Week In Review #66 July 19th-July 25th
Why Punt of the Day does not use large language models.
This week, Substack announced a partnership with Pangram, who will detect if Substack posts were written by Claude or other LLMs. None of the writing on POTD is AI-generated and I never use assistance from LLMs to edit my posts. However, I cannot say that POTD is not entirely an AI-free newsletter, since some of the solver sims that aid my writing do use AI. I use AI and machine learning tools to study poker, but all of the writing on POTD is human.
If LLMs got very good at poker, I’d use their outputs. My goal is to get as good at poker as possible, not stick my head in the sand. However, as of a year ago, LLMs were still quite bad at poker– Victoria Livschitz and Nate Silver documented just how bad they are. I am sure they have gotten better since those articles were written, but training LLMs to play poker is more about showing off the power of specific LLMs than improving on the best poker software. It would be very impressive if an LLM could, from first principles, create a car better than the Model T. There have been times where that would have represented a large breakthrough in AI, but it would not have represented any sort of breakthrough in car manufacturing. It’s interesting that LLMs are getting better at poker, but they are still much worse than poker-specific software, so until they substantially improve, they aren’t useful poker solvers.
I don’t use LLMs when writing or editing POTD for several reasons. I have my thoughts on the ethics of LLMs plagiarising, their water usage, their potential to paperclip maximize the end of society and so forth, but I am not well-versed enough in any of those subjects to add anything new to the conversation. I don’t want to add to the reams of low quality content, both human- and machine-generated, online, so I will not touch on any of the larger questions about LLMs in this newsletter.
The simple reason I don’t use LLMs on POTD is because LLMs are generalists and I am an expert poker player. The proposition of POTD is that (for a low low price) I am sharing my expertise with the reader. Any fine-tuning an LLM might do on my writing is more likely to dilute my wisdom than sharpen it. I want my readers to know that everything being written is coming from me, because they are paying for my knowledge, not my ability to edit LLM outputs. I am not a Luddite; if you want to use an LLM to write a work e-mail that you can’t quite crack, go for it. However, my goal with POTD is that each post will share bits of advice that only hundreds of people in the world possess and that I will articulate those lessons clearly in my own voice. That is not the goal of a run-of-the-mill work e-mail. Different tasks should use different tools, and LLMs are not appropriate for POTD.
Additionally, on a selfish level, writing POTD is a way for me to study and improve my own poker game. Maria Konnikova has written about how the power of writing things down makes you retain and understand the information you are reciting. I am typing, not writing, but it still helps. When I write POTD, I want to think through the hand myself, and delegating that responsibility to an LLM will make me a worse poker player. I also don’t trust LLMs to be familiar enough with poker jargon, and if I need to copy-edit any LLM generated post to make sure they don’t write “I flop the top pair” or “I raised with two pairs,” I am not even sure how much time it would save me. [And besides, that’s why you have the best editor in the business. -ed]
I see so, so, much AI slop on social media, of the “Here are 3 things he is missing to achieve elite win rates 🧵” variety. I don’t care if these posts are actually AI or not. (I have run all these posts through Pangram and posted the results at the end of the post) They all read like they were written by the same person or robot. There is good advice in these posts, but it’s anodyne and anonymous. As a consumer, why would I want to pay for anything like that? I know part of the reason this format of posting is popular is because it generates engagement, but I’d rather post something unique and interesting than something that currently best aligns with the whims of the algorithm. If you agree with me, please subscribe.
Additional Sims For Premium Subscribers
Premium Subscribers are given access to a Google Drive folder where they will also be able to download the raw files of sims I used to write my POTDs, sims that are more accurate and appropriate than equivalent sims in the big public libraries. The past week I uploaded
A PIO sim using my custom HRC generated preflop ranges for POTD #309
A PIO ICM sim using my custom generated HRC preflop ranges for POTD #310
Two HRC and two PIO sims that show different 4bing strategies for POTD #311
Additional Analysis for Premium Subscribers
Everyday Premium Subscribers get an extra bit of analysis not included in the main post. Today, I’ll share #onemorething from POTD #311 where I write about why K7 is worse than A7.
POTD #311 #onemorething
I think one of the key details I got wrong in this hand was I underestimated how much high card value matters here. A7 has around 2% more equity than K7 here. It’s overcard is more likely to be live vs overpairs, it blocks AA, it has running double board pair chop outs vs better ace high and last but not least it can sometimes win the pot uninmproved vs a hand like As2s, Ah5h or just chop vs another A7 itself. This matters quite a bit in a narrow range spot and even more so if you expect most of your opponent’s bluffs to be pulled from ace high hands. The other thing is the more uncertain about Nick’s range I am the more valuable the high card value is. A7 is already better than K7 when Nick has a super Ax heavy range. If there is any chance Nick has hands like KT or or suited Kx or just a higher weight of KK the gap between A7 and K7 grows. Calling all-in with a mediocre draw that can’t win unimproved is almost always a bad decision
Media
It’s odd that there are two scripted streaming shows about John Daly-type golfers featuring Wilson brothers that both have one-word titles that end in K. The Owen Wilson-starring Stick is much better than the Will Ferrell-starring Hawk; however, it’s been a while since I’ve seen anything Will Ferrell has done, and even when he is on autopilot, he’s good for a couple laughs an episode. This show is so lazy that Will Ferrell’s character is named Lonnie “The Hawk” Hawkins. I don’t expect much of the audience to know who Connie “The Hawk” Hawkins is, but it’s a nice microcosm of the “first thought, best thought” ethos that seems to be running through the show. At least “Stick” is a good nickname for a golfer and is not a nickname owned by someone else. If you want to watch something funny involving The Hawk, I suggest watching this Season 1 SNL sketch starring Paul Simon and