Here's a story of how I created the world’s first No-Limit Holdem poker solver and made $500k by age 23
I had to keep d
Enjoyed watching that interview. Perhaps an interesting fact considering that I was apparently one of the very first people in the world to have access to a solver is that I did very little HU solver analysis in my whole career. By the time I got access to Oleg's solver, I could only really get HU action from sauce and ike (jungle wasn't playing on stars at that point in time and this was before Doug rose right to the top so he'd taken a few losing shots at me but didn't really play me). I knew sauce and his dad entered bots in the computer bot competition and I had an inkling that ike might have a solver or at least be developing one so I thought if I used it to work on HU, I'd just be in a solver battle vs sauce and ike, which didn't sound too appealing. I haven't had confirmation either way from them about it, but it sounds like they may have not had access to a solver at that point in time, so maybe it would have been worth my while. I knew about the academic papers posted on solvers though so I assumed some of the other top guys must be doing something in that direction too but perhaps I overestimated how well people were doing with it.
Anyway, I decided to focus on 6 max and did an absolute ton of work with one other guy looking at a bunch of 6max spots. If I remember right we were paying Oleg for each flop studied to some degree and each calculation took days so we were trying to run the minimum spots initially and spent a lot of time and effort extracting concepts, often to find that there was a flaw in our thinking when some other situation didn't play as we expected. We wrote well over a hundred k words of analysis (maybe much more, I can't remember) in total and definitely did pretty well in that period of time before public solvers came out.
It's probably hard to imagine these days, now that so much coaching material has been published using solvers but it was pretty overwhelming initially. All of a sudden we just had access to huge amounts of information and had to work out what to focus on, how to take concepts from it, how to implement any of it in our games. I remember the first thing I did was decide I was going to work things out in detail to get a full understanding of the game and spent a while trying to work out why KhQs was betting so much more than KsQh on the first flop I looked at etc. Quickly abandoned that approach after writing pages on one flop and then realising there were now 40+ turns to analyse and then thousands of rivers just from that one flop. Nowadays you could probably get as much info from an hour long coaching video as we could get from a month of analysis but there were some fairly crucial early insights we had which definitely gave us a nice edge on the competition for a while.
Here's a story of how I created the world’s first No-Limit Holdem poker solver and made $500k by age 23
I had to keep details secret since 2013, but now you can read how I went from near broke mid-stakes player to making business with Kanu, Trueteller and RaulGonzalez.
Short version below. [COLOR="blue"]Full version here[/COLOR]
My journey began in 2008 when I discovered online poker. I instantly recognized the opportunity and in just 18 months, I progressed from microstakes to regularly playing up to $1000 buy-in games, won $40k, and moved out of my parents' house
What really bothered me though was that I had to grind poker every day.
In one-on-one poker, there's an unbeatable strategy called the Nash Equilibrium. As a math student, proving its existence in poker was easy, but actually calculating it was another matter
Eventually I succeeded!
In January 2013, following months of experimentation and research, I finally developed the program. It approximated the Nash Equilibrium for No-Limit Hold'em from the flop with two active players, assuming a limited amount of bet sizes.
The program started with a random strategy and improved with each iteration as it played against itself, and increased the frequency of more profitable actions, relying on the Monte-Carlo Counterfactual Regret Minimization algorithm
I aimed to sell my software for $50,000 per person to elite players in high-stakes games. These top players couldn't hire coaches because anyone better than them was also a competitor at the tables, making my software a unique learning tool in their arsenal.
Despite being able to verify the optimality of calculated strategies programmatically, I still faced skepticism from potential customers. The significance of the breakthrough led some to outright dismiss my demo offers, suspecting a scam.
Ultimately, it was a specific unconventional move that convinced my first and most important customer, Trueteller
Back in 2012, consensus poker wisdom dictated always checking the turn after a check-call on the flop. Yet, the solver suggested a 11% bet here
recognized that such a unique move could only be produced by a genuine solver, not a set of hardcoded rules or simple heuristics
I also reached out to Alex Millar through a DM here on the twoplustwoforum. He and his group of friends expressed significant interest, noticing solver's deep understanding of blocker cards and a precise balance between bluffs and strong hands
Over several weeks, we negotiated a $200,000 deal. By the end of it, I had only $6,000 left in my account, as I've poured everything I had into developing the software. I was determined not to rush the negotiations, understanding that haste could result in accepting a lower offer
Later that year, I created an Omaha solver, which I eventually sold for $300k, bringing my total earnings for the year to $500k. This was six times more than what I had earned as a poker player the previous year, marking a significant leap in my career and financial well-being.
You can read the full story here:
https://medium.com/@olegostroumov/worlds...
And play a solver demo here: https://holdem.olegsolvers.com/
Twitter where I'm going to post more stories: https://twitter.com/OlegOstroumov
https://twitter.com/Kanupoker/status/173...
https://twitter.com/TTimofey_K/status/17...
Happy to answer your questions.
–Oleg
That comparison to investment model output is very interesting. I'm actually spending my work time investing these days, using a fairly basic home-made quantitative approach based on academic research. Luckily (sort of) I have a small amount to invest in the grand scheme of things so I get to invest in the stuff which is generally considered too small to be investible. I like it when I see criticism of academic work saying "well of course you got a positive result there, you included all these tiny, uninvestable stocks where everyone knows the results will be really good!"
I can imagine that with more scale you would be trying to refine models and include many more parameters than I do in order to maximise performance though so I can imagine it being very similar indeed to what we were doing with solver output. There were lots of times when we spent a long time puzzling over why certain things were happening or why the results from spot X were so different to a similar seeming spot Y. It was really interesting and rewarding work though, if frustrating at times. I imagine you feel similarly?
I came up through the stakes in the pre-GTO awareness era, never mind pre-solvers, so I reached 25/50 without even knowing what a Nash Equilibrium was. Then I learned about that side of the game and made the breakthrough to being one of the very best HU players in the world by being ahead of the curve in thinking about GTO concepts and trying to apply them the best I could but without having any sort of confirmation as to what was correct. I guess I look back on that time nostalgically as the "best" time for poker but that's no doubt heavily influenced by personal bias.
There was lots more action though because people had very different ideas of what was correct so it was kind of a battle of ideas at the tables and the skillset needed to be one of the best was to think correctly about the ideas and then find ways to implement them successfully while exploiting your opponents mistakes. Memorising ranges was not really my strength or interest so the top players these days no doubt play far more accurately than I ever did or would have done if I continued. I'm sure they think the game is great these days because it suits their strengths, it's all just personal opinion I guess.
I think I was on board with the solver's accuracy pretty quickly, I can't remember all the reasons for that and their weightings but I know I was a bit familiar with the academic research on the topic and I knew Oleg had used that research in the solver. I also imagine it confirmed some of the things that I had been working out independently, similarly to Oleg's description of trueteller and turn leading when the board pairs. Once I got into looking at it and started to see new things and then work out why they were correct theoretically, then confidence is sky high of course. Before this point in time, pretty much everyone had a standard cbet size I think which didn't vary by board texture. Seeing the different sizes, when they were used and working out it had to do with how the ranges interacted with each other was a pretty big learning moment.
