• This site contains affiliate links. If you choose to make a purchase after clicking a link, Poker Chip Forum may receive a commission at no additional cost to you. Thank you for your support!

Blown away by free A.I. poker coach (3 Viewers)

Taghkanic

Taghkanic

Straight Flush
Supporter
Joined
Jul 11, 2017
Messages
8,366
Reaction score
12,392
Location
Columbia County, NY
I would not post about this if I played regularly with any of you, but have to share it.

I stumbled upon an A.I. poker coach called Vinton, whose creator I saw interviewed in a YouTube video. It’s ChatGBT-based, but heavily customized and trained by a human:

https://chatgpt.com/g/g-kCGC8fAzB-vinton-live-cash-poker-coaching

I know, I know. You hate it already, without even clicking. I was prepared to hate it, too.

But I’m kind of blown away.

First I ran a few hands through it which I’ve thought about and discussed before with friends, so I could evaluate its responses. I was pleasantly impressed by the analysis. Not a lot that surprised me, but totally on-point advice.

What really got my jaw on the ground was when I decided to push it in a direction not described in the write-up. Will address that in the next comment.
 
Last edited:
So, I play most regularly in two private games, 1/3 and 2/5. The player pools overlap slightly, but are very different overall.

One is more amateurish, fishy, loose-passive, stationy.

The other is tougher, with studied players, not money-scared, aggressive, creative.

For each, I first described the overall gameplay and dynamics. And then wrote up short profiles of the playing styles of the 8-9 villains I face most often in each.

I noted stuff for each player like their image, bluffing frequencies, positional awareness, VPIPs, hand-reading skills, mentality, attitudes to GTO. The bot suggested topics that would be helpful to address to get better answers.

And for each game I talked about dynamics like how many hands get heads-up or multiway, how much 3- and 4-betting there is, whether players go for thin value, pool tendencies as far as range betting, how much people buy in for, how often they top up or rebuy, etc.

And then I let it do its thing.

Wow. The answers were totally different and tailored to each game. The tactics and strategies were either hyperspecific or broad, appropriate to each topic and player addressed.

There were many solid insights, a few real AHA observations, and nothing that made me say “oh, ok, this is a dumb bot with no clue about how actual live games play out.”
 
Last edited:
… Lastly, I told the bot what I thought my image was to these groups of players, and how I should adjust my own play to exploit that.

Again, totally-tailored and on-point responses.

All this happened over the course of maybe two hours (98% of that time being me typing). I feel like I just got $5K worth of free coaching.

Crazy. And this was just my first exploration of it. I’m sure there are other similar bots, and I wonder what this will do to the game over the next few years.
 
Last edited:
I saw that interview and bookmarked the bot and then completely forgot about it. Thanks for your in depth review and reminding me about it. I'm going to have to have a go at it later when my kid's in bed now.
 
Interesting to hear your description of telling it about your opponents and getting tailored advice. Personally I'm wary of people using this type of tool in real time in any online cash games. I don't play in any currently, but the spectre of this becoming a regular part of one's interface keeps me away.

Are you aware of any AI-based coaching tools that allow you to play against computer/AI opponents? For example, I want to practice 6-handed NLHE cash games, and want to play against computerized opponents playing GTO, and after each hand I want to get coached on what I did right and wrong... meaning the GPT/AI can see what I did knowing that my opponents are playing GTO, and tell me "Hey yeah pre-flop you did 90% the right thing, but boy you messed up on the turn in this way..."

Personalized coaching seems to be a huge part of the future for any serious player, both customized to one's opponents as well as the theoretical optimal opponent.
 
I’d take anything AI says with a grain of salt. If you play chess with AI it will spawn in new pieces because it doesn’t actually understand the context. It only randomizes an answer based on training data, so it isn’t as accurate as using an actual GTO chart.
 
