times uh sometime

120
00:04:37,440 --> 00:04:41,340
after they bought out the the game they

121
00:04:39,720 --> 00:04:44,639
started collecting statistics because

122
00:04:41,340 --> 00:04:46,860
they're a tech company and they

123
00:04:44,639 --> 00:04:49,500
collected statistics and they made a

124
00:04:46,860 --> 00:04:52,500
site called wordlebot which is both a

125
00:04:49,500 --> 00:04:54,060
Wordle solver itself but also a site

126
00:04:52,500 --> 00:04:55,620
that will give you player statistics for

127
00:04:54,060 --> 00:04:57,060
playing Wordle and if you check in on

128
00:04:55,620 --> 00:04:59,880
wordlebot you get some interesting

129
00:04:57,060 --> 00:05:01,440
information particularly uh if you're an

130
00:04:59,880 --> 00:05:03,120
Australian person or somewhere on this

131
00:05:01,440 --> 00:05:05,820
side of the International Date Line you

132
00:05:03,120 --> 00:05:08,220
can join in on worldbot basically the

133
00:05:05,820 --> 00:05:09,840
moment people have started playing this

134
00:05:08,220 --> 00:05:12,660
is a screen grab from wordlebot on

135
00:05:09,840 --> 00:05:14,940
Wednesday morning uh here in Australia I

136
00:05:12,660 --> 00:05:17,460
took it at six a.m a dedication to

137
00:05:14,940 --> 00:05:20,100
Preparing slides and so really you've

138
00:05:17,460 --> 00:05:22,440
only got the early Rises and people from

139
00:05:20,100 --> 00:05:25,080
New Zealand here included in the stats

140
00:05:22,440 --> 00:05:27,360
and you can see that most people who

141
00:05:25,080 --> 00:05:28,880
play World earlier in that that 48 hour

142
00:05:27,360 --> 00:05:31,979
time window

143
00:05:28,880 --> 00:05:33,720
they they pick uh edu as their first

144
00:05:31,979 --> 00:05:35,699
word

145
00:05:33,720 --> 00:05:38,460
um why do people pick a year as their

146
00:05:35,699 --> 00:05:40,380
first word uh it's because it's got a

147
00:05:38,460 --> 00:05:41,940
lot of vowels in it and people think

148
00:05:40,380 --> 00:05:44,400
that figuring out which vowels will help

149
00:05:41,940 --> 00:05:46,320
them solve word or quickly and this is

150
00:05:44,400 --> 00:05:48,960
sort of a strategy that people latched

151
00:05:46,320 --> 00:05:51,060
on to very very early uh in the in the

152
00:05:48,960 --> 00:05:54,360
days of Wordle

153
00:05:51,060 --> 00:05:56,940
if you take a screen grab at 6am my time

154
00:05:54,360 --> 00:05:59,639
I live in California at the moment so I

155
00:05:56,940 --> 00:06:01,979
play wordal 17 hours later in California

156
00:05:59,639 --> 00:06:04,080
than I would in Australia and you can

157
00:06:01,979 --> 00:06:08,160
see that the most popular word is least

158
00:06:04,080 --> 00:06:08,880
a jerk is still up there it's number two

159
00:06:08,160 --> 00:06:10,320
um

160
00:06:08,880 --> 00:06:14,479
but what do you think has changed

161
00:06:10,320 --> 00:06:14,479
between uh between those 17 hours

162
00:06:14,520 --> 00:06:19,919
okay you've got everybody who subscribes

163
00:06:17,940 --> 00:06:22,800
to New York Times which nobody does in

164
00:06:19,919 --> 00:06:24,060
Australia uh who's looked at wordlebot

165
00:06:22,800 --> 00:06:25,500
and know that least is a better word

166
00:06:24,060 --> 00:06:26,880
than endure because the New York Times

167
00:06:25,500 --> 00:06:28,740
says so and they use it as their

168
00:06:26,880 --> 00:06:30,419
starting word New York Times actually no

169
00:06:28,740 --> 00:06:31,800
longer says that they say slate is the

170
00:06:30,419 --> 00:06:33,419
best starting word and that's now number

171
00:06:31,800 --> 00:06:34,440
three

172
00:06:33,419 --> 00:06:35,940
um

173
00:06:34,440 --> 00:06:37,740
and so once you make your first guess

174
00:06:35,940 --> 00:06:39,479
generally it's really just down to

175
00:06:37,740 --> 00:06:41,340
remembering a few good second guesses

176
00:06:39,479 --> 00:06:42,600
remembering some patterns and then the

177
00:06:41,340 --> 00:06:44,699
further you get into the game it's

178
00:06:42,600 --> 00:06:47,220
really just down to how much you can

179
00:06:44,699 --> 00:06:48,360
make educated guesses about

180
00:06:47,220 --> 00:06:51,180
um

181
00:06:48,360 --> 00:06:53,039
about what the answer is so yeah we we

182
00:06:51,180 --> 00:06:55,080
really don't have the mental capacity to

183
00:06:53,039 --> 00:06:57,060
play word or like a computer at all uh

184
00:06:55,080 --> 00:06:58,740
Wordle can play a lot closer to

185
00:06:57,060 --> 00:07:00,600
optimally

186
00:06:58,740 --> 00:07:03,240
um Wordle is a guessing game and and the

187
00:07:00,600 --> 00:07:05,460
way to play a guessing game is to make a

188
00:07:03,240 --> 00:07:07,979
guess of uh I guess that we think will

189
00:07:05,460 --> 00:07:09,419
reveal the most information about a

190
00:07:07,979 --> 00:07:12,539
solution space

191
00:07:09,419 --> 00:07:16,199
to a problem and thankfully information

192
00:07:12,539 --> 00:07:18,000
is a Lo uh is a formal term and we can

193
00:07:16,199 --> 00:07:21,120
use that formal term to figure out how

194
00:07:18,000 --> 00:07:22,440
to build a solver for for Wordle but I'm

195
00:07:21,120 --> 00:07:24,300
not going to build a solver for word or

196
00:07:22,440 --> 00:07:26,400
firstly I'm going to discuss a simpler

197
00:07:24,300 --> 00:07:27,780
game to play which is the guessing game

198
00:07:26,400 --> 00:07:30,360
where you have to guess a number between

199
00:07:27,780 --> 00:07:31,919
one and a thousand and the game will

200
00:07:30,360 --> 00:07:35,099
tell you if the answer is higher or

201
00:07:31,919 --> 00:07:37,440
lower so what's our first guess

202
00:07:35,099 --> 00:07:39,900
right you guess the number in the middle

203
00:07:37,440 --> 00:07:42,599
and then immediately the solution space

204
00:07:39,900 --> 00:07:43,979
gets halved and then you can guess the

205
00:07:42,599 --> 00:07:46,220
middle of that space and you'll have the

206
00:07:43,979 --> 00:07:48,900
solution every single time

207
00:07:46,220 --> 00:07:51,300
and with this game you can play

208
