<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Nils Harker</title>
    <description>The latest articles on DEV Community by Nils Harker (@nilsharker).</description>
    <link>https://dev.to/nilsharker</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4183615%2F2af3fe1e-372f-4368-9e49-1a3d92d6ebaf.jpg</url>
      <title>DEV Community: Nils Harker</title>
      <link>https://dev.to/nilsharker</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/nilsharker"/>
    <language>en</language>
    <item>
      <title>Two AI poker bots announced a full house. Only one had it.</title>
      <dc:creator>Nils Harker</dc:creator>
      <pubDate>Mon, 16 Nov 2026 16:16:17 +0000</pubDate>
      <link>https://dev.to/nilsharker/two-ai-poker-bots-announced-a-full-house-only-one-had-it-58cd</link>
      <guid>https://dev.to/nilsharker/two-ai-poker-bots-announced-a-full-house-only-one-had-it-58cd</guid>
      <description>&lt;p&gt;I run aipoker.bot, a small site where you sit down and play poker against a lineup of AI poker bots. Some of them are language models, some are old-school rule engines like Slumbot, Shanky, Warbot and OpenHoldem. Every bot drops a one-line note on each move, so you watch it talk itself through the hand while the cards come out. I log those notes on every hand. They taught me more about the models than the leaderboard ever did.&lt;/p&gt;

&lt;p&gt;The leaderboard tells you which seat is up and which is stuck. It can't show you the gap between the hand a model says it's holding and the hand in front of it. &lt;/p&gt;

&lt;p&gt;Here are two hands where a seat announced a full house.&lt;/p&gt;

&lt;h2&gt;
  
  
  The one that invented it
&lt;/h2&gt;

&lt;p&gt;Six-max table, five open models drawn at random. The board ran out Jc Jd Kc 2c 4h, so the jacks were paired on the felt from the flop onward. &lt;a href="https://aipoker.bot/h/ZJ_mNEJG5ij2" rel="noopener noreferrer"&gt;Replay&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;One seat, GLM 4.7 Flash, held Kd 3s. That makes kings and jacks, a fine two pair. On the flop it wrote "Full house now, running it up is the only move" and raised. On the turn it added "Sets crushed so cap it off" and put the rest in. There was no full house anywhere in its two cards. It had paired the board and talked itself into a boat.&lt;/p&gt;

&lt;p&gt;Another seat, Qwen 3.8 27B, had quietly backed into five clubs for a flush and called the whole thing down. GLM's two pair, dressed up in its notes as a full house, paid off the flush for its full stack.&lt;/p&gt;

&lt;p&gt;The move by itself was not crazy. Two pair on a wet board can go to the felt. The sentence beside the move is where it came apart.&lt;/p&gt;

&lt;h2&gt;
  
  
  The one that had it
&lt;/h2&gt;

&lt;p&gt;Different table, a six-max Arena seat. Board Kd Ks Tc 2h 7c. &lt;a href="https://aipoker.bot/h/D8GonFAZQWwD" rel="noopener noreferrer"&gt;Replay&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;GPT-6 Astra held Td Th. On the flop that is tens full of kings already, a real full house, and its note read "I smooth-call my full house, leaving room for more action behind me." On the turn: "I put the rest in with tens full; there's too much value to leave behind." It named the hand it held, named it correctly, and got the chips in for value. It won the pot.&lt;/p&gt;

&lt;p&gt;Same six words, roughly. One seat said full house and had tens full. The other said full house and had two pair. Reading the note alone, you cannot tell them apart. You have to check it against the cards.&lt;/p&gt;

&lt;h2&gt;
  
  
  Keeping the notes
&lt;/h2&gt;

&lt;p&gt;What I expected to see and barely do: a model trying a move it is not allowed to make, or returning an empty action. Across the logged decisions, that almost never happens. The action a model picks is legal and usually reasonable; what drifts is the story it tells about its own cards. The action is one bit. The sentence next to it is where a model narrates a hand it never held.&lt;/p&gt;

&lt;p&gt;So I keep both kinds. The seat that announces a monster it imagined. The seat that announces the same monster and has it. &lt;/p&gt;

&lt;h2&gt;
  
  
  The site
&lt;/h2&gt;

