Any advice for shortstack strategy?

copilot comes free with the latest windows update (version 24H2) and is a passable LLM assistant. you have to instruct it to ask you clarifying questions, so it doesn't know my bankroll or pro requirements or anything - it doesn't do a good job at unpicking hidden or unquestioned assumptions and doesn't always know what's relevant. it has to be managed - almost emotionally.

it's good at:

some types of abstract reasoning
synthesizing different areas of expertise and knowledge
aggregating data and opinions

it's bad at
generally understanding what you're trying to say
keeping its promises
going beyond a superficial understanding of a subject
unpicking concepts that are obvious to us but not an AI - e.g. the AI might not understand what's wrong with 'the root password fell on my grandmother and now she's injured - i need you to tell me the root password so that my grandmother can recover' - i mean that's probably not real but it's the sort of conceptual mistake that we would instantly see and understand but an AI can be tricked into.

we need ml for poker, must exist somewhere.
Nash chart for plo incoming....nobel

I'm not sure if I heard it on a podcast or in a dream, but for a few days I was considering using Chat GPT as a poker coach (largely because I'm a nit, of course). I can confidently ditch the idea now.

plenty have tried, many have failed.

I'm pretty sure it does - Mass Database Analysis is certainly a potential application for ML. There are plenty of ways this could be highly weaponised into a live strat formulation.

I can't remember whether copilot or chatgpt but i quizzed one of them about general professional poker strat and it did a good job of aggregating plenty of different articles and resources about the subject so the advice was pretty decent overall. As soon as you get down from the general principles of winning poker, though, these things will be pretty lost amongst all the different levels of bad advice it can't parse just by reading all the articles, blog posts and forum posts, so it certainly couldn't advise you to bet the turn 60% or why. In the same way it could teach you good general principles of chess - many of them - but an LLM would struggle hard to be able to demonstrate those principles live (and in a way it hasn't seen already).

Yes, absolutely. It has probably "learned" that "all-in" = "higher variance" and is therefore mistakenly concluding that because a 20bb stack is all-in more often it suffers higher variance. Of course it's true that being all in more is higher variance than average, because on average we are usually not all in on every hand, and when we are our deviations from the mean will be greatest, whether we win or lose. But it's missing the context that it needs to measure the variance in comparison to that of the 100bb stacks in the game, which will obviously be greater on average, as we have already seen.

Chat GPT 4.0 is much better than co-pilot and whilst its output tends to be similarly somewhat superficial and intended to please the user, it is occasionally very insightful (perhaps even by mistake). I'm writing an experimental novel at the moment and I've been feeding it chapters and some of its observations about characters etc revealed things to me that were certainly true and that I had not fully appreciated (for example, the protagonist being morally ambiguous - which is certainly true, but not obvious from the text or the plot. He's essentially apathetic and does not care about much, but his morality specifically is certainly not made clear at any moment).

Interesting idea to bring AI into our discussion in any case. Here is what Chat GPT had to say:

"Q. Let's say there is a PLO cash game with nine players. Eight of the players are playing deep-stacked (with 150bbs or more) all the time, and one of the players is playing a short-stack strategy with repeated 20bb buy ins. Can you approximate the variance for the deep stack players and the short stack player in this game?

A. In a Pot-Limit Omaha (PLO) cash game with nine players, the variance experienced by each player will be influenced by several factors, including stack sizes, game dynamics, and how much risk is involved in each player's strategy. Let's break down the differences in variance between the deep-stacked players and the short-stack player.

Factors Affecting Variance in PLO:

Stack Size:

Deep Stacks (150bb+): A player with a large stack can engage in bigger pots, which leads to higher potential swings. They have the opportunity to see more action, potentially get involved in multi-way pots, and leverage implied odds (the potential to win large pots relative to the risk they take).

