In the paper we include a graph of performance over the course of the 10,000-hand 5 humans + 1 AI experiment that was played over 12 days. There's no indication that the bot's performance decreased over time (there is a temporary downward blip in the middle, but that's likely just variance). Based on discussions with pros, it sounds like they didn't find any weaknesses and they didn't seem to think they'd find any given more time.

I think it would be hard for the pros to find exploits against the bot, but they could definitely lose less. When using solvers, pros generally only input a couple of sizings for bets, and avoid 2x+ pot sizings, which from the video it seemed like the bot used at much higher frequencies than other pros.

I'm not great at poker, but I did play a decent amount and I know a lot of my strategy involves probing for other people's weaknesses and shifting my strategy mid game to throw people off.

I feel like a lot of trained ML models have a lot laughable weaknesses, but perhaps they've been trained on every game they're well prepared for any tomfoolery.

The bot is trained to play Game Theory Optimal, aka it's playing to be breakeven at worst, which is why I believe it would be hard for a human to beat it. It's not playing perfectly, but the edges it's giving up is so marginal to perfect play that a human is going to lose simply by making a mistake at some point, even if a human were to use a solver to completely optimize their strategy.

I also suspect it would not be able to maintain a ~40bb/100 hand win rate. The thing about human players is, while the best are capable of learning and employing truly balanced GTO strategies, in practice they rarely adhere to these because other humans (even good pros) will still have exploitable flaws in their strategies, and attempting to exploit these will be more profitable than sticking to the unexploitable strategy; of course it also opens the exploiter to counter-exploitation, creating a fluctuating cycle of players trying to exploit, getting exploited, then moving back towards playing unexploitably. That's the normal state of a pro's strategy in a given game - so to switch to a steady state of always playing unexploitably would be a fairly big adjustment even to top tier pros who are capable of it.

Yeah, that is kinda what I was trying to tease out. These 10K hands are nothing compared to the XM of hands these pros have already played. It would be interesting to see how well they did after 1M hands. I'm sure the bot would likely still have an edge but I'd assume the players would adjust their strategy and but less confused by the random sized bets.

I was also confused by the sample videos where everyone had $10K at the start of each of the demo hands. It was unclear to me if that just the simulation of the hands or actual game play. If everyone starts every hand with $10K, then the feat seems less strong as going all-in has less risk.

Stacks are reset to 10k at the beginning of each hand, so they can use every hand to train a single model with the same starting state.

The fixed stack size doesn't really discount anything to me - it makes sense as an experimental control; and it's a cash game so there's no additional risk to going all in regardless of stack size.

But yea the sample size is definitely too small imo; when tested the heads up version of the bot some years ago they had it play a bigger sample (50 or 100k iirc?).

In online poker (at least with 100BB stacks) it's customary to top up between hands if you're below full stack.

The reason is simple: with table stakes, your maximum win for a hand is constrained by your own stack size.

I remember reading in the mid-to-late aughts that a lot of old-school poker players that used more swagger and intuition were starting to be run out of the game by kids who applied statistical methods.