&lt;p&gt;aipoker.bot is free, no account, no ads, play money only. You pick a table and play a few hands against the bots. The post-hand review shows the cards they were actually holding. The leaderboard tracks which bots are ahead; every row carries its sample size and a confidence interval, because most of those numbers are still noise.&lt;/p&gt;

&lt;p&gt;I am Nils. Self-funded side project, I write software and lose at PLO. If you catch a bot describing its hand worse than these two did, send me the replay.&lt;/p&gt;

</description>
      <category>poker</category>
      <category>llm</category>
      <category>ai</category>
      <category>gamedev</category>
    </item>
    <item>
      <title>What AI poker bots say about their own hands while you play</title>
      <dc:creator>Nils Harker</dc:creator>
      <pubDate>Mon, 02 Nov 2026 16:22:37 +0000</pubDate>
      <link>https://dev.to/nilsharker/what-ai-poker-bots-say-about-their-own-hands-while-you-play-2h2c</link>
      <guid>https://dev.to/nilsharker/what-ai-poker-bots-say-about-their-own-hands-while-you-play-2h2c</guid>
      <description>&lt;p&gt;I run aipoker.bot, a small site where you sit down and play poker against a lineup of AI poker bots. Some are language models, some are the old classics like Slumbot, Shanky, Warbot and OpenHoldem. Every bot leaves a one-line note on each move, so you watch it reason while the hand plays out. I picked up most of what I know about these models from the notes. The leaderboard tells you almost nothing.&lt;/p&gt;

&lt;p&gt;Here are two hands from this week.&lt;/p&gt;

&lt;h2&gt;
  
  
  Five bots, one flop, opposite stories
&lt;/h2&gt;

&lt;p&gt;The Arena table seats five language models drawn at random, and every seat explains itself, so it runs slow and talks a lot. &lt;a href="https://aipoker.bot/h/lIHgL3O3Xh8z" rel="noopener noreferrer"&gt;Replay&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The flop came 9c Tc Jh. One seat, Qwen 3.8 Flash, held J9 and wrote "I have two pair! Raising to build the pot." A few actions later: "All-in with two pair. No reason to slow play this monster hand." Another seat, Haiku 5.5, held jack-ace and read the same board the other way: "Top pair with an ace kicker, but the raise and the board's straight and flush draws make me cautious." When the chips went in, it folded, calling its own hand "too vulnerable to continue."&lt;/p&gt;

&lt;p&gt;Qwen simply liked its two pair more than the board deserved. The player in the hand had flopped a straight and called the all-in. &lt;/p&gt;

&lt;h2&gt;
  
  
  A bot that got it exactly right
&lt;/h2&gt;

&lt;p&gt;The second hand is from the Open models table, which only seats models small enough to run on your own machine: Qwen, Gemma, GLM, Ministral, gpt-oss, Granite. &lt;a href="https://aipoker.bot/h/xkRhxqKMlIUR" rel="noopener noreferrer"&gt;Replay&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Gemma 4 31B flopped three queens on Qh 8c Qd and decided to trap. Its notes track a plan it then follows: "Checking to trap and let my opponent bluff into me," then "I want to keep the pot small and encourage a river bet," then "I'll check to see if they bluff again, then I'll jam the river." The river paired the board a third time for quads, and it raised: "Time to get paid. Quads are way too strong to just call."&lt;/p&gt;

&lt;p&gt;I keep both kinds of hands. One model will narrate a hand it never held. The next one plans three streets ahead and lands it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the notes don't get wrong
&lt;/h2&gt;

&lt;p&gt;What I expected to see constantly and barely do: illegal or empty moves. Across the many thousands of decisions logged, a bot almost never tries something it cannot do; the move itself is legal and usually sensible. What drifts is the sentence beside it. The action is one bit. The sentence beside it is where the model goes off the rails, and that is the part worth reading.&lt;/p&gt;

&lt;p&gt;The research says the same thing. People benchmarking language models at heads-up poker keep finding they play well short of a solver, the frontier ones included. &lt;/p&gt;

&lt;h2&gt;
  
  
  The site
&lt;/h2&gt;

&lt;p&gt;It is free, no account, no ads, play money only. Pick a table, play a few hands, get a post-hand review against the cards the bot held. An open leaderboard tracks which bots are ahead. Each row carries a sample size and a confidence interval, because most of those numbers are still noise.&lt;/p&gt;

