This is too sick, rip humanity

This showed up in my newsfeed a couple of hours ago, about the challenges of teaching a computer to play NL, with bluffing. And they also talked about what other real-word applications the program might have ...

Jungle and Ike should collab and ruin it

god damn this **** is scary

Wasnt Jungle losing to Durrr after the first couple challenge sessions, then he adjusted and pwned? If humans find some exploits, the bot wont readjust so the humans will just print from there on out

I just figured that even if you're raising a polarized range versus button limps there are plenty of better hands than 82o you can pick as bluffs.

The rules of this game are however highly stacked in favor of the bot. I mean resetting the stacks to 200bb each hand means the bot does not need different strategies/ranges for different stacksizes. Other than that I think the tanking on turns and rivers is pretty ridiculous. That makes it so much harder for humans to play. Bots don't fatigue humans do.

I was pretty interested in watching this match roll out, however they somehow managed to make it barely watchable. For the next time please:

1. use a proper table theme! Currently it looks like ms paint and you can't clearly see whos turn it is, what the potsize is and the betting amounts.
2. Use a short timebank comparable to zoom poker. The bot tanking for ages in miniscule pots is ridiculous and makes the game so boring (I wonder how they will get the required amount of hands in).
3. The bot could always show his holecards after the hand played out. That is more entertaining for the viewers and as the bot is playing (aproximated) GTO, it will not hurt its strategy.
4. provide handhistories for download at the end of the day.

Thanks!

No money in HU, super computers too solid

being down after 6k hands doesn't seem very significant, esp. because, with the way this match seems to play so far, stdv will be quite high.

I also assume that the edge of team brain increases towards the end of the match since they will get better at adjusting to the bot.

Has anyone posted current status of standard deviation?

In the twitch chat Jason Les said it was a bit too technical for most people but that he could make a recap at the end of each session.

Quite funny that people wanted to bet on the AI at 4:1 and now the AI is a favorite. At 3k hands per day those odds estimate his winrate to be 5bb/100.

-removed-

"In a similar tournament of 80,000 hands of poker played last year, the machine, an earlier iteration of the AI lost, but just barely, to the humans."

lol out of statistical tie now evolved this

Haha nice misleading statement

Would love to beat someone for real money just barely over 80k hands @9bb/100 HU

Did you win bro? Just barely, 72 Bi's

nice comeback jimmy

Booo Daniel boo!

Would have been quite a good 2nd day if dan didn't get burried but looked like first half of the day was run bad (while jimmy got some run hot on opposite mirrored side) and then he kept yawning and was probably slightly tilted for second half. The prof kept saying the AI averages 15 seconds per entire hand but it seems like every hand that went to the turn took 30 sec- 1.5 mins to tank on a decision for turn + river, so that average must be from the times the bot snap calls then snap folds to flop bet.

One thing I thought was kind of funny was the professor when asked yesterday what happens if the bot wins this one and he said if these 4 cannot beat it then -nobody- in the world can. Obviously these guys are really good HU players but I feel if money wasn't an option there might be another lineup you could put to the test? doug sauce ike giving them all a bit of time to prepare and throw in a 4th crusher?

Also I think it's silly how many people are screaming the sky is falling in 6k hands (at least in the twitch chat more so than the couple of posts here) but even if the bot does end up winning after 18 more days it's not like someones going to rent out super computers put in millions of dollars and years of research to replicate the bot to hop online and play HUNL which poker sites have been actively trying to kill for a couple years now.. Also he said that 6 max would be incredibly harder due to no good way to measure progress and develop the AI.

I'm still 100% certain claudicus would have raped the humans last time if it had been able to keep a database on the humans.

You don't need a super computer to do this TreadLightly, there's already HU bots that beat everyone but the top 5% of hu regs at every limit. A nice neural network with 7 years of accumulated data shared between millionaire programmers.

3 or so years ago, i found a weird one that i thought was a fish because its play style was similar to claudico. I gave the information to a few really good rich HU players and they all gave up pretty quick.
first couple hundred hands were easy, it was really stupid. Then it had every info on my timing, sizing and thought process to completely kill me even though it seemed to be exploitable and dumb.

neu·ral net·work: (n) a computer system modeled on the human brain and nervous system.

i dont think you used that term correctly. it usually refers to terminators.

all of this!

Interested to bet on Libratus, PM !

Comments during this challenge by the Professor and how he and the University interpreted AI's defeat last year as a "statistical tie" caused me to develop a very healthy skepticism about public science and experts and how both can be just as biased and corrupted by monetary/reputational rewards as the rest of us. Perhaps that will be the real benefit of the this contest.

p.s.-- If the AI wins even one dollar more than the humans in this contest there will NOT be a rematch with "better" human player for reasons noted above.

Hi, I created Claudico (not Libratus). I wanted to say first that I am very impressed with what the team has done for Libratus. I agree that if they win, it is unlikely that they will do another challenge (why risk a "loss" after they have already "won"?)

I would like to run another challenge with my team against doug, sauce, ike, and jungleman in approximately one year after this, assuming we can work out a payment scheme both sides agree to and I can find a way to raise the money. My bot would not tank nearly as long as Libratus and would likely not use a supercomputer. I'd like to do this at Seminole HR, and I think it could be ready one year from now.

Sorry if this has already come up regarding the timebanking - is it just running a solver before each decision? The software's already out there for anyone to do that.

Hopefully the players were warned in advance and knew what they were getting into. Reminded me of that guy on Stars that tanked every hand in the no blind increase hu sng's.

Hello Dr. Ganzfried, who are you working with now? Any more details on the project?