I'm thinking eddie as LLinus too although I haven't seen enough hands to say for sure.

MrBlue managed to win solidly at 15bb/100 over almost 10K hands so I was originally thinking it could be him but as I've looked at more of the hands it's become clear its not. (For example, MrBlue is calling really wide in the SB, not 3Betting much, not betting much on flops etc).

Yeah, I don't disagree - their stats would be different in that scenario.

What I meant though was if you got small stakes regs (who have looked at their own leaks in PT/HM a bit, say) to play in a tough reg game, I think they'd often have more solid stats than these pros tbh.

Watched both video, didn't see a single example of a multiway pot. Are there some in the HHs?

This doesn’t seem like any sort of consequential, valuable, advancement in AI to me. PokerSnowie probably could have beat top pros in a 6max NL match in a more convincing way, and could’ve done so for years now.

Apparently this Noam guy is supposed to be a respected researcher, I’m sure he’s a smart guy. But this just seems like a bunch of guys are incentivized to market themselves as genius AI researchers doing valuable, productive work, by being the guys that created “the FIRST AI to BEAT the TOP POKER PROS” when really they didn’t do anything genius, or productive, or valuable, at all. They did a rehashed version of PokerSnowie that lost 700bb over a small sample against some mediocre players and Linus.

Certainly doesn’t seem like any sort of new, consequential, AI advancement. Just seems like nerds marketing themselves as productive AI researchers for their own personal gain while doing nothing of actual value. They saw an opportunity to “conquer the prestigious milestone of beating humans in a multiplayer hidden information game” and did it in a half assed, super unconvincing way lol.

Wouldn’t this imply that there is basically no room for improvement for top players? Top players in 5-10 years aren’t going to be that much better than the top players today?

Intuitively this seems wrong to me. Linus probably makes a ton of mistakes. They’re just subtle enough that no human can spot them all and take advantage.

I don't have PT4 but I've done some analysis. Can anyone confirm if these winrates are correct for the hand sample? Why don't ya'll think Linus is Budd?

https://imgur.com/a/cWpIpTC?

His flop cbet is too low for one. My candidate for being Linus is MrPink from a quick look at stats.

Seat 1: Pluribus (10000 in chips)
Seat 2: MrWhite (10000 in chips)
Seat 3: MrBlonde (10000 in chips)
Seat 4: MrBlue (10000 in chips)
Seat 5: MrOrange (10000 in chips)
Seat 6: MrPink (10000 in chips)
Pluribus: posts small blind 50
MrWhite: posts big blind 100
*** HOLE CARDS ***
Dealt to Pluribus [Ts 2h]
Dealt to MrWhite [2d 9c]
Dealt to MrBlonde [Ac 7c]
Dealt to MrBlue [5d 9s]
Dealt to MrOrange [Ad Jh]
Dealt to MrPink [8c 3d]
MrBlonde: raises 150 to 250
MrBlue: folds
MrOrange: folds
MrPink: raises 800 to 1050
Pluribus: folds
MrWhite: folds
MrBlonde: folds
Uncalled bet (800) returned to MrPink
MrPink collected 650.0 from pot

MrPink is clearly beyond awful I don't know what you're talking about

I've looked at quite a few of MrPink's hands and I'd definitely say he seems good. He does seem to call pretty wide from the SB against raises (for instance with 22) and not 3Bet much from there which isn't what I do, but it might be ok. He's also bet flop then given up on hands on the turn in a couple spots where I wouldn't have, but again, he could just be better than me and I'm not fully understanding his strategy. He definitely hasn't made any clear big mistakes in the hands I've looked at (barring the 83o but something makes me think he didn't do that on purpose to 'balance his range').


Out of interest- was there anything that made you think that eddie wasn't Linus? He seems solid from the hands I've seen.

lul. yeah it could be eddie, don't see anything obvious that would disqualify

I downloaded PT4 trial and had a look. I think it's pretty clear the best players are Budd, MrBlue and Eddie. Linus could be any of these three I think.

Cbet flop % by players
Budd 54.5%
MrBlue 46.7%
Eddie 61%
Pluribus 53.7%

There's no way you can dismiss Budd as Linus based on this stat alone.

I agree. MrBlue has some leaks this after looking more closely. The best players are Budd and Eddie.

why are these regs making terrible preflop plays? they are obviously not bad regs so why are they limping k5o pre utg, ect. 34s open MP, the Q9o hand, 83o.

This is terrible news for the poker community.