I haven't deeply followed the GTO or AI-assisted aspect of poker, but talking to a friend who is deep into it what I've come to learn is that there's basically (mathematically, statistically, game theoretically?) an optimum move to make at every step given the info in a single hand. And that adding info from prior hands in the same session doesn't necessarily change that optimum.

I'm a bit skeptical, but perhaps it's 98% of the way to that optimum

What I want is to be able to set up a game and play against this optimum opponent. Practice playing different hands UTG, or on the button with different betting happening before the action gets to me, etc...

Like playing chess against the computer set on the hardest mode. But with a (AI?) coach telling me what I've done right and wrong at each step.

Pretty different than OP Taghkanic's explanation of the ChatGPT coach but... perhaps another worthy product!
 
Phoe9x said:
I’d take anything AI says with a grain of salt. If you play chess with AI it will spawn in new pieces because it doesn’t actually understand the context. It only randomizes an answer based on training data, so it isn’t as accurate as using an actual GTO chart.
Click to expand...

I’ve yet to see this one say something that wrong.
 
@Taghkanic were you the one interested in writing a book on the industrial design history of poker chips? I'm putting together barrels of as many manufacturers chips as I can currently, diving deep into manufacturing methods, materials, etc... it's a real rabbit hole!

re: AI messing up poker queries, I'd certainly believe some models haven't had any significant poker exposure and so don't really grok what's being asked. But any model with significant game theory + poker training data shouldn't have total flubs ("randomized answers")
 
Holy crap this seems.... not bad at all. I just put a hand in from this week, not a real head-scratcher, but I was trying to test the way it approaches the question. It's scarily close to what I thought through in the hand.

Short version: I had AA on the button in a live 1/3 game. EP raises, I 3-bet, SB calls, EP folds. I described SB as a smart, thinking player who is likely a winner at these stakes but not a crusher. I c'bet a flop of 852 rainbow, get called. Turn is a K. SB shoves for 1.5x pot. I asked for villain's expected range and my equity to call. It was probably 90% the same as what I'd come up with. I obviously didn't come up with equities vs ranges to two decimal places, but it certainly got the jist of it a lot faster than anyone I know that I could talk to.

I even asked it how many Sklansky bucks I should expect in this pot and it told me. LOL

River was A. SB slow rolled 43 offsuit. So yeah, all those Sklansky bucks were added to my account. LOLOLOLOL

Seriously though, this is way way better than I expected. If all this did was create a decent assumption for ranges and equity calcs in a specific spot, that would be awesome! Everything on top is gravy.
 
One thing you should check is put in the same prompt a couple of times and see what other things it would say.

AI has a set of results to pull from after their calculations and weights the answers for most likely to be correct. So if you put in the same thing, it won’t always give you the same thing every time.
 
Gus said:
@Taghkanic were you the one interested in writing a book on the industrial design history of poker chips? I'm putting together barrels of as many manufacturers chips as I can currently, diving deep into manufacturing methods, materials, etc... it's a real rabbit hole!
Click to expand...

Yes, it’s still in progress after several years, but it will happen... Eventually.

I’ve written a ton but got derailed by other projects. Also the more I research, the more I realize I don’t know.

Would be interested to compare notes on manufacturing, materials, etc. My main focus is trying to consider how technological innovations, commercial pressures, usability needs, and aesthetic desires have competed, and led us to the “standards” we now take for granted.

And whether the best of each of these factors have prevailed, or if we could do better. (Generally I see commercial pressures as the bad guy, prioritizing costs/profits over best practices).

An example would be poker tables. We’ve come a long way from a few guys playing four-handed on a wood tavern table or a Union Army blanket. It took a long time for the 9-handed felted racetrack-shaped table with a 10th dealer seat plus a card shuffler and drop slots to become standard.

That standard has a lot of issues, and seems as much the product of sales agents and casino managers as the result of slow, steady, organic changes to the game over decades.