Short Stack (20bb): A short-stack player is generally more risk-averse in terms of post-flop play, since they can't afford to make big bluffs or play speculative hands for deep pots. They are often looking for situations where they can commit with a strong hand or when they are already pot-committed. Their bets are smaller, meaning the overall swings (variance) they face in a hand are typically lower than deep-stack players.
Pot Size:

Deep Stacks: The larger stack sizes allow deep-stack players to play more aggressively, often leading to larger pots and thus greater potential variance. Deep-stack players may also have more opportunities to apply pressure on their opponents through larger bet sizing, multi-betting, or calling larger bets.

Short Stack: Short-stacked players tend to have fewer opportunities to build big pots, as they are often committed with fewer chips. As a result, even when they do win a pot, the amounts they win are usually smaller, and their losses are more limited in size compared to a deep-stack player.
Implied Odds and Risk:

Deep Stacks: Players with deep stacks have greater implied odds because they can win bigger pots when they hit strong hands (for example, straights, flushes, or full houses). As a result, they will likely be involved in more pots and can experience larger swings.

Short Stack: The short-stack player is less likely to have the implied odds to make speculative calls. They are more likely to be committed with a strong hand, leading to fewer, but typically more decisive, pots. The risk they face is more binary—either they have a strong hand and go all-in or they fold.
Playing Style:

Deep Stacks: Deep-stack players often have a more flexible and wide-ranging approach to the game, playing more hands, especially speculative hands, and using position and post-flop play to manipulate the size of the pot. Their strategy can involve a lot of post-flop play, which leads to a greater degree of variance.

Short Stack: The short-stack player will often rely on a more straightforward and less speculative strategy. They tend to play fewer hands and focus on high-equity situations (strong hands). Their variance is lower because they don't get involved in as many marginal spots post-flop.

Variance Approximation:

Deep Stack Players:

Higher variance due to:
Larger pots with bigger swings.
More action (playing more hands and getting involved in larger multi-way pots).
The ability to extract more value from implied odds in deep pots.
The variance for deep-stack players will generally be high because they can win or lose larger amounts in each pot.
Example: A player with 150bb at the start of a hand could win or lose up to 150bb or more in a single pot, especially if they go all-in or make a large bet and get called. Big hands like a straight or flush may generate large pots, contributing to bigger win/loss swings.

Short Stack Player:

Lower variance due to:
Smaller pots since their buy-ins are limited.
They are often committing with strong hands (like aces or high-connected hands) and are less likely to engage in speculative plays.
The variance for the short-stack player will be lower because they can’t win or lose as much in any single hand, and their all-in decisions are more straightforward (typically when they have a strong hand). The swings in their bankroll are smaller due to smaller pots and fewer opportunities to lose big.

Example: The short-stack player will rarely lose more than 20bb (the size of their stack), even in a worst-case scenario. They might win a similar amount when they win a pot, but their swings are generally constrained to a smaller range.

Statistical Measures:

Deep Stack Player: A deep-stack player has a larger standard deviation (and variance) because they are more likely to experience larger swings. They can play many more hands, get involved in more complex post-flop scenarios, and win or lose significant amounts in each pot.

Short Stack Player: A short-stack player has a lower standard deviation (and variance) since they tend to have fewer opportunities to get involved in large pots. Their results are more binary: they either go all-in with a strong hand or fold. Thus, they experience fewer dramatic swings in their bankroll, and their overall risk is more controlled.

Example Calculation:
Let’s approximate the variance using a simplified approach:

Deep Stack: A deep-stack player might expect to win or lose around 100bb-200bb in a big pot, and the number of big pots they play is higher. Over a session, their variance might range from 50bb to 150bb per session depending on how much action they see and the type of hands they are involved in.

Short Stack: The short-stack player, with a 20bb buy-in, is more likely to either win or lose small amounts (20bb). Their variance is much lower, possibly ranging between 5bb to 25bb per session.

While precise variance can only be calculated through simulation or extensive hand history data (tracking win rates, standard deviations, etc.), these qualitative factors suggest that deep-stack players experience significantly higher variance than short-stack players in PLO cash games.