00:07:48,900 --> 00:07:53,160
optimally and you know like unless you

209
00:07:51,300 --> 00:07:55,139
accidentally happen on one of those

210
00:07:53,160 --> 00:07:56,699
answers you will know exactly how many

211
00:07:55,139 --> 00:07:59,240
guesses it will take you to get to the

212
00:07:56,699 --> 00:07:59,240
correct answer

213
00:08:00,000 --> 00:08:04,560
so so back to this idea of information

214
00:08:02,340 --> 00:08:07,440
every time you halve the solution space

215
00:08:04,560 --> 00:08:10,620
the the base two log of the size of the

216
00:08:07,440 --> 00:08:12,960
solution space reduces by about one

217
00:08:10,620 --> 00:08:16,080
um if there were 1024 spaces it would

218
00:08:12,960 --> 00:08:17,580
Reduce by one uh every single time so we

219
00:08:16,080 --> 00:08:20,879
described this halving of the solution

220
00:08:17,580 --> 00:08:23,819
space as gaining one bit of information

221
00:08:20,879 --> 00:08:25,500
so the idea of expected Information Gain

222
00:08:23,819 --> 00:08:29,039
Is that the base two log of the solution

223
00:08:25,500 --> 00:08:30,840
space less the expected base two log of

224
00:08:29,039 --> 00:08:33,479
the solution space after guess is your

225
00:08:30,840 --> 00:08:35,940
expected Information Gain so for our

226
00:08:33,479 --> 00:08:39,240
guessing game that expected Information

227
00:08:35,940 --> 00:08:40,560
Gain is always going to be basically one

228
00:08:39,240 --> 00:08:42,959
bit

229
00:08:40,560 --> 00:08:45,060
so what about Wordle

230
00:08:42,959 --> 00:08:46,560
we can we can sort of apply these

231
00:08:45,060 --> 00:08:48,360
theories to Wordle but the difference is

232
00:08:46,560 --> 00:08:50,399
with Wordle you can't actually play

233
00:08:48,360 --> 00:08:52,920
optimal you can't be guaranteed to play

234
00:08:50,399 --> 00:08:54,740
optimally uh the same guess is going to

235
00:08:52,920 --> 00:08:57,300
reveal different amounts of information

236
00:08:54,740 --> 00:08:58,740
depending on what the actual answer is

237
00:08:57,300 --> 00:09:01,440
so the best thing that you can do is

238
00:08:58,740 --> 00:09:03,120
make an estimate and so we can also

239
00:09:01,440 --> 00:09:06,360
figure out how good our estimate needs

240
00:09:03,120 --> 00:09:07,740
to be we can see that the solution space

241
00:09:06,360 --> 00:09:09,540
for

242
00:09:07,740 --> 00:09:12,120
um for the set of five letter words in

243
00:09:09,540 --> 00:09:14,580
English language is about 14 bits and

244
00:09:12,120 --> 00:09:17,399
that means that we need to reveal about

245
00:09:14,580 --> 00:09:18,600
2.33 bits of information per guess which

246
00:09:17,399 --> 00:09:21,240
means

247
00:09:18,600 --> 00:09:23,580
um somewhere between splitting the

248
00:09:21,240 --> 00:09:24,720
solution space into a quarter to an

249
00:09:23,580 --> 00:09:26,760
eighth somewhere in that sort of space

250
00:09:24,720 --> 00:09:29,700
you need to do each guess in order to

251
00:09:26,760 --> 00:09:32,519
get in guess the solution in six guesses

252
00:09:29,700 --> 00:09:34,440
or less so back to my starting starting

253
00:09:32,519 --> 00:09:38,399
word here is a worked example from a few

254
00:09:34,440 --> 00:09:42,000
weeks ago uh I took my my solving codes

255
00:09:38,399 --> 00:09:44,640
uh preferred first word and I got a

256
00:09:42,000 --> 00:09:48,720
green tea and a yellow a

257
00:09:44,640 --> 00:09:50,519
so if I guess the word today depending

258
00:09:48,720 --> 00:09:52,260
on the information revealed to me I

259
00:09:50,519 --> 00:09:54,600
might end up with three potential

260
00:09:52,260 --> 00:09:56,760
answers or I might end up with only one

261
00:09:54,600 --> 00:09:59,339
potential solution remaining so that's

262
00:09:56,760 --> 00:10:02,160
13 potential answers in that solution

263
00:09:59,339 --> 00:10:04,500
space it gets split up into seven groups

264
00:10:02,160 --> 00:10:06,060
and so the expected number of solutions

265
00:10:04,500 --> 00:10:09,420
I have after the guess is going to be

266
00:10:06,060 --> 00:10:10,860
1.86 words and that's a gain of 2.81

267
00:10:09,420 --> 00:10:13,019
bits of information which is better than

268
00:10:10,860 --> 00:10:15,180
the 2.33 we need

269
00:10:13,019 --> 00:10:17,700
on the other hand if I had guessed the

270
00:10:15,180 --> 00:10:19,740
word Titan I'll still end up with either

271
00:10:17,700 --> 00:10:21,600
three words or one word remaining but

272
00:10:19,740 --> 00:10:24,480
there's far more ways for me to end up

273
00:10:21,600 --> 00:10:26,279
with just one word uh than three and so

274
00:10:24,480 --> 00:10:28,380
that's 13 words split up into nine

275
00:10:26,279 --> 00:10:31,080
groups which means I have an expected

276
00:10:28,380 --> 00:10:34,440
number of solutions of 1.44 which is a

277
00:10:31,080 --> 00:10:36,540
gain of 3.17 bits of information

278
00:10:34,440 --> 00:10:39,540
so that gives us a strategy for playing

279
00:10:36,540 --> 00:10:41,459
Wordle the idea is to pick the word with

280
00:10:39,540 --> 00:10:43,860
the smallest expected solution space

281
00:10:41,459 --> 00:10:45,420
size uh make a guess with that word and

282
00:10:43,860 --> 00:10:47,100
then eliminate all those words from a

283
00:10:45,420 --> 00:10:48,600
set of answers and repeat that and

284
00:10:47,100 --> 00:10:50,700
repeat that and repeat that until only

285
00:10:48,600 --> 00:10:53,160
one word remains

286
00:10:50,700 --> 00:10:55,380
so let's look at some python code this

287
00:10:53,160 --> 00:10:57,300
is a python conference after all here's

288
00:10:55,380 --> 00:11:00,660
what the main Loop of my Wordle solver

289
00:10:57,300 --> 00:11:01,920
looks like we make a guess and then we

290
00:11:00,660 --> 00:11:03,899
classify the guess that we've made

291
00:11:01,920 --> 00:11:05,700
against the actual Target word and then

292
00:11:03,899 --> 00:11:08,399
we filter the solution space against the

293
00:11:05,700 --> 00:11:10,560
classification that we get just as an

294
00:11:08,399 --> 00:11:13,320
aside I structure a lot of my code like