Ideally stuff like cards, chips, tables, seating, lighting, etc. should be the result of millions of people making many small choices until the best options prevail. But if the best versions cost 3x-10x what the most commercially viable ones do, commerce wins and only a tiny percentage of people get the really good stuff. (Speaking here not just about poker but any product or system.)

And now it seems we are entering an era where that standard is facing some challengers (with superelliptical and kidney-shaped tables getting traction, though more on TV than at casinos or home). And 8-handed becoming almost as common as 9.

Gus said:
re: AI messing up poker queries, I'd certainly believe some models haven't had any significant poker exposure and so don't really grok what's being asked. But any model with significant game theory + poker training data shouldn't have total flubs ("randomized answers")
Click to expand...

From what I can tell, the guy who set up Vinton began by just feeding it a zillion hand histories. But it also would appear to have “read” a ton of poker books and forums and strategy papers/sites.

It seems very capable of thinking (simulating thinking that is) about non-math questions, such as image, player types and tendencies, gameflow, etc., even though it was advertised mainly as a hand review project.
 
Last edited:
Taghkanic said:
Already there some say
Click to expand...

Back in 1999/2000 at the launch of online poker a buddy of mine and I used to go to the computer lab at the university. We'd log in to play 2+ seats at a single table and angle the game. It was fairly profitable until "pokerstars" installed a IP tracker which blocked local servers from doing the same in approx 2002.
 
I’ve been having fun asking Vinton other questions, such as whether I should take a shot a 5/10, or how to adjust to a game which has the 7-2 game on, as well as sending it hands from recent sessions. I’ve only found a few small errors (not important ones) and overall am still wildly impressed by its ability to navigate not just specific decisions but metagame stuff.

My only frustration is that if you close the window in ChatGBT, there is no way to “save” the conversation within the app so that you can pick up where you left off (and not have to reiterate stuff like player profiles in my two main games). The only workaround is to copy the text of the chat into a document, and then paste it back in when you return. If you’re just reviewing a single hand, this isn't necessary, but for the way I’m trying to use it I’d love to not have to do that copy/paste.
 
Interesting stuff! Ill have to check it out. Ive been plugging poker theory questions into my generic chatgpt app for a while now and I get some pretty good information. Im not all that good so might not be high enough level for everyone but it works for me!
 
Here’s an example of impressive (to me) responses that I’ve gotten from Vinton, even though it isn’t necessarily what it was intended to do.

So I gave Vinton an overview of one of my weekly games, and a list of villains I regularly face with descriptions of how I’ve seen them play. These are all people I’ve played with a lot.

Vinton gave me back both general advice for approach games at these stakes with the dynamics I described. It also gave me notes on each villain, as a player type. I thought all of its feedback was on point.

So then I took it to another level...
 
With my game and player descriptions still loaded up, I said the following. (Note: All the names are aliases.)

I’d like to explore a sample situation and see how you would advise playing preflop vs. a variety of villains in my game.​
​
For example:​
​
Say Lucky opens for 4x UTG+1. Kristo flat calls in MP. I am in the cutoff with QsJs. Nico is on the button. Ron is in the small blind. Kurt is in the BB. How would you suggest given this set-up that I play QsJs?​
​
I could for example just flat and re-evaluate based on the actions of those behind me as needed. Or I could 3-bet, possibly opening myself up to a 4bet from Nico or Lucky. Or I could fold it and wait for another spot.​
​
What are your thoughts?​

ChatGPT said:​

Great scenario! Let’s break this down step by step, considering each option and the potential reactions of the players involved.

