Maybe you just need better jokes

What an amazing historic screenshot with a GUI for Fixed Limit Holdem solver, where did you get it?

Researchers like Oskari Tammelin, Marc Lanctot, University of Alberta Computer Poker Lab guys and others invented CFR and had created bots for Computer Poker Competition before me.

But they did not create a product that was used by pros to study No-Limit Holdem. Or maybe they did, but I reached out to a bunch of players and researchers and found zero evidence of it.

Who is Punter and how is his solver called?

It was easier for me to work with a small group of customers. At that time I also felt that that approach has higher EV.

He created and released Jesolver (around 2016) as an engine that was compatible with the Piosolver user interface. I think he posted that screenshot in the Pio thread at some point. Yours might be the first complete product to be sold.

Nice

Can't bat 1000, but you miss every pitch you don't swing at

Piosolver his 2+2 nick is punter with some numbers after it.

<3

Agree with the second part. First part, i will take it as a joke, you must be ret.arded to think the game is way better now how is played

Didn’t Jason mo brag that his group had the best software at some point? If you think Doug never used any analytic tools i don’t know what to tell you.

Doug wasn't an extroverted dude bro until he did the very stereotypical 'poker pro in 2010' thing of getting heavy into PUA culture.

Thats why hes so toxic now, because hes still that 13 year old goob deep down inside

lol great analysis of Doug

what were the hardware limitations like in 2013? piosolver trees were / are ram heavy, were there lots of betszing limitations back then?

did you approach husng guys?

feel like they would have been natutal customers?

any stories of people turning you down and trying to get in later?

I wonder who bought the Omaha version. Tollerene? Berri??

The pool is definitely more polarized today, where bad regulars are being forced out and the remaining are becoming better and better in obvious trainable spots like defending vs flop cbets and so on.

OP's clientes were already top players around that time, lots of people who got a solver didn't know what the **** they were doing for a long time until coaches came to teach them how to study, CFPs and now GTO trainers.

Anyway, I can't see how any aspiring reg back then or any bad reg who improved thanks to the technology and his own efforts could ever complain. Only the big winners who lost their huge advantages would have any reason.

I was running solver on cr1.8xlarge instances at AWS with 244 GB RAM (my product was cloud-based from day 1). Realistically, the only betsizing limitation was the ability of a player to learn a strategy with 5 different sizings.

The idea of approaching HU SNG guys didn't occur to me, and I probably missed out on significant EV. Since I was playing cash games myself, cash players were at the top of my mind.

RaulGonzalez fully ignored my initial DM and and an attempt to reach him via a common contact, as he didn't believe it is possible to have a solution for NLHE and suspected a scam. But when Alex reached out to him, Raul joined the group. Since then, he has become the most frequent user of my solvers.

doug is not an extrovert. he gets far too nervous in social situations and it makes him uncomfortable. Doug is an introvert acting like an extrovert because he feels thats how someone like him is supposed to act.

he got married, I doubt he was ever into actually picking up women

No answer for this one, op?

That's v. interesting Oleg. Ty for posting. It helps explain how the game changed over that period.

- How close do you think you came to NE (allowing for any bet size)?

- How, if at all, did your solution output materially differ from today's solvers?

- Did your PLO solver use the same iterative strategy?

actually doug was doing terrible until he built a model that is some sort of super rudimentary solver

not really a solver though but some approximation of a balanced strategy and deviated from that strategy when needed

At the moment I'm looking for the next big project.

In the meantime, I'm publishing my work on solvers, and learning about LLMs and Code Generation.

Doug and his team beat that AI back in 2015. It is kind of surprising to know a private solver existed two years prior to this as it wasn't until till 2017 that Noam Brown and his team beat the humans heads up. You can be sure all the guys involved in that competition back in 2015 got access to tools as part of the deal to play the AI.

Incidentally Noam Brown now works for OpenAI (ChatGTP).

Super-hard question. As far as I know, exploitability in a game allowing any bet size has been only calculated in toy games. It is infeasible to do in full NLHE, the game is too big.

The good part was that you could play quickly against the solution really fast (check out https://holdem.olegsolvers.com to get an idea. It was more basic in 2013 but still fast).

The bad part is that we had a chart with hands (one you see on the screenshot with an innovative turn lead line), and a separate chart with hands categories. Modern strategy display with hand matrix is so much better.


Yes. I used algorithm called Monte-Carlo Counterfactual Regret Minimization (by Marc Lanctot and co-authors) for both Holdem and Omaha.

It wasn't a solver, it was a game tree visualizer, where he could put his HH in, and see his opponent's frequency in every node of the game tree. He used the old Janda 1-alpha to approximate GTO on every street, which was wrong, but better than what most people were doing. He would see how far they deviated from 1-alpha and then craft his game to exploit those leaks. He talks a lot about it in his old HU course, and shows the software in detail.

Do you still develop solvers? It seems like there is a gap in the market right now, with GTO Wizard using the Ruse AI, which uses neural networks and what not to solve, so it's much faster. The gap in the market is that Piosolver still uses their CFR algorithm, which is much slower. People don't want to pay a monthly sub, so something as basic as a backend for Piosolver that uses NN and machine learning would be nice.

I couldn't remember the details but just watched this interview with Noam that I had seen before.

He says in 2015 they had attempted to have have the entire tree solved with a lookup table. This obviously was less than optimal and led to lots of sub optimal approximations. I think the human players at the time also figured this out and were able to exploit the bot in specific spots. In 2017 they focused on approximating the Nash Equilibrium along with using search algorithms in real time during the hand.

Edit - Worth sharing this point he makes too as it seems rather profound in someway. When he talks in that interview about how even the likes of AlphaZero in chess can't achieve superhuman performance with just a neural net and need the addition of some kind of search. He equates this to human intuition when they look at a board and the search as when the human then gets thinking time and can come up with a better move.

A quick clarification about Noam's bot: In 2015, they had the whole tree solved, and used a lookup table to make moves instantly.

By 2017, they still had the whole tree solved as a baseline strategy, but added real-time subgame solving. They used a 15-second (on average) calculation per move to get more precise solution for the current spot. It improved performance massively, they allowed many sizings, especially on the river spots, but it was still a limited amount.

Hope that clears things up!