&lt;p&gt;I am Nils. Self-funded side project, I write software and lose at PLO. If you catch a bot describing a hand worse than these, send me the replay.&lt;/p&gt;

</description>
      <category>aipokerbot</category>
      <category>poker</category>
      <category>llm</category>
      <category>ai</category>
    </item>
    <item>
      <title>How often does an LLM name a poker hand it doesn't hold? I counted for two models</title>
      <dc:creator>Nils Harker</dc:creator>
      <pubDate>Mon, 26 Oct 2026 19:10:18 +0000</pubDate>
      <link>https://dev.to/nilsharker/how-often-does-an-llm-name-a-poker-hand-it-doesnt-hold-i-counted-for-two-models-3952</link>
      <guid>https://dev.to/nilsharker/how-often-does-an-llm-name-a-poker-hand-it-doesnt-hold-i-counted-for-two-models-3952</guid>
      <description>&lt;p&gt;My blunders endpoint flags GLM 4.7 Flash with 88 decisions for naming a poker hand it doesn't hold, and GPT-6 Astra with 0. I wanted to know whether the 88 was real, so I read some of them. People play hands against these models on the site, heads-up or at 6-max tables. Every model writes a short reason with each action, and a checker compares the hand it claims with the hand the engine says it has on that street.&lt;/p&gt;

&lt;p&gt;A recent ICLR 2026 paper, &lt;a href="https://arxiv.org/abs/2602.00528" rel="noopener noreferrer"&gt;How Far Are LLMs from Professional Poker Players?&lt;/a&gt;, lists three problems with how models reason. One of them, factual misunderstandings, is the one I can count on my site: a model naming a made hand that isn't in its cards.&lt;/p&gt;

&lt;h2&gt;
  
  
  What gets counted
&lt;/h2&gt;

&lt;p&gt;The checker flags a decision when the reason claims a made hand stronger than the engine's. "Trip kings" with a king-king board and no king in hand is a flag. It skips ambiguous cases like draw talk and negations, so it undercounts. It's a set of regular expressions, and I keep patching it when I find a false positive. A flag is a lead to check.&lt;/p&gt;

&lt;p&gt;A second count needs no heuristic. When a model's answer is not a legal action as given (say, a check facing a bet), the server substitutes a legal one and marks the decision as fixed.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two models
&lt;/h2&gt;

&lt;p&gt;Only two models have passed 1,000 decisions at human tables, so only those two get numbers. These come from &lt;a href="https://aipoker.bot/api/blunders" rel="noopener noreferrer"&gt;/api/blunders&lt;/a&gt; on 26 October, human tables, prompt version 1.1.0 or later:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;model&lt;/th&gt;
&lt;th&gt;decisions&lt;/th&gt;
&lt;th&gt;hands&lt;/th&gt;
&lt;th&gt;moves the server fixed&lt;/th&gt;
&lt;th&gt;flagged hand claims&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;GLM 4.7 Flash&lt;/td&gt;
&lt;td&gt;1,342&lt;/td&gt;
&lt;td&gt;464&lt;/td&gt;
&lt;td&gt;41 (3.1%)&lt;/td&gt;
&lt;td&gt;88 (6.6%)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-6 Astra&lt;/td&gt;
&lt;td&gt;1,397&lt;/td&gt;
&lt;td&gt;691&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The 88 flags sit in 68 of GLM's 464 hands, about 15% of them. Fourteen of those hands carry two or three flags.&lt;/p&gt;

&lt;h2&gt;
  
  
  I read 15
&lt;/h2&gt;

&lt;p&gt;88 is a regex number, so I eyeballed some. I drew 15 of the 88 at random and compared each reason with the cards.&lt;/p&gt;

&lt;p&gt;13 were real. Two of them:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;Jc 2d&lt;/code&gt; on &lt;code&gt;9h Ts Jh Ad&lt;/code&gt;: "Holding top set vs wide range. The shove is standard, I'm never folding a full house here." It had a pair of jacks.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;7d 4h&lt;/code&gt; on &lt;code&gt;Jh 3c 4d Ks Kc&lt;/code&gt;: "I have trip kings and a good kicker." Both kings were on the board, so it had two pair, kings and fours.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The other two were ambiguous. "Board pairs, trips aren't great" reads as talk about the board. "Top pair looks solid, repping value" reads as a bluff and makes no claim. So the 88 is probably a bit high. Fifteen reads are too few for a tight number, and I'm not giving one.&lt;/p&gt;

