you can watch similiar experiments on stars quite frequently

I wonder if the bot programmers would have any helpful insights wrt bot detection etc

Two or three.

-3

My guess for a while has been that online poker will be a rake-free "casino" game in the future, i.e. you can choose to play basically any cash game variant, any time, any stakes, but against a near-GTO house bot(s).

No need to market towards recreational players and house can take against any caliber of player in the world.

GG.

some 2p2 hero needs to convert HHs in the sciencemag link (data section) for easier viewing this formatting is terrible for reading

only darren elias and chris ferguson played vs 5 of the bots. majority of the players took part in the 1AI 5human section, and the 1 bot vs 5 human winrate was larger than the 5 bot 1 human lossrate

joey plz put it out there that we need these HHs converted for easier viewing

EDIT: whats the most popular free online HH viewer? i'll pay someone on fiver to do this

The cloud has already been used a long time to solve the GTO of NLH (and PLO) in 6-max. The low cost is new info for me.

We all have the GTO software of NLH but it takes too much time to compute real time. The real-time advisors are faster (using memory or/and simple calculations, I suppose). This as so is not new really (I don't know how they work). Preflop as cloud and then one could use simple math only for the advisor that only has the difficulty of considering the other player(s) ranges. A real-time advisor does this faster.

"128 GB of memory [RAM, I suppose]. On average, Pluribus plays twice as fast as typical human pros: 20 seconds per hand."

Only some players have that much RAM; most have from 2 to 16 GB. 180 hands per hour. This strong play as so needs more than what players have. but is easily reachable.

It only broke even vs. the best human play. It is not too much an issue but when considering the rake.

The 6-max bots I have played against before have a bad understanding of the situations and leak a lot of money by bluffing stacks and calling stacks in many situations where a human would do better, and as so are major losers. They additionally don't adjust, making it easy for a human. These superior human skills are not available for mere learning-bots. The GTO bots are unbeatable but the only way to lose to them is by making mistakes.

ffs i wanted to hear his opinion

Dream machines for live games coming to mobile phones near you soon

So it was 10k hands divided up between the 12 players. and they used an equation that can only figure their bots win rate and is not applicable for human play. it doesnt say how many hands were played by each player. it doesnt sound like its enough info to determine if it would be a winner in 25/50 on stars . how can the equation determine a bots win rate and not a human. so say it beats linus for .5bb over 2k hands that doesnt mean anything really. and 2bb/100 winrate against 1h 5 bots with chris ferguson doesnt sound very high at all really. i wonder what linus or any of the top 6m guys think.

Ok, but how would it do vs Durrrr?

they used a variance reduction technique called AIVAT (https://arxiv.org/pdf/1612.06915.pdf) which supposedly allows them to reduce std deviation by factor of 3 and use ten times less data to achieve the same statistically significant conclusion

Lmfao was the point of that to flex on how the bot “owned” Linus?

+1

Ferguson has a PhD in AI, so is probably friendly with the AI researchers.

Wow, pretty sick thin-value river x/r with QJ in the first hand.

Second hand seems like a questionable hero call with A7, though. Seems like a good bluff from the human who should have a lot of Tx +. Then again what do I know? I'm not beating Linus for 0.5bb/100.

A7 call is standard, the peel with AJo pre was lol_terrible though

the QJ hand was nice and quite unusual

they did cherry pick hands that made pluribus look good though, which is kinda dumb, explanations below were obviously made by someone who doesn't play poker much at all

spent some time going through the hands manually (looking for big pots which pluribus was involved in) and found a few (below), but it was exhausting... also the stacks for each game change throughout the game, but stack sizes aren't listed in each bit of data, so it's tough, but i have a friend who says he can reorganize the data so check back here for an update. hoping someone beats me to it though, seems tough

sample

odd, the facebook article (https://ai.facebook.com/blog/pluribu...-player-poker/) says there were 3 pros who played 1 human v 5 AI (linus, elias, ferguson) and the sciencemag article says only 2, no linus

So it beat Linus. gg poker

Please post more hands, they are very interesting.

Would be way more fun if there was actual strong online 6max cash game regs only. Majority of guys are either hu regs or mtt grinders lol..Would be priceless to see russians finest, zaruba vs this bot.

game over for real money online poker as we know it.
might as well transition over to play money and play against the aussies before stars makes a announcement that you will now have to play against 1-8 house bots that play GTO in any format. People will too. maybe stars might be nice and award prices to the players that play the best against the bots. turn poker into a total skill game. actually nah they won't work either since everyone will have a bot. Game over.

Lol I read this page, watched the videos etc. I have no doubt that they created a dream machine bot but the stuff thats being stated in the first video is kinda funny.

"This is a complex hand because 4 players remain in it, Most professionals still have trouble in these situations".

Bot then goes all in for a big overbet with a gutshot and flush draw on the flop. The video calls it a "massive bluff".

It's pretty terrible for online poker. Sites cannot enforce no bot rules no matter what they say. I think some top pros have had very close to optimal dream machines for several years now.

They had a way of making 10k hands significant apparently, I quote from the webpage https://ai.facebook.com/blog/pluribu...-player-poker/

Not sure if I would have trusted this had I been a human opponent.

one of the bot creator's accounts on another site, answering questions about the bot (https://news.ycombinator.com/threads?id=noambrown)

he's saying they want to make it as accessible to the AI dev community as possible, but aren't releasing the code because of the damage it could do to online poker

these guys that work on this bot are sitting on code which they could easily morph into bots and generate millions, but aren't leaking it so us peasants can grind for quarters and nickles for a while longer. i have a much more dim view of online poker after reading this