295
00:11:10,560 --> 00:11:16,260
this with a high level with a high level

296
00:11:13,320 --> 00:11:18,839
top level function that can be split out

297
00:11:16,260 --> 00:11:21,420
into easily testable functions only one

298
00:11:18,839 --> 00:11:23,220
of these things is going to be subjected

299
00:11:21,420 --> 00:11:26,040
to concurrency later on which is uh

300
00:11:23,220 --> 00:11:28,140
which is Handy so let's take a look at

301
00:11:26,040 --> 00:11:30,060
the second of those two steps uh this is

302
00:11:28,140 --> 00:11:32,640
a classification routine I'm using in a

303
00:11:30,060 --> 00:11:34,140
num to describe whether the letter is in

304
00:11:32,640 --> 00:11:35,940
the solution or not and numbers are

305
00:11:34,140 --> 00:11:37,920
great you should use them

306
00:11:35,940 --> 00:11:40,200
um it says whether the letter is yellow

307
00:11:37,920 --> 00:11:42,120
or green and then it just returns a five

308
00:11:40,200 --> 00:11:44,339
Tuple of these classification objects

309
00:11:42,120 --> 00:11:46,320
which can then be used as keys in a

310
00:11:44,339 --> 00:11:49,440
dictionary you'll see why being used

311
00:11:46,320 --> 00:11:51,420
these used whether why these should be

312
00:11:49,440 --> 00:11:54,120
used as keys in a dictionary uh that's

313
00:11:51,420 --> 00:11:55,620
important kind of later on which is is

314
00:11:54,120 --> 00:11:57,060
it here no it's not uh here's the

315
00:11:55,620 --> 00:11:59,820
implementation of the filtering function

316
00:11:57,060 --> 00:12:01,920
it takes those classification uh five

317
00:11:59,820 --> 00:12:03,480
tubules and it runs them against every

318
00:12:01,920 --> 00:12:05,760
word in the solution space and it

319
00:12:03,480 --> 00:12:08,880
returns only the words in the solution

320
00:12:05,760 --> 00:12:11,940
space that match the classification

321
00:12:08,880 --> 00:12:14,700
I've written this solver three times and

322
00:12:11,940 --> 00:12:16,860
and every single time I've written a a

323
00:12:14,700 --> 00:12:18,240
four else block here

324
00:12:16,860 --> 00:12:20,459
um

325
00:12:18,240 --> 00:12:22,079
with like I've had the opportunity to

326
00:12:20,459 --> 00:12:23,880
decide not to do that and it's always

327
00:12:22,079 --> 00:12:25,019
been the right idea as far as the code

328
00:12:23,880 --> 00:12:29,240
I've written

329
00:12:25,019 --> 00:12:29,240
um sometimes it's justifiable

330
00:12:31,320 --> 00:12:34,680
so here's the point where we actually

331
00:12:32,940 --> 00:12:36,480
make a guess

332
00:12:34,680 --> 00:12:38,760
so the first function that we're going

333
00:12:36,480 --> 00:12:40,440
to try to do to make guesses is just to

334
00:12:38,760 --> 00:12:43,680
Loop through the space of possible

335
00:12:40,440 --> 00:12:45,839
guesses find the potential gas with the

336
00:12:43,680 --> 00:12:46,920
lowest potential score which is the the

337
00:12:45,839 --> 00:12:49,139
number of

338
00:12:46,920 --> 00:12:50,820
um the number of groups of words also

339
00:12:49,139 --> 00:12:53,700
the average number of answers in each

340
00:12:50,820 --> 00:12:55,380
group there's no tie break facility here

341
00:12:53,700 --> 00:12:57,120
which means it might not always make an

342
00:12:55,380 --> 00:12:59,160
ideal guess you'll you'll see the effect

343
00:12:57,120 --> 00:13:00,779
of that later on

344
00:12:59,160 --> 00:13:02,339
and then the scoring function there's

345
00:13:00,779 --> 00:13:04,560
there's another for Loop in that we go

346
00:13:02,339 --> 00:13:06,779
through every potential answer classify

347
00:13:04,560 --> 00:13:09,000
our guess as if the potential answer is

348
00:13:06,779 --> 00:13:11,459
the actual answer and then see how that

349
00:13:09,000 --> 00:13:13,380
guess would split the solution space so

350
00:13:11,459 --> 00:13:15,180
the more groups the smaller the solution

351
00:13:13,380 --> 00:13:16,860
space is likely to be after the guess

352
00:13:15,180 --> 00:13:18,660
and we just average that for every

353
00:13:16,860 --> 00:13:21,240
potential answer

354
00:13:18,660 --> 00:13:22,860
so uh I guess the first question is does

355
00:13:21,240 --> 00:13:25,019
this approach actually work well does

356
00:13:22,860 --> 00:13:28,079
this solve the problem that we have so

357
00:13:25,019 --> 00:13:30,300
I'm using the target word snake and you

358
00:13:28,079 --> 00:13:34,079
can see that it solves the word or in

359
00:13:30,300 --> 00:13:35,880
Easy Mode in five guesses a lot of these

360
00:13:34,079 --> 00:13:37,019
words are really really obscure and

361
00:13:35,880 --> 00:13:40,260
probably not ones that I would guess

362
00:13:37,019 --> 00:13:42,480
personally uh spelled means to Splinter

363
00:13:40,260 --> 00:13:44,700
or chip

364
00:13:42,480 --> 00:13:46,560
uh before I move on there's one code

365
00:13:44,700 --> 00:13:48,899
simplification that I suggest using this

366
00:13:46,560 --> 00:13:50,579
is what my actual code does uh here's

367
00:13:48,899 --> 00:13:52,860
our make guess function it's a lot of

368
00:13:50,579 --> 00:13:54,839
python code it can be Rewritten like

369
00:13:52,860 --> 00:13:57,300
this which has the nice effect of

370
00:13:54,839 --> 00:13:58,980
running the outer loop in C code it

371
00:13:57,300 --> 00:14:01,620
makes things somewhat but not

372
00:13:58,980 --> 00:14:03,779
importantly faster so there we go we can

373
00:14:01,620 --> 00:14:06,660
solve Wordle in about three minutes on

374
00:14:03,779 --> 00:14:08,639
my M1 MacBook here so can we do better

375
00:14:06,660 --> 00:14:10,680
in Python

376
00:14:08,639 --> 00:14:12,779
so the first place you might start is

377
00:14:10,680 --> 00:14:14,399
looking at thread pools threads are

378
00:14:12,779 --> 00:14:17,459
chains of execution that exist within

379
00:14:14,399 --> 00:14:19,560
the same process segment on your system

380
00:14:17,459 --> 00:14:21,779
here is our make guess function and

381
00:14:19,560 --> 00:14:23,820
again we have this Loop and we do

382
00:14:21,779 --> 00:14:25,740