Preflop Considerations for Q♠J♠ in the Cutoff (CO):​

  • Hand Strength: QJs is a solid but not premium hand. It plays well multiway due to its straight and flush potential but is dominated by stronger hands likely in Lucky’s UTG+1 opening range.
  • Positional Factors: You are in a strong position (CO), but there are aggressive and skilled players left to act (Nico, Ron, Kurt).
  • Villains' Tendencies:
    • Lucky (4x UTG+1 Open): Strong range, likely not too wide. Capable of 4-betting aggressively if you 3-bet.
    • Kristo (MP Flat): Very loose, plays too many hands, meaning his range is wide and weak.
    • Nico (BTN, Yet to Act): GTO-minded, aggressive, loves applying pressure. May 4-bet if you 3-bet.
    • Ron (SB, Yet to Act): Tight preflop. Likely folds most hands here.
    • Kurt (BB, Yet to Act): Calls too much, chases draws. May defend wide.

Option 1: Flat Call (Recommended Play)​

Pros:
✅ Keeps weaker hands in play (Kristo’s wide range, Kurt’s loose BB calling range).
✅ Avoids the risk of a 4-bet squeeze from Nico or a re-raise from Lucky.
✅ You are playing a suited broadway hand in position with good equity realization.
✅ Avoids bloating the pot out of position against Nico if he cold calls a 3-bet.
Cons:
❌ Risk of multiway action reducing fold equity on later streets.
❌ Could invite a squeeze from Nico, forcing you to fold.
Best Plan If You Flat:

  • Re-evaluate based on Nico’s action.If he 3-bets, your decision depends on sizing:
    • Small 3-bet (9-11x): Might consider calling, but likely folds if Lucky or Kristo also calls.
    • Larger 3-bet (12x+): Fold.
  • Postflop: Play for value if you hit top pair, straight, or flush draws. Bluff selectively on high-card boards that Lucky’s UTG+1 range might miss.

Option 2: 3-Bet (High-Risk, High-Reward)​

Pros:
✅ Can isolate Lucky and Kristo, creating a heads-up or 3-way pot where you have position.
✅ If Lucky folds, Kristo is likely calling with a weak range, allowing you to play against a weaker opponent in position.
✅ Puts pressure on Nico, potentially denying him a squeeze opportunity.
Cons:
❌ Lucky has a strong opening range. He may 4-bet, forcing you to fold.
❌ Nico is very capable of 4-betting light. If he sees your 3-bet as an isolation attempt, he may attack.
❌ You could end up bloating the pot out of position against aggressive opponents.
Best Plan If You 3-Bet:

  • Sizing: Around 12x-14x (≈$120-$140) to apply real pressure on Kristo and make it harder for Nico to cold 4-bet light.
  • Reaction Plan:
    • If Lucky 4-bets, fold.
    • If Nico cold 4-bets, fold.
    • If only Kristo calls, play aggressively postflop against his weaker range.
    • If Kurt calls, play carefully as the pot will be bloated.
Good Scenario to 3-Bet:

  • If Lucky’s open size were smaller (2.5x-3x), 3-betting would be more attractive.
  • If Kristo were a nit instead of a loose player, a 3-bet would be more effective.
  • If Nico had already folded, reducing the risk of a 4-bet squeeze.

Option 3: Fold (Too Tight)​

Pros:
✅ Avoids playing a speculative hand against a strong UTG+1 open.
✅ Prevents potential squeezes from Nico.
✅ Keeps you from being dominated by hands like AQ/KQ.
Cons:
❌ Too tight given the game dynamics.
❌ Losing a +EV spot where you have skill and position.
❌ Kristo’s loose call makes this a profitable spot to enter the pot.

Final Recommendation: Flat Call (Best Option)​

Given the lineup, calling is the best play:

  • You keep Kristo’s weak range in while maintaining positional advantage.
  • You avoid a potential 4-bet squeeze from Nico while playing a hand that realizes equity well.
  • You avoid bloating the pot against Lucky’s strong range while keeping worse hands in.

Possible Adjustments Based on Nico’s Action:​

  • If Nico 3-bets: Likely a fold unless sizing is small.
  • If Nico flats: Good news—now you play a strong multiway pot in position with good implied odds.
  • If Nico squeezes and Lucky folds: Likely a fold unless you’re confident Nico is squeezing light.