Sure, it was all announced and we knew it was coming but not so immediately and with the low resources needed to run it. It is actually insane for being the first bot that can beat multiway pots. Imagine when they improve it, It's going to be possible to run it on the phone.


So is this the end of The Poker Online carers for all humans?

What can do the online casinos to fight back bots, this one is even more affordable than the previous ones... People can use the bot in a different computer via screen reader so I can't think about a way to detect bots.

What do you guys think? It´s gg wp already or we should still work hard.

I think this has little to do with real advancement in poker AI. There is a war going on between Google and Facebook — along with other Silicon Valley startups — to hire and retain top AI talent. Facebook hired Noam in November in part to run this AI recruiting campaign. They need something to compete with AlphaGo and AlphaZero from Google which crushed humans in Go and Chess. Facebook marketing and recruiting teams need some catchy headlines to convince the reinforcement learning research community that they are not a laggard. I doubt Noam would have supported these claims and conclusions if he were still at a CMU lab.

I doubt it, in 2015 they claimed to have just weakly solved heads-up NLHE (using supercomputers?) and solving six-handed is several magnitudes more difficult if it even is possible. But when/if they solve NLHE six-handed it's time to quit online NLHE poker and perhaps move to a more complex online poker game or live.
https://science.sciencemag.org/content/347/6218/145

It would be easy to come up with games that are a trillion times more complicated than NLH. Suppose for instance five cards are exposed before the flop and five more after.

Edit: they solved LIMIT heads-up, sorry for the error

Top players are not balanced with optimal frequencies everywhere, but a balanced ai that isn't exploiting them will only beat them when them when they choose a 0 frequency action.

So imagine a looking at a piosolver sim. A bunch of stuff is mixed. Vs a balanced opponent that won't adjust, it doesn't matter what frequencies you take those mixed actions. Spots where someone would lose to a balanced ai are spots where they would choose an action that they should never do at any frequency. Sure top players still do this, poker is complicated etc, but these spots just aren't common enough for them to lose at 40bb/100.

Ai can keep improving and getting closer to optimal, but its winrate won't grow substantially.

Pre-computed solver databases have been a thing for years now (wasn't the Skier5 database for HU SNGs available from like 2016?) and those are much more accurate and dangerous than this bot. Nothing has really changed.

High stakes has been dead for a long time now. Reg on reg battle in 2019 is just an exercise in who has the more accurate solutions. The only reason Oborra was caught and banned was because he went about it in the most asinine way imaginable. People with a bit of subtlety can easily mask the use of real time assistance without giving up much if any EV.

^like shoving 17x pot.

Sure, it was used from some time now BUT this new AI doesn't require high resources to run it like the previous one, meaning that almost anyone could use it from home without spending a lot of money, as long as they can get access to it. So now any pleb could run it even in middle stakes.

I was under the impression that this AI was capable of exploitative play. It can have a pseudo-GTO baseline from trillions of hands of self play, but it also has the capability of making in-game adjustments to maximize win rate.

I don’t think Linus would ever lose 40bb/100. I think a bot like this could beat a table full of Linuses for 40bb/100. I think a good number of the exploitations it could come up with would lead hands down branches of the game tree that no human has really studied much. IIRC there was a HUNL bot that crushed some pros and it’s play style seemed pretty bizarre. Obviously there’s quite a bit less room for this sort of thing in a multiway game, but 6-max human games play pretty much by rote when there’s no fun players in the mix.

In a co v. bb pot, the bot could do something like c-bet 3.75x pot on the flop, 71% pot on the turn, and 2.55x pot on the river. Even if the human played mdf perfectly, they could still end up getting massively exploited because of the peculiarities of the range asymmetry on that exact board. No human is ever going to be able to figure out the nuances of something like that in real time, and I’d be willing to bet that no human has ever thoroughly studied how to play against that line on any flop texture/runout.

So you think Linus, Katya, gale etc are all cheating?

Not saying the particular players that are mentioned are or aren't but surely it is just as suspicious having players battling it out all day every day as a new player suddenly appearing and winning. Just because players are 'old school' or have played HS for a while should not discount them from having assistance. It would imo mean that they are much more likely to be able to hide it better because they have the knowledge on how to deviate correctly vs fish/weaker players

There is equilibrium in multiplayer 0 sum games. Equilibrium means that no player can improve on their own (changing their strategy without others doing so as well). There might be many solutions like that all with different EVs for all participants. This and the fact that even if you follow the equilibrium you can get less EV if other players don't play the equilibrium strategy means that multiway equilibrium is "shaky".