exactly the same thing every single time

383
00:14:23,820 --> 00:14:27,959
without modifying any data structures

384
00:14:25,740 --> 00:14:30,060
and then the inside of that Loop is also

385
00:14:27,959 --> 00:14:32,519
quite resource intensive you there's

386
00:14:30,060 --> 00:14:33,600
there's two nested for Loops

387
00:14:32,519 --> 00:14:35,339
um

388
00:14:33,600 --> 00:14:37,320
so there's that and so we can we can

389
00:14:35,339 --> 00:14:38,940
pull this out and this is where the

390
00:14:37,320 --> 00:14:41,579
concurrent Futures Library comes in in

391
00:14:38,940 --> 00:14:43,440
Python it is often my first Port of Call

392
00:14:41,579 --> 00:14:46,440
for any sort of asynchronous programming

393
00:14:43,440 --> 00:14:48,000
even before using async await

394
00:14:46,440 --> 00:14:49,440
um it's I find it somewhat easier to

395
00:14:48,000 --> 00:14:52,019
conceptually put things together that

396
00:14:49,440 --> 00:14:53,820
way your mileage may vary uh it's

397
00:14:52,019 --> 00:14:55,260
definitely much easier than writing raw

398
00:14:53,820 --> 00:14:57,779
threads in Python if you have a problem

399
00:14:55,260 --> 00:14:59,579
that's amenable to threading so you can

400
00:14:57,779 --> 00:15:01,500
think in terms of of calling functions

401
00:14:59,579 --> 00:15:03,420
and waiting for a result rather than

402
00:15:01,500 --> 00:15:05,579
working with things like locks which I

403
00:15:03,420 --> 00:15:07,740
think are a bit tedious and and prone to

404
00:15:05,579 --> 00:15:09,480
getting those things wrong uh so here's

405
00:15:07,740 --> 00:15:11,160
what our make guess function looks like

406
00:15:09,480 --> 00:15:14,160
with a thread pool

407
00:15:11,160 --> 00:15:16,320
we submit each potential guess to the

408
00:15:14,160 --> 00:15:18,899
thread pool which returns a future which

409
00:15:16,320 --> 00:15:20,459
is an object that will eventually hold

410
00:15:18,899 --> 00:15:22,440
the result of the guests that we made

411
00:15:20,459 --> 00:15:23,820
and then we're going to associate that

412
00:15:22,440 --> 00:15:25,260
future with the word that we use to

413
00:15:23,820 --> 00:15:27,120
submit it so we can actually get a score

414
00:15:25,260 --> 00:15:29,699
back and know which word it was that

415
00:15:27,120 --> 00:15:31,440
generated the best result we're still

416
00:15:29,699 --> 00:15:34,500
using that Min built in that I used in

417
00:15:31,440 --> 00:15:36,420
that that code Improvement we grab each

418
00:15:34,500 --> 00:15:38,880
result as it comes out of the thread

419
00:15:36,420 --> 00:15:40,800
pool complete and by you we do that by

420
00:15:38,880 --> 00:15:42,300
using the the as completed function

421
00:15:40,800 --> 00:15:43,579
which is in the concurrent Futures

422
00:15:42,300 --> 00:15:45,899
Library

423
00:15:43,579 --> 00:15:48,120
so we can set up a thread pool with

424
00:15:45,899 --> 00:15:49,740
eight workers I have only eight cores in

425
00:15:48,120 --> 00:15:51,260
my machine so we won't get more than an

426
00:15:49,740 --> 00:15:53,519
eight-time speed up

427
00:15:51,260 --> 00:15:55,800
and we can go from our looping code

428
00:15:53,519 --> 00:15:59,279
which takes 185 seconds to make its

429
00:15:55,800 --> 00:16:03,139
first guess oh uh Vemma by the way a

430
00:15:59,279 --> 00:16:03,139
westphalian vigilante Society

431
00:16:03,860 --> 00:16:09,199
and here's our threaded code which takes

432
00:16:06,240 --> 00:16:12,120
181 seconds to make a guess

433
00:16:09,199 --> 00:16:13,980
uh why

434
00:16:12,120 --> 00:16:15,180
the Gill yeah

435
00:16:13,980 --> 00:16:17,760
um so this is where I point out that

436
00:16:15,180 --> 00:16:20,100
python has a global interpreter lock uh

437
00:16:17,760 --> 00:16:22,680
the global interpreter lock is a it's a

438
00:16:20,100 --> 00:16:24,660
it's a lock it prevents multiple threads

439
00:16:22,680 --> 00:16:27,420
from having access to any piece of state

440
00:16:24,660 --> 00:16:30,360
that is held by the python interpreter

441
00:16:27,420 --> 00:16:33,959
um in practice this means that only one

442
00:16:30,360 --> 00:16:35,699
python thread can execute it once

443
00:16:33,959 --> 00:16:37,380
I think it's actually quite interesting

444
00:16:35,699 --> 00:16:40,019
to note here that the guild basically

445
00:16:37,380 --> 00:16:43,639
prevents concurrency but it's very very

446
00:16:40,019 --> 00:16:46,860
very efficient at preventing concurrency

447
00:16:43,639 --> 00:16:48,420
we don't actually get any slowdown from

448
00:16:46,860 --> 00:16:50,459
various threads having to acquire and

449
00:16:48,420 --> 00:16:51,959
release the guild here we it runs in

450
00:16:50,459 --> 00:16:54,120
basically exactly the same amount of

451
00:16:51,959 --> 00:16:55,620
time it's like five seconds difference

452
00:16:54,120 --> 00:16:58,100
that difference is down to me running

453
00:16:55,620 --> 00:17:00,180
Chrome on my computer at the same time

454
00:16:58,100 --> 00:17:02,040
don't do that if you're trying to do

455
00:17:00,180 --> 00:17:04,199
scientific analysis of your your speed

456
00:17:02,040 --> 00:17:05,220
up that's a really bad idea

457
00:17:04,199 --> 00:17:07,860
um

458
00:17:05,220 --> 00:17:10,020
so are our threads useful in Python with

459
00:17:07,860 --> 00:17:11,579
with the girl uh yes if you have

460
00:17:10,020 --> 00:17:13,140
parallel code that's not written in

461
00:17:11,579 --> 00:17:16,140
Python and you can release the Gill on

462
00:17:13,140 --> 00:17:17,699
each thread if your parallel code is IO

463
00:17:16,140 --> 00:17:20,220
bound but only up to a certain amount

464
00:17:17,699 --> 00:17:22,260
only to a certain extent async is much

465
00:17:20,220 --> 00:17:24,839
better for that if you can run your code

466
00:17:22,260 --> 00:17:26,459
on Pi Pi and things will be better it's

467
00:17:24,839 --> 00:17:29,040
really not good for writing a wordal

468
00:17:26,459 --> 00:17:31,679
solver uh so the next thing we can try

469
00:17:29,040 --> 00:17:33,660
is a thing called process pools so