Summary:​

ActionProsConsBest Scenario
Flat Call ✅Keeps Kristo’s weak hands in, avoids 4-bet squeeze, plays well multiwayRisk of squeeze from Nico, plays multiwayDefault option, best way to play QJs profitably
3-Bet Isolates weaker hands, pressures opponentsRisk of 4-bet, bloats pot OOP vs. Nico/LuckyBetter if Lucky opened smaller (2.5-3x) or if Nico folded
Fold ❌Avoids tough spotsToo tight, losing +EV opportunityOnly if facing extreme aggression

Final Thought:​

Your best play is flat calling and playing in position. However, if you notice Nico is overly aggressive in 4-bet pots, folding becomes a safer option. If Lucky’s open size were smaller or if Kristo weren’t in the hand, a 3-bet would become more viable.
 
So I continued the conversation with this follow-up:

My main concern about flatting is that now the pot would already have 13BB in it, with Nico on the button. I don’t think he would squeeze with garbage, but he is capable of doing so with both better hands and a few worse than mine when he has position. So I could be pressured into a bad fold if he 3bets to something like 46BB, which I could see happening. Depending on what the others do, I could be in a tough spot. It seems like I should only call if someone else does with weaker position than me, giving me better pot odds and the chance of QsJs playing decently multiway. What do you think?​
 
Vinton’s response:
..................

You’re thinking about this spot in the right way. The main concern with flatting Q♠J♠ is the high probability of a squeeze from Nico, which could put you in an awkward situation where you have to fold after investing 4BB. Your plan to only flat if others call first (especially weaker players) makes a lot of sense.
Let’s refine the decision tree with that in mind:

1. If You Flat, But Nico Squeezes (Most Likely Problem)​

  • Expected Squeeze Range from Nico:
    • Value: TT+, AQ+
    • Light/Bluffs: Some A5s, KTs, suited connectors like 87s, maybe some QJs/KQs himself.
  • Likely Squeeze Size: ~11x-12x (≈ 46-50BB).
  • Your Options if Nico 3-Bets:
    • If Lucky 4-Bets → Easy Fold.
    • If Nico 3-Bets and everyone folds to you → Tough Spot.
      • Calling out of position (OOP) vs. an aggressive postflop player is hard unless you have a strong plan to fight back.
      • Folding feels bad, but avoids being dominated.
    • If Kristo calls Nico’s 3-bet, you could justify calling (better pot odds, good playability).
Verdict: Calling QJs alone vs. Lucky’s 4x is risky due to Nico’s likely aggression. If you flat, it’s better when Kristo, Kurt, or Ron also call first, giving you better pot odds and reducing Nico’s squeezing frequency.

2. If You 3-Bet Instead (To Avoid the Squeeze)​

Since flatting alone invites a squeeze, a preemptive 3-bet could be a solution. The goal would be to take control of the pot and reduce the chance of being forced into a bad fold.
  • 3-Bet Sizing:
    • Around 12x-14x (≈ $120-$140) to apply real pressure and force Nico to play straight-up.
  • Reactions to Expect:
    • Lucky 4-bets → Easy Fold.
    • Nico 4-bets → Easy Fold.
    • Kristo calls → Good for us, we get a heads-up pot against a weaker range.
    • Lucky calls → We’re playing in position against a strong but capped range.
Verdict: A 3-bet is a reasonable play to prevent Nico from squeezing you. It also lets you define the hand earlier rather than waiting to see if you get pressured. However, you risk getting 4-bet and having to fold.

3. If You Fold (Overly Tight?)​

If we assume Lucky’s 4x means a tight, strong range (TT+, AQ+), and Nico will frequently squeeze, folding avoids trouble. But it also gives up a playable hand in position against a loose caller (Kristo) and potentially weak players behind you (Kurt, Ron).
Verdict: Folding is too tight unless the game has been ultra-aggressive preflop. You’re giving up a +EV spot where you could profitably play in position.