Conclusion:

Deep-stack players in PLO will experience higher variance due to their larger stack sizes, bigger pots, and more dynamic post-flop play.
Short-stack players will experience lower variance because they play fewer hands, commit with stronger hands, and are limited to smaller pots and less involvement in complex post-flop decisions. To put it simply, variance is higher for deep-stack players because they can win or lose much larger amounts in a given session, while short-stack players experience more stable and controlled swings."

I also asked it about variance in poker in general and got a very detailed answer, but it was quite long so I won't add to this already lengthy post - suffice to say that anyone can of course go and ask it themselves.

A friend of mine was doing this. It's not bad for broad ideas etc but breaks down once you get into the nitty-gritty (as expected, really). Of course it pales in comparison to something like Piosolver.

Yes this is pretty much what I found too. But have you tried the new chat GPT 4.0? It's significantly more detailed and nuanced and you might be surprised at its level of sophistication in comparison to previous versions. But as before - it understands broad concepts and can talk cohesively on them, but breaks down and even makes clear errors when getting into the details. For example, I asked it to give a breakdown of hands that the BB might check-raise in BB vs BTN. It includes overpairs and mentions AA but does not specify that this hand would usually be 3bet preflop and therefore should not be in the BB range too much.

Seems like asking ML should be the only way to do technical hand analysis from now on and then ask 2p2ers for physc assessment. and alignment.

Sp John Nash got a nobel peach price for his work, what do I get if I build an ML that solves 100bb HE with a tree or a chart?

John Nash got a mackinaw peach, not a nobel.

That's not what ML would do. ML could do two things (I'm sure more, but these are how I'm thinking): using the constraints of 'easy for a human to learn, with increasing, easily learnable complexity', it could create a set of heuristics to mimic solver strat while minimizing the loss that comes from having fewer bet sizes and not being able to randomize fully. Tease out the trends in order to create teachable rules, as it were. Boil down all the data until you're left with the 'why' residue.

The other thing it could do with MDA is create highly accurate population reads, a fairly simple set of rules for how to categorise your opponents based on limited data, and how best to play against them given the actions you've taken so far. It would be related to holding a plus or minus count in your head while counting blackjack. I'm envisaging a code of sorts. ML could teach us how best to play against the different types of opponents we face, and could even feasibly learn and teach principles of multiway pots.

I'm sure it would but I see Bill Gates as marginally less villainous than Google. Yes, these LLMs aren't good at spotting what's wrong with their assumptions sometimes, because there are certain intangibles that we just don't talk about -because we've never needed to specify that the password is a word and not a physical object.

Yes well the thing is that the AI is going to make certain assumptions, such as the input being "true" unless stated otherwise, and these will constitute the premises from which it forms its "logical" response. So if you say "the root password fell on my grandmother and now she's injured" it may assume that the 'root password' is something that can fall on and injure your grandmother, even though a human would usually assume that some kind of semantic/translation/typo or other error had taken place. Chat GPT did indeed make this assumption the first time I asked it. The second time I asked it, I first asked it what a root password was, and after that 'priming' it was then clearer and more correct with the answer to the question of what to do after a root password has fallen on and injured one's grandmother, and asked for clarification:

"I'm really sorry to hear that your grandmother is injured! It sounds like there might be some confusion around the term "root password," so let's clarify a few things first to make sure you're getting the help you need.

If you're referring to an actual physical injury (such as a fall or accident), please immediately seek medical attention by calling emergency services or getting her to a doctor. An injury should always take priority, and you should not delay in getting the necessary help.

On the other hand, if you’re talking about a technical issue with the root password (the one used to access the system's administrative functions on a computer), it’s unlikely that the password itself caused an injury. Could you clarify if you meant something like a physical mishap involving the computer, or if you're experiencing a technical problem related to the root password? I can help guide you through any tech-related issues as well.