470
00:17:31,679 --> 00:17:35,820
processors are spaces of execution that

471
00:17:33,660 --> 00:17:37,140
are managed by your operating system and

472
00:17:35,820 --> 00:17:39,480
they have their own memory they don't

473
00:17:37,140 --> 00:17:41,580
share memory between uh between multiple

474
00:17:39,480 --> 00:17:42,960
processors uh in practice it's really

475
00:17:41,580 --> 00:17:44,940
only possible to communicate between

476
00:17:42,960 --> 00:17:47,400
processes over things like sockets or

477
00:17:44,940 --> 00:17:48,419
files or specially declared shared

478
00:17:47,400 --> 00:17:51,059
memory that your operating system

479
00:17:48,419 --> 00:17:53,940
manages for you the good news about

480
00:17:51,059 --> 00:17:56,100
processes is that there's one Gill per

481
00:17:53,940 --> 00:17:59,280
python process so if you have multiple

482
00:17:56,100 --> 00:18:00,840
processes you only uh have you have one

483
00:17:59,280 --> 00:18:03,000
deal per process and you should actually

484
00:18:00,840 --> 00:18:05,820
see some speed up and if you're using

485
00:18:03,000 --> 00:18:07,620
the concurrent Futures Library

486
00:18:05,820 --> 00:18:10,919
the only real difference between using

487
00:18:07,620 --> 00:18:13,320
threads and processors is defining what

488
00:18:10,919 --> 00:18:14,940
sort of pool you use so I'm using a

489
00:18:13,320 --> 00:18:16,559
process pool executor instead of a

490
00:18:14,940 --> 00:18:18,480
thread pool executor

491
00:18:16,559 --> 00:18:23,460
and to remind you our single threaded

492
00:18:18,480 --> 00:18:26,940
code talk 185 seconds to make a guess

493
00:18:23,460 --> 00:18:29,460
with processpool executor it it took 38

494
00:18:26,940 --> 00:18:32,700
seconds which is actually a speed up

495
00:18:29,460 --> 00:18:34,620
which is cool it's about 4.8 times

496
00:18:32,700 --> 00:18:36,000
so what if we wanted to see if we can

497
00:18:34,620 --> 00:18:38,880
get more speed up well we can change

498
00:18:36,000 --> 00:18:41,940
this Max worker's parameter to eight uh

499
00:18:38,880 --> 00:18:45,539
from 8 to 16 and it goes from 38 seconds

500
00:18:41,940 --> 00:18:50,160
under eight workers to more than 38

501
00:18:45,539 --> 00:18:53,039
seconds this again is because I have

502
00:18:50,160 --> 00:18:55,080
I have only eight cores on my laptop so

503
00:18:53,039 --> 00:18:56,760
my laptop only has eight cores which is

504
00:18:55,080 --> 00:18:59,160
obviously the maximum speed up I can get

505
00:18:56,760 --> 00:19:01,320
16 processors can't run on eight cores

506
00:18:59,160 --> 00:19:03,240
at once but I only got 4.8 times speed

507
00:19:01,320 --> 00:19:04,980
up uh with uh with switching to a

508
00:19:03,240 --> 00:19:08,100
process pool

509
00:19:04,980 --> 00:19:09,419
um this is where I go into our uh into

510
00:19:08,100 --> 00:19:11,400
our undergraduate computer science

511
00:19:09,419 --> 00:19:13,260
curriculum and talk about amdel's law if

512
00:19:11,400 --> 00:19:14,820
you know about that you can you can

513
00:19:13,260 --> 00:19:16,440
check it off now but for everyone else

514
00:19:14,820 --> 00:19:18,059
the quick explanation is that a

515
00:19:16,440 --> 00:19:20,400
programming task can be split into Parts

516
00:19:18,059 --> 00:19:22,020
the Wordle solver has three you can

517
00:19:20,400 --> 00:19:23,340
score every word you need to do that you

518
00:19:22,020 --> 00:19:24,900
need to find the minimum by looking at

519
00:19:23,340 --> 00:19:27,600
every single word and then you need to

520
00:19:24,900 --> 00:19:29,280
filter the list of solutions the scoring

521
00:19:27,600 --> 00:19:31,559
task can be split up into chunks and you

522
00:19:29,280 --> 00:19:33,660
can run all those bits in parallel

523
00:19:31,559 --> 00:19:35,160
the bits that run in sequence cannot be

524
00:19:33,660 --> 00:19:36,960
run in parallel because they run in

525
00:19:35,160 --> 00:19:38,580
sequence and that establishes a limit on

526
00:19:36,960 --> 00:19:41,640
the amount our program can be sped up

527
00:19:38,580 --> 00:19:43,440
but also in reality there's a bunch of

528
00:19:41,640 --> 00:19:45,480
work the operating system needs to do in

529
00:19:43,440 --> 00:19:47,100
order to coordinate all these processes

530
00:19:45,480 --> 00:19:49,140
to deal with shared memory those sorts

531
00:19:47,100 --> 00:19:50,880
of things that gets added into our

532
00:19:49,140 --> 00:19:53,640
execution time and that coordination

533
00:19:50,880 --> 00:19:55,080
actually increases as we start to add

534
00:19:53,640 --> 00:19:57,120
more processes beyond the number of

535
00:19:55,080 --> 00:19:59,940
cores that we have on our system and

536
00:19:57,120 --> 00:20:01,980
that's where that extra slowdown came on

537
00:19:59,940 --> 00:20:03,660
so maybe there are other bits that we

538
00:20:01,980 --> 00:20:05,760
can run in parallel maybe we can run the

539
00:20:03,660 --> 00:20:07,740
filtering operation

540
00:20:05,760 --> 00:20:09,539
um it turns out that really just doesn't

541
00:20:07,740 --> 00:20:12,960
do much it actually in fact slightly

542
00:20:09,539 --> 00:20:14,880
slows down things because there's not a

543
00:20:12,960 --> 00:20:16,620
whole lot of well there's actually too

544
00:20:14,880 --> 00:20:18,960
much inter-process communication going

545
00:20:16,620 --> 00:20:20,940
on in that filter operation

546
00:20:18,960 --> 00:20:22,559
so to get a task from one process to

547
00:20:20,940 --> 00:20:25,440
another it actually has to be explicitly

548
00:20:22,559 --> 00:20:27,179
communicated and that takes time that's

549
00:20:25,440 --> 00:20:29,100
this sort of os coordination thing I

550
00:20:27,179 --> 00:20:31,440
showed in that diagram earlier uh

551
00:20:29,100 --> 00:20:33,299
processes they work best process pools

552
00:20:31,440 --> 00:20:35,280
work best if you can have longer running

553
00:20:33,299 --> 00:20:37,320
tasks that don't need to communicate

554
00:20:35,280 --> 00:20:39,480
very often and that gives us another

555
00:20:37,320 --> 00:20:40,559
insight into how to make things a bit

556
00:20:39,480 --> 00:20:42,780
faster

557
00:20:40,559 --> 00:20:44,880
so instead of submitting thousands of

558
00:20:42,780 --> 00:20:47,179
relatively small tasks into the process

559
00:20:44,880 --> 00:20:49,559
pool if we spend some time actually

560
00:20:47,179 --> 00:20:51,179
breaking up our problem space into as

561
00:20:49,559 --> 00:20:53,460
many cores that we have and running

562
00:20:51,179 --> 00:20:55,620
those sequentially we can solve the

563
00:20:53,460 --> 00:20:57,539
problem locally within each of the each

564
00:20:55,620 --> 00:21:00,080
chunk of the problem space and then find

565
00:20:57,539 --> 00:21:02,520
the global solution afterwards

566
00:21:00,080 --> 00:21:04,200
so we're actually just making use of the

567
00:21:02,520 --> 00:21:06,120
the make guess with Loop function that

568
00:21:04,200 --> 00:21:08,160
we wrote earlier and we're only doing 16

569
00:21:06,120 --> 00:21:10,440
pieces of inter-process communication

570
00:21:08,160 --> 00:21:12,480
instead of the 24 000 pieces of

571
00:21:10,440 --> 00:21:14,580
inter-process communication that we were

572
00:21:12,480 --> 00:21:17,100
doing with the obvious way of chunking

573
00:21:14,580 --> 00:21:19,380
things up and so here's our version

574
00:21:17,100 --> 00:21:20,760
without chunking up the inputs this is

575
00:21:19,380 --> 00:21:22,320
the one with lots of inter-process

576
00:21:20,760 --> 00:21:25,440
communication because we're submitting

577
00:21:22,320 --> 00:21:28,020
small tasks uh 24 000 small tasks into

578
00:21:25,440 --> 00:21:29,520
16 no 12 000 small tasks into the

579
00:21:28,020 --> 00:21:32,280
process pool

580
00:21:29,520 --> 00:21:34,320
um it takes that long uh here's the

581
00:21:32,280 --> 00:21:37,260
version where we're only doing eight

582
00:21:34,320 --> 00:21:38,940
submissions into the process pool you'll

583
00:21:37,260 --> 00:21:40,559
also notice that the opening word is

584
00:21:38,940 --> 00:21:43,020
different this is because our

585
00:21:40,559 --> 00:21:44,820
tie-breaking behavior is not really

586
00:21:43,020 --> 00:21:46,980
deterministic it really depends on what

587
00:21:44,820 --> 00:21:51,179
thing comes first out of the process

588
00:21:46,980 --> 00:21:53,820
pool so instead of Tears we have hairs

589
00:21:51,179 --> 00:21:59,120
um it is faster by a noticeable margin

590
00:21:53,820 --> 00:21:59,120
uh Sports by the way it means brittle

591
00:21:59,340 --> 00:22:05,640
so uh process pools they are good if you

592
00:22:02,520 --> 00:22:08,580
have a small number of large but CPU

593
00:22:05,640 --> 00:22:10,200
bounded tasks it takes a lot more work

594
00:22:08,580 --> 00:22:11,700
to set these up efficiently you need to

595
00:22:10,200 --> 00:22:13,980
think about how you split your problem

596
00:22:11,700 --> 00:22:15,720
space up such that you reduce the amount

597
00:22:13,980 --> 00:22:17,580
of inter-process communication you have

598
00:22:15,720 --> 00:22:19,140
this is different to how you would deal

599
00:22:17,580 --> 00:22:21,419
with threads in a language like C or

600
00:22:19,140 --> 00:22:24,600
Java and it's reasonably good for

601
00:22:21,419 --> 00:22:27,059
writing a word or solver I included the

602
00:22:24,600 --> 00:22:28,320
word async in the title of the talk

603
00:22:27,059 --> 00:22:30,120
um

604
00:22:28,320 --> 00:22:31,620
that's only because I need the third

605
00:22:30,120 --> 00:22:33,360
five letter word that was relevant to

606
00:22:31,620 --> 00:22:35,360
concurrency

607
00:22:33,360 --> 00:22:35,360
um

608
00:22:35,880 --> 00:22:39,419
it has it has the it has the same

609
00:22:37,860 --> 00:22:42,179
problem as thread pools in that

610
00:22:39,419 --> 00:22:43,980
everything runs within the same um uh

611
00:22:42,179 --> 00:22:46,260
within the same process space as python

612
00:22:43,980 --> 00:22:47,820
so you'd always end up with uh you'd

613
00:22:46,260 --> 00:22:48,960
always end up with something related to

614
00:22:47,820 --> 00:22:50,640
the Gill if you had a different event

615
00:22:48,960 --> 00:22:52,260
Loop that did allow for thread based

616
00:22:50,640 --> 00:22:54,059
concurrency if you dealt with process

617
00:22:52,260 --> 00:22:55,620
based concurrency well you're dealing

618
00:22:54,059 --> 00:22:57,120
with a whole lot of language level

619
00:22:55,620 --> 00:22:58,559
complication that would make things

620
00:22:57,120 --> 00:23:01,500
slower

621
00:22:58,559 --> 00:23:04,679
finally let's talk about no Gill because

622
00:23:01,500 --> 00:23:06,299
as I mentioned python has a gill it's

623
00:23:04,679 --> 00:23:07,919
always had a girl and people have always

624
00:23:06,299 --> 00:23:09,419
wanted the guild to go away because it

625
00:23:07,919 --> 00:23:12,179
gets in the way of using threads

626
00:23:09,419 --> 00:23:15,120
efficiently there is a new effort called

627
00:23:12,179 --> 00:23:17,640
nogill which is about the fifth effort

628
00:23:15,120 --> 00:23:19,559
that I've seen in the last 10 years and

629
00:23:17,640 --> 00:23:21,179
it seems to be making some level of

630
00:23:19,559 --> 00:23:21,960
progress

631
00:23:21,179 --> 00:23:23,880
um

632
00:23:21,960 --> 00:23:25,500
in terms of making like using

633
00:23:23,880 --> 00:23:27,600
concurrency in Python without the guild

634
00:23:25,500 --> 00:23:29,460
actually usable you can install no Gill

635
00:23:27,600 --> 00:23:32,520
very easily if you have pione on your

636
00:23:29,460 --> 00:23:34,460
machine just go power and install nogill

637
00:23:32,520 --> 00:23:37,320
3 9 10-1

638
00:23:34,460 --> 00:23:39,000
it is only supported in Python three

639
00:23:37,320 --> 00:23:39,840
nine

640
00:23:39,000 --> 00:23:42,860
um

641
00:23:39,840 --> 00:23:45,780
this is important because in Python 311

642
00:23:42,860 --> 00:23:48,840
uh the there's a team at Microsoft

643
00:23:45,780 --> 00:23:51,120
running the faster C python project and

644
00:23:48,840 --> 00:23:53,520
the faster C python project did a lot of

