Minor readme fix

Author: lightvector

README.md

@@ -474,7 +474,7 @@ The following test shows that the strength of play scales massively and rapidly
     value28-140-6v    : -745.3  75cf (-780.3,-713.3)  95cf (-810.3,-684.3)  ( 57.0 win, 588.0 loss) 
     value28-140-3v    : -893.6  75cf (-937.6,-855.6)  95cf (-976.6,-821.6)  ( 25.0 win, 623.0 loss) 
 
-Displayed above are the stats for the bots in those tournaments where primarily I was testing the effect of varying the number of visits. "value28-140" and "value33-140" are neural nets trained on Leela Zero games LZ105-LZ142 and ELF games as of about June 27 for 140 million training samples, differing only in some insignificant simplifications to the head architecture. They have both converged well and appear to be well within statistical noise of each other in strength. You can see in the table above the maximum likelihood Elo ratings of these bots ranging from 3 visits to 1200 visits given the data assuming the standard [Elo rating model](https://en.wikipedia.org/wiki/Elo_rating_system) where the probability of a win given a certain rating difference follows a [logistic curve](https://en.wikipedia.org/wiki/Logistic_function) scaled so that 400 points difference implies to a 10:1 odds.
+Displayed above are the stats for the bots in those tournaments where primarily I was testing the effect of varying the number of visits. "value28-140" and "value33-140" are neural nets trained on Leela Zero games LZ105-LZ142 and ELF games as of about June 27 for 141 million training samples, differing only in some insignificant simplifications to the head architecture. They have both converged well and appear to be well within statistical noise of each other in strength. You can see in the table above the maximum likelihood Elo ratings of these bots ranging from 3 visits to 1200 visits given the data assuming the standard [Elo rating model](https://en.wikipedia.org/wiki/Elo_rating_system) where the probability of a win given a certain rating difference follows a [logistic curve](https://en.wikipedia.org/wiki/Logistic_function) scaled so that 400 points difference implies to a 10:1 odds.
 
 Also displayed are the symmetric 75% and 95% confidence intervals for each given bot *assuming all other bots' Elo ratings are taken as given* (so it does not account for the joint uncertainty or covariance between the various bot ratings, but is still a good indicator of the error bounds on these ratings). Several of the bots have many more games than the others, this is because those bots also participated in many other round robin tournaments against other bots not shown.