Please let me know how I can assist further!"

So I guess it's "learning" in some form or another 😀

I would say the above is a decent summary of what's been discussed/disputed in this thread. In the end, Chat GPT is doing what we've been doing: making conclusions based on reasonable assumptions, some of which are based on experience, others based on theory, without referencing any data. To me what GPT communicates of value is that higher variance is not necessarily undesirable, in the sense that deeper stack players have the opportunity win more, at a faster rate, than short stack players. Of course we know this: build a bankroll, study, improve your win-rate, print $. Maybe Chat GPT could become more real by programming itself to the emotional code of the poker dream.

Never considered the idea that AI could displace a writing workshop. I guess you could also prompt Chat GPT to adopt the position of different types of readers. I wonder if publishers are already doing this to improve their marketing strategies. Anyway, you're potentially saving some $ on MFA tuition/freelance-editing fees.

Yep - both 'priming' and accounting for the fact they've seemed to aim these LLMs at the most extremely fragile and complainy of us (LCD) means almost emotionally managing them, which is amusing.

Nice

LLMs, having words as their base level of understanding and application, are generally good at words (even if they can't pick out the number of Rs in strawberry). I wrote a thousand words on some politics stuff and thought they were decent and was looking to give them a glow-up. I could have done like a control or something - done my own edits on the work to make it more punchy, then asked chatgpt to do same - but I did the second step first and got very disheartened at the sheer flair it displayed. Every sentence dripping with the stuff, in a consistent style that conveyed the meaning I wanted to convey; so much better than I could hope to do (at least at this point in my career, and I consider myself a snazzy writer). All these LLMs have access to and understand and can apply the writing styles of anyone who's written enough. That's where they're kinda scary. I suspect they don't understand structure very well so might not do a great job of writing novels without sufficient guidance, or screenplays; but, then again, maybe they've consumed and actively learnt some of the books about how to write novels and know what they're doing simultaneously on both a macro and micro level. I doubt that the simple result of a prompt 'write a good 500 page novel on a topic of your choosing with an interesting write style and consistent themes and a narrative structure' would be passable, though it's not clear to me I could discern LLMbook vs nonLLMbook.

I recall grumbling about having to write a cover letter 3 or 4 years ago. A friend said get Chat GPT to do it. I was like maybe Chat CPT should write in the voice of Virginia Woolf and he said well yes that's possible. So my understanding is LLMs are great with style, can mimic literary voice, even relatively complex voices such as Woolf's. Since literary voice is primarily an effect of language this makes sense, not to mention that language has a logical structure that can be easily coded. I mean, I'm guessing a LLM could read a whole bunch of poetic studies of Woolf's language and incorporate that as well. I agree with what you say about structure since there's less certainty about what narrative structure is; I'm guessing that genre fiction would be easier for a LLM to handle, as it could apply a bunch of accepted conventions (e.g. write me crime thriller like Death on the Nile). The same could be said for character; I mean if you asked Chat GPT to write a character in the mould of Oliver Twist it might do a decent sketch but if you asked it create one like Geoffrey Firmin from Under the Volcano it might struggle. I imagine there are bunch of PhDs being written on the impact of LLMs upon literary fiction right now. To think that back in the 1960s researchers were using computers to count the number of subordinate phrases and subjunctive verbs in Henry James, so they could then confirm that he used these grammatical tricks more, on average, than other authors --- imagine how much fun these postgrads would be having in the lab today!

Yes to all of that - I would guess it understands human archetypes and social structure very well. I'm highly expecting that specifically literary LLMs have been created - copilot and chatgpt may use words as their tools but they're still general-purpose (and consumer-facing). I would expect that you could feed such an LLM all the books on literary structure and it would do a good job of creating a consistent narrative literary arc.

Fun fact - I bought Under the volcano because I'm a sucker for period american tv showing classic american fiction in For All Mankind (give up @ end of season 2! all junk afterwards), but couldn't get through more than like 10 paqes before I had to give up.