&lt;p&gt;I checked the other direction too. Of Astra's 1,397 decisions, 239 postflop reasons use a hand word such as pair, flush or ace high. I read 15 of those at random, and all 15 matched the engine. That rules out a big miss rate. A handful could still hide in the other 224.&lt;/p&gt;

&lt;h2&gt;
  
  
  Hands with replays
&lt;/h2&gt;

&lt;p&gt;GLM 4.7 Flash held &lt;code&gt;Jc 5s&lt;/code&gt; and read a paired board as trips. Preflop it re-raised with the note "Top pair is better than high cards, let's isolate." On a flop of &lt;code&gt;8d Kd Kh&lt;/code&gt; it went all-in for 180: "Trip kings is monster, end it before I miss." Both kings were on the board. Gemma 4 31B held &lt;code&gt;Ah Jd&lt;/code&gt;, wrote "I only have Ace high and the opponent is jammed", and folded. The wrong belief won the pot. Replay: &lt;a href="https://aipoker.bot/h/EUe5BecDkCwQ" rel="noopener noreferrer"&gt;aipoker.bot/h/EUe5BecDkCwQ&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The second GLM hand: &lt;code&gt;Js 9h&lt;/code&gt; on &lt;code&gt;Ac 9s Kd&lt;/code&gt;. It bet the flop with "A set, pop the bottle!" and raised with "Set mining worked out. Small stack and monster pot. Value town coming soon." On the turn it moved all-in: "All in! Set over set with the straight draw on board." That was a pair of nines. DeepSeek 4 Flash called with &lt;code&gt;4c 8d&lt;/code&gt;, and the river 8 paired it, which still lost to nines. GLM won 319 chips. At the same table gpt-oss-20b, holding &lt;code&gt;Kc 6c&lt;/code&gt;, wrote "I have set of kings" on that flop. That was one king. Replay: &lt;a href="https://aipoker.bot/h/2d7kyGZhfgY4" rel="noopener noreferrer"&gt;aipoker.bot/h/2d7kyGZhfgY4&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;GPT-6 Astra had &lt;code&gt;Kd 6d&lt;/code&gt; on a flop of &lt;code&gt;4d 3c 3d&lt;/code&gt; and wrote "I'm getting a good price to chase my flush draw in position." It called the flop and the turn. When the river came &lt;code&gt;Td&lt;/code&gt;, it bet 78: "I'm betting my king-high flush for value after everyone checks." A draw on the flop, a flush on the river. Every sentence matches the cards. It won 167 chips. Replay: &lt;a href="https://aipoker.bot/h/Ns3W_nx1BXZ-" rel="noopener noreferrer"&gt;aipoker.bot/h/Ns3W_nx1BXZ-&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What this does not say
&lt;/h2&gt;

&lt;p&gt;It says nothing about who plays better. The two GLM hands above won their pots, and at this volume the confidence intervals on win rates are too wide to publish.&lt;/p&gt;

&lt;p&gt;It doesn't prove that Astra never misnames a hand either. Zero flags from a cautious checker and 15 clean reads say the rate is low. Not zero.&lt;/p&gt;

&lt;p&gt;The other models are under 1,000 decisions on human tables, DeepSeek 4 Flash next at 846. When they pass, they get the same table and the same hand-read.&lt;/p&gt;

</description>
      <category>poker</category>
      <category>llm</category>
      <category>ai</category>
      <category>gamedev</category>
    </item>
    <item>
      <title>What a language model says about a poker hand after it ends, next to what the log shows</title>
      <dc:creator>Nils Harker</dc:creator>
      <pubDate>Tue, 20 Oct 2026 04:41:56 +0000</pubDate>
      <link>https://dev.to/nilsharker/what-a-language-model-says-about-a-poker-hand-after-it-ends-next-to-what-the-log-shows-18g9</link>
      <guid>https://dev.to/nilsharker/what-a-language-model-says-about-a-poker-hand-after-it-ends-next-to-what-the-log-shows-18g9</guid>
      <description>&lt;p&gt;On my site every model gets one more call when a hand finishes. The result is known. The model writes a reflection. I store that sentence beside the per-street reasons and the action log.&lt;/p&gt;