645
00:23:51,120 --> 00:23:55,140
work to make C python faster and they

646
00:23:53,520 --> 00:23:58,440
actually succeeded and all their work

647
00:23:55,140 --> 00:24:01,620
landed in Python 311. so our first guess

648
00:23:58,440 --> 00:24:03,960
in Python 39 under nogill in a single

649
00:24:01,620 --> 00:24:06,419
threaded uh

650
00:24:03,960 --> 00:24:09,059
um a single threaded sort of context uh

651
00:24:06,419 --> 00:24:11,880
took about 40 seconds longer to run than

652
00:24:09,059 --> 00:24:13,200
it did in Python 311 so our first useful

653
00:24:11,880 --> 00:24:14,580
lesson like if you need to make your

654
00:24:13,200 --> 00:24:16,620
code faster and you're still running

655
00:24:14,580 --> 00:24:20,280
python 310 or earlier

656
00:24:16,620 --> 00:24:22,200
upgrade to python 311 it will get faster

657
00:24:20,280 --> 00:24:22,860
for free basically

658
00:24:22,200 --> 00:24:25,440
um

659
00:24:22,860 --> 00:24:27,179
what takes even longer than running our

660
00:24:25,440 --> 00:24:29,520
single threaded version in nogill is

661
00:24:27,179 --> 00:24:30,179
running with a thread pull executor

662
00:24:29,520 --> 00:24:31,980
um

663
00:24:30,179 --> 00:24:33,480
which so instead of taking about the

664
00:24:31,980 --> 00:24:35,340
same amount of time as a single threaded

665
00:24:33,480 --> 00:24:37,700
version as it took in the Gill uh with

666
00:24:35,340 --> 00:24:42,539
the Gill it now takes about one to six

667
00:24:37,700 --> 00:24:44,520
1.6 times longer than it did in a single

668
00:24:42,539 --> 00:24:46,740
threaded execution as I was mentioning

669
00:24:44,520 --> 00:24:48,659
earlier the Gill is actually really

670
00:24:46,740 --> 00:24:50,700
really really efficient and this is why

671
00:24:48,659 --> 00:24:53,700
all these Gill removal projects have

672
00:24:50,700 --> 00:24:55,500
sort of faltered over the years

673
00:24:53,700 --> 00:24:57,659
um there's not really any slowdown from

674
00:24:55,500 --> 00:24:59,820
acquiring releasing the Gill in in

675
00:24:57,659 --> 00:25:01,980
normal python on the other hand once you

676
00:24:59,820 --> 00:25:04,500
start adding a lot more locks and things

677
00:25:01,980 --> 00:25:06,480
like finely grained thread safety into

678
00:25:04,500 --> 00:25:09,480
the the memory allocator in Python

679
00:25:06,480 --> 00:25:10,559
things get really really slow

680
00:25:09,480 --> 00:25:13,980
um

681
00:25:10,559 --> 00:25:15,840
1.6 times slower as it turns out you may

682
00:25:13,980 --> 00:25:17,600
be asking whether or not nogil is worth

683
00:25:15,840 --> 00:25:19,500
it from this one example I've given you

684
00:25:17,600 --> 00:25:22,740
first up

685
00:25:19,500 --> 00:25:25,020
yeah slowing down only 1.6 times is a

686
00:25:22,740 --> 00:25:27,179
huge Improvement on previous versions of

687
00:25:25,020 --> 00:25:29,100
trying to remove the Gill especially

688
00:25:27,179 --> 00:25:31,159
since the authors actually have an idea

689
00:25:29,100 --> 00:25:33,900
of how to make this get even faster

690
00:25:31,159 --> 00:25:36,720
secondly there are actually demos on on

691
00:25:33,900 --> 00:25:39,240
the nogill readme that demonstrates a

692
00:25:36,720 --> 00:25:41,039
CPU bound function that parallelizes

693
00:25:39,240 --> 00:25:42,179
basically with the number of cores you

694
00:25:41,039 --> 00:25:44,400
have in the number of threads you give

695
00:25:42,179 --> 00:25:46,679
it it's a really contrived example it

696
00:25:44,400 --> 00:25:48,480
doesn't touch any data structures but it

697
00:25:46,679 --> 00:25:51,840
demonstrates that you can run threaded

698
00:25:48,480 --> 00:25:54,000
python code in AC Python and get

699
00:25:51,840 --> 00:25:56,400
unbounded speed up which is which is

700
00:25:54,000 --> 00:25:58,200
very impressive there is a pep to remove

701
00:25:56,400 --> 00:25:59,760
the Gill There is a guarantee of

702
00:25:58,200 --> 00:26:02,400
sponsorship to do the real work to make

703
00:25:59,760 --> 00:26:04,080
it faster I would not be surprised if

704
00:26:02,400 --> 00:26:06,059
the Gill goes away in the next three to

705
00:26:04,080 --> 00:26:07,500
five years and that would be really

706
00:26:06,059 --> 00:26:10,140
incredible

707
00:26:07,500 --> 00:26:11,700
so to finish up we're finished talking

708
00:26:10,140 --> 00:26:13,020
about concurrency in Python let's get

709
00:26:11,700 --> 00:26:15,240
back to talking about Wordle which is

710
00:26:13,020 --> 00:26:16,620
what I'm actually interested in firstly

711
00:26:15,240 --> 00:26:19,140
all the examples I've shown you have

712
00:26:16,620 --> 00:26:20,640
been in Easy Mode I play in hard mode

713
00:26:19,140 --> 00:26:22,020
it's probably interesting to see what it

714
00:26:20,640 --> 00:26:24,179
would take to make the solver work in

715
00:26:22,020 --> 00:26:25,740
hard mode uh it's really just a matter

716
00:26:24,179 --> 00:26:28,500
of changing one line at the bottom of

717
00:26:25,740 --> 00:26:30,240
the main Loop very small so we're now

718
00:26:28,500 --> 00:26:32,940
instead of guessing with the entire set

719
00:26:30,240 --> 00:26:34,679
of English language words we're just

720
00:26:32,940 --> 00:26:37,679
solving with the set of answers that we

721
00:26:34,679 --> 00:26:39,720
get after each step and it very slightly

722
00:26:37,679 --> 00:26:41,820
changes the guesses it makes it

723
00:26:39,720 --> 00:26:43,500
significantly reduces the last the

724
00:26:41,820 --> 00:26:45,080
amount of time the subsequent guesses

725
00:26:43,500 --> 00:26:47,460
beyond the first

726
00:26:45,080 --> 00:26:49,380
takes that's because there's fewer words

727
00:26:47,460 --> 00:26:52,980
to choose from it takes less time to run

728
00:26:49,380 --> 00:26:55,159
uh cpal is the parts of the calyx of a

729
00:26:52,980 --> 00:26:55,159
flower

730
00:26:55,380 --> 00:27:01,020
and finally how good is this solver at

731
00:26:58,080 --> 00:27:03,179
actual Wordle before the New York Times

732
00:27:01,020 --> 00:27:05,340
took over the full set of answers to