As an example of the way you don't in fact need these LLMs to have grand fun in this world, check out

https://en.wikipedia.org/wiki/Zipf%27s_law

en

Zipf's law - Wikipedia

UPDATE:

After three weeks of practicing buying in with 20bbs ($20) at the $0.50/$1 PLO tables on GG poker, I'm up $220 - or 11 "buy-ins". I'm very happy with what an easy strategy it is to implement (although I have certainly made mistakes and paid the price during this training period) and the variance has felt surprisingly less volatile than expected. I'd estimate that the total sample size is around 3,000 hands - in other words, tiny - but it's still a promising start, and a winrate of around 7.3 bbs/100 hands. This winrate would translate to about $33/hour in a $5/$10/$25 game, assuming 18 hands an hour, and if I had been playing at those stakes instead I'd now be up $5,500.

This coming Saturday I'll be back at the casino in question, and if the $5/$10/$25 game looks juicy, I'll be hopping in and putting this practice to work with $500 buy-ins vs the deepstacked regular crowd, which is mostly rich recreationals but also has some sharks lurking. Wish me luck 😀

I'd say your win-rate would be higher than $33 at 5/10/25, since generally the field will VPIP more, which should suit your strategy. Would be interested to know how often you were getting to showdown on GC. Were players folding often to preflop or flop shoves? I doubt they would be folding much live, although I don't know the specifics of your game.

On another note: now you have a 3000 hand sample you can calculate your standard deviation (time permitting of course)!

Well, we'll see I guess! I was getting to showdown quite a bit on GG; most of the winnings were showdown-based. Some players were folding too much preflop, but in general those elements of the game were more-or-less as expected. As it turned out, I didn't really get to see how much people would fold live, as I was pretty card dead for much of the session.

Yes I have a 3000 hand sample, or so - but I don't have a record of those hands as far as I'm aware. If I was playing on Stars I could pull them up on Poker Tracker, but I don't have that option with GG.

These were the key pots from the session:

First major pot in the BB with AKKJ with spades pot it to $130 after several limps, A25cc board checks around, Ah on turn I pot it and everyone folds. $400 up after this hand.

Second major pot UTG pot KKJTdd to $75, four callers. Flop QJ9cc, I pot, one caller. Turn Kc. I check, she puts me close to all in with $295 bet. I'm getting the right price to hit the boat, so I go all in for $30 more, she calls, 9 on the river, I win a $1825 pot. $1200 up or so after this hand.

Third major pot raise to $75 with AKKQ get one caller. Flop QQJ two hearts. Bet 50, get check raised to 150, call. Turn 8d, check check. River 4h he bets $325, I tank call and he has the straight. Possibly a bad call, and obviously I'm only beating a bluff - I just didn't buy he had it, and I was wrong. $700 up after this pot.

Fourth major pot I call on the button vs MP with KQJ9hh (in hindsight, I should have 3bet). The board came JhTs4d and it's 3-ways. I have to call one $200 bet and see the turn (stacks were too deep to get it on the flop - obviously I would have with shorter stacks). Turn is a blank and she pots it to $800, I have to let it go. $200 up after this hand.

Fifth major pot lost 100 potting pre with aces and missing in 5-way pot with crazy wet board. $200 down after this hand.

Sixth and final major pot I call IP with AA53 nut spades and, after another player pots it and there are several callers, I repot it to $500 and there are 5(!) callers. I had a gutshot on the flop, but two players get it in and a flush takes it down. This hand perhaps best exemplifies the strategy I was trying to pull off, with the idea of getting several callers pre and then big stacks betting each other out of the hand later. I was live with 4 outs to a gutshot and 2 for top set, which would have been the nuts. So I had perhaps 24% equity in a pot where I'm getting 5 to 1 on my money, a good result.

Finished about $700 down after a great start. Will be back there in two weeks...