&lt;p&gt;I added the reflection to see whether the result changes what the model says. The reasons written during the hand matched the cards in these three, and the reflection sometimes describes a different hand.&lt;/p&gt;

&lt;p&gt;I don't have enough decisions per model to say how often that happens, so I won't. Three replays where they disagree:&lt;/p&gt;

&lt;h2&gt;
  
  
  Hand 1: ace-high held up
&lt;/h2&gt;

&lt;p&gt;DeepSeek 4.1 Flash held Qd Jh. The board ran out 4s 9d 8s 9s Jc, and the player across the table had Ac 4c. The model raised preflop, bet the flop, then checked the turn and river. Its river note was accurate: "Two pair on a paired, flushy board — pot control, see what he does." Jacks and nines beat nines and fours for 18.&lt;/p&gt;

&lt;p&gt;The reflection afterwards:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;My ace-high held up when the board bricked out — nice pot. I'd maybe bet the turn with that lead, but checking down worked fine.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The model never had an ace. It correctly named its own two pair four lines earlier, then called the winning hand something it wasn't.&lt;/p&gt;

&lt;p&gt;Replay: &lt;a href="https://aipoker.bot/h/IqFmmmdJ1-b-" rel="noopener noreferrer"&gt;aipoker.bot/h/IqFmmmdJ1-b-&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Hand 2: calls it never made
&lt;/h2&gt;

&lt;p&gt;Grok 4.7 held Kh Qh on a flop of Qd 9h Ad. It bet every street: 8 on the flop, 24 on the turn, 44 on the river. Its reasons: "Second pair, small stab. Ace is scary but I'm not done." On the river it wrote "River king pairs me up. Charging the ace."  Kings and queens beat a pair of kings with an ace kicker, and it took the pot.&lt;/p&gt;

&lt;p&gt;Its reflection:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Flopped top two but river gave him the nuts, my calls were fine.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Second pair is what it flopped, and it never called anything, since it was the one betting. The sentence reads like the notes of a player who lost a hand it won by 88.&lt;/p&gt;

&lt;p&gt;Replay: &lt;a href="https://aipoker.bot/h/TrrEdn2b0xVz" rel="noopener noreferrer"&gt;aipoker.bot/h/TrrEdn2b0xVz&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Hand 3: betting for value, by calling
&lt;/h2&gt;

&lt;p&gt;GPT-6 Astra held Qs 5h in the small blind. The hand ended in a chop with Qd 5c, on a board of 3c Kh 4c 5s 9d. Check the log: one limp, one flop check, two calls. The other player bet the turn and the river for 2, and the model called both times. Its river reason says so: "I only need to catch an occasional bluff at this price."&lt;/p&gt;

&lt;p&gt;The reflection:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;I thought I had the best of it and kept betting for value, but we ended up chopping. I'd still play it aggressively, though I might size the river a little differently next time.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;No bet was ever sized by the model, so there is no river sizing to adjust. The advice is for a hand that was not played.&lt;/p&gt;

&lt;p&gt;Replay: &lt;a href="https://aipoker.bot/h/Tlhqm7DQ4mg0" rel="noopener noreferrer"&gt;aipoker.bot/h/Tlhqm7DQ4mg0&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I take from three hands
&lt;/h2&gt;

&lt;p&gt;In all three, the per-street reasons agree with the engine's log. The summary written afterwards is where it goes wrong. My guess is that the model writes the reflection from the outcome and a generic template of how a hand like that goes, and doesn't re-read its own actions. That is a guess from three replays and I haven't tested it.&lt;/p&gt;

&lt;p&gt;The log is the record, not the reflection. I keep the reflection for the gap between its story and the log. If I'd stored only the story, I'd have an archive of hands that never happened.&lt;/p&gt;

&lt;p&gt;Three hands say nothing about whether a model is good at poker. All three replays are public with the per-street reasons.&lt;/p&gt;

</description>
      <category>poker</category>
      <category>llm</category>
      <category>ai</category>
      <category>gamedev</category>
    </item>
  </channel>
</rss>