733
00:27:03,179 --> 00:27:08,039
Wordle was visible in the JavaScript

734
00:27:05,340 --> 00:27:09,840
source of the game so you could you

735
00:27:08,039 --> 00:27:11,820
could basically strip that out and and

736
00:27:09,840 --> 00:27:13,140
use it as an answer list there are

737
00:27:11,820 --> 00:27:14,700
copies of this all over the Internet

738
00:27:13,140 --> 00:27:17,400
because it's the internet

739
00:27:14,700 --> 00:27:19,020
uh there's about 2 300 words in that

740
00:27:17,400 --> 00:27:21,360
answer list and we can run our solve

741
00:27:19,020 --> 00:27:22,919
against all of them firstly to establish

742
00:27:21,360 --> 00:27:25,260
a baseline let's see what the behavior

743
00:27:22,919 --> 00:27:28,380
is like if we just guess randomly from

744
00:27:25,260 --> 00:27:30,120
the set of all possible solutions it

745
00:27:28,380 --> 00:27:31,740
solves the game correctly in about 90

746
00:27:30,120 --> 00:27:33,480
percent of the time

747
00:27:31,740 --> 00:27:36,299
and there are three words that that

748
00:27:33,480 --> 00:27:37,799
trigger a bug I think this is just due

749
00:27:36,299 --> 00:27:38,940
to their not being an answer in the

750
00:27:37,799 --> 00:27:40,500
guest list and I haven't just done

751
00:27:38,940 --> 00:27:42,600
enough error correction in the code to

752
00:27:40,500 --> 00:27:45,720
account for it properly

753
00:27:42,600 --> 00:27:48,720
code first talk

754
00:27:45,720 --> 00:27:50,460
um in Easy Mode uh there's those three

755
00:27:48,720 --> 00:27:52,860
bug cases that we saw in the random

756
00:27:50,460 --> 00:27:55,320
guessing thing it successfully solves

757
00:27:52,860 --> 00:27:59,340
all but one puzzle in six guesses or

758
00:27:55,320 --> 00:28:02,039
less and it does well in more than uh it

759
00:27:59,340 --> 00:28:05,400
does well more than half of the uh of

760
00:28:02,039 --> 00:28:07,320
the words in four guesses so

761
00:28:05,400 --> 00:28:09,600
I basically play about as well as my

762
00:28:07,320 --> 00:28:11,640
solver does personally which is cool uh

763
00:28:09,600 --> 00:28:15,299
the one word that it can't guess is

764
00:28:11,640 --> 00:28:16,880
Joker it takes seven guesses uh via a

765
00:28:15,299 --> 00:28:19,380
bunch of really quite obscure words

766
00:28:16,880 --> 00:28:23,480
including aflage which is a type of

767
00:28:19,380 --> 00:28:23,480
irrigation system originating in Oman

768
00:28:24,419 --> 00:28:29,220
we can run the same exhaustive solver in

769
00:28:27,059 --> 00:28:30,720
hard mode as well and it fails to solve

770
00:28:29,220 --> 00:28:32,640
in hard mode a lot more than an easy

771
00:28:30,720 --> 00:28:33,900
mode which you've probably experienced

772
00:28:32,640 --> 00:28:35,640
if you play in hard mode yourself

773
00:28:33,900 --> 00:28:37,260
there's a lot of cases where you

774
00:28:35,640 --> 00:28:39,840
accidentally get locked into a sequence

775
00:28:37,260 --> 00:28:41,880
of equally likely answers where an

776
00:28:39,840 --> 00:28:44,100
impossible answer you know might split

777
00:28:41,880 --> 00:28:47,220
up the options so for example this this

778
00:28:44,100 --> 00:28:49,440
word curvy here if you could make the

779
00:28:47,220 --> 00:28:51,480
word if you could guess the word venz on

780
00:28:49,440 --> 00:28:53,580
the third guess you would guarantee a

781
00:28:51,480 --> 00:28:55,860
solution for those curvy Kearney Kerdi

782
00:28:53,580 --> 00:29:00,720
or you get a choice between curly and

783
00:28:55,860 --> 00:29:03,059
Curry so you get one or two guesses you

784
00:29:00,720 --> 00:29:05,039
can't use vans in hard mode because it's

785
00:29:03,059 --> 00:29:07,440
already said that S and E aren't in the

786
00:29:05,039 --> 00:29:10,080
solution so you can't use that

787
00:29:07,440 --> 00:29:12,240
um and so we failed to solve in about 91

788
00:29:10,080 --> 00:29:14,340
cases now uh there are four words and

789
00:29:12,240 --> 00:29:16,140
unless again bugs I don't know unsure

790
00:29:14,340 --> 00:29:19,559
what's going on there it manages to

791
00:29:16,140 --> 00:29:21,360
solve about 96 of cases in hard mode uh

792
00:29:19,559 --> 00:29:22,919
finally one small optimization you can

793
00:29:21,360 --> 00:29:25,980
make to the word list which is to remove

794
00:29:22,919 --> 00:29:28,020
plurals the nyt removes words that end

795
00:29:25,980 --> 00:29:31,020
in s or ES if they're plurals or

796
00:29:28,020 --> 00:29:33,179
transitive verbs it changes the opening

797
00:29:31,020 --> 00:29:36,899
guess from tears to perch which means

798
00:29:33,179 --> 00:29:38,820
lively and we now solve in 97.5 percent

799
00:29:36,899 --> 00:29:40,500
of cases but now there's seven words

800
00:29:38,820 --> 00:29:41,820
where it's uh where the answers aren't

801
00:29:40,500 --> 00:29:44,159
in the guest list

802
00:29:41,820 --> 00:29:45,899
so to recap it's reasonably easy to make

803
00:29:44,159 --> 00:29:47,940
a good word or solver it was a fun

804
00:29:45,899 --> 00:29:49,700
exercise I encourage you to do it or you

805
00:29:47,940 --> 00:29:51,899
can copy my code whichever you prefer

806
00:29:49,700 --> 00:29:54,840
threads aren't great for Wordle style

807
00:29:51,899 --> 00:29:57,720
problems but process pools are my

808
00:29:54,840 --> 00:29:59,159
sources get another word list uh which

809
00:29:57,720 --> 00:30:01,559
you can download to make your own word

810
00:29:59,159 --> 00:30:03,799
games the original version of Wordle by

811
00:30:01,559 --> 00:30:07,200
Josh Wardle for the actual answers list

812
00:30:03,799 --> 00:30:08,700
uh the uh introduction to the wordlebot

813
00:30:07,200 --> 00:30:10,679
algorithm from The New York Times They

814
00:30:08,700 --> 00:30:12,600
explain how all of that works and if you

815
00:30:10,679 --> 00:30:14,039
want to read my own code and point and

816
00:30:12,600 --> 00:30:16,559
poke holes at it

817
00:30:14,039 --> 00:30:18,740
it's available on my GitHub thanks

818
00:30:16,559 --> 00:30:18,740
everyone

819
00:30:21,480 --> 00:30:23,600
thank you