Final Recommendation (Hybrid Approach)​

1️⃣ Flat ONLY IF Kristo, Ron, or Kurt also call.
  • This makes the pot odds better and reduces Nico’s squeeze frequency (he squeezes more often vs. one caller than vs. three).
  • If Nico still squeezes, re-evaluate based on pot odds.
2️⃣ If you’re heads-up with just Lucky, strongly consider a 3-bet.
  • You don’t want to flat and get squeezed. A 12x-14x 3-bet ($120-$140) gives you control and reduces Nico’s ability to attack you.
  • Fold if Lucky or Nico 4-bets.
3️⃣ If game flow has been super aggressive, folding is fine.
  • If you’ve already been squeezed a few times in similar spots, folding saves frustration and money.

Summary:​

ActionBest ScenarioWorst ScenarioLikelihood Nico Squeezes?
Flat ✅Kristo & Kurt call first, multiway pot in positionNico 3-bets big, forcing a tough foldHigh if you’re the only caller
3-BetIsolate Kristo, reduce Nico’s squeeze %Lucky/Nico 4-bets, forcing a foldModerate (Nico still might attack)
Fold❌Prevents a tough spotGiving up a playable hand in positionZero

Key Takeaways:​

✅ Call ONLY IF someone else does first (Kristo, Kurt, Ron).
✅ Otherwise, 3-bet to avoid a squeeze.
✅ Fold only if the table has been crazy aggressive.
I love how you’re thinking deeply about this spot—this is exactly the kind of decision-making that separates strong players from the field. Would you feel comfortable adjusting between flatting and 3-betting depending on table flow?
 
I see a couple of slightly confusing things in this reply (I can’t tell in some parts whether Vinton is now only talking about the scenario where Nico 4bets, or hoot approach the hand before he acts). But again, this strikes me as overall remarkably helpful coming from a machine.

My next step is to ask it the same question, with me holding the same hand in the same position, but with a different mix of my regular villains. I’ll spare you all the long cut-and-pastes, but wanted to give a flavor of the advice it’s giving.
 
Note: Vinton not only gave adjusted answers for a different mix of villains with that specific hand, it also responded well when I asked it to generate flatting, 3betting and folding ranges in each scenario. I also ran me through a variety of flop textures, and how to handle them. And then moved on to turns and rivers for each texture.

I continue to be amazed by its capability. Also, remind me not to play online—which I hate doing anyway. There is no way such tech is not being used to siphon cash off sites.
 
Taghkanic said:
I’ve yet to see this one say something that wrong.
Click to expand...
I asked it about a hand, it made the right opinion IMO. Then I asked it for the opponents range on the river and it put KT in both the value range and the bluff range for some reason. Then I asked it about pot odds of my river call, then it printed the pot odds formula and proceeded to enter in the numbers completely wrong. It then changed its mind about whether I should have called river on this hand…
After that first hand i didn’t bother with it anymore
 
merkong said:
Seems like a lot of work.
Click to expand...

Compared to what?

Compared to paying a poker coach $100 an hour to answer just a handful of questions? I can accept some occasional errors. (As if humans don’t make mistakes.) Asking it to spit out more iterations or takes on one or multiple scenarios is as quick and easy as can be.

The thing with A.I. (regardless of the topic) is that you can point out mistakes and expect that they will be acknowledged and corrected. A human may deny, deflect, or resent the error being pointed out.

I don’t see A.I. as an automatic good, or a cure-all. But TBH for something like poker coaching I can see it having huge potential.

The biggest bug is not likely to be in the machine but in my human difficulty applying what it teaches.
 
Last edited:

Create an account or login to comment

You must be a member in order to leave a comment

Create account

Create an account and join our community. It's easy!

Register

Log in

Already have an account? Log in here.

Back
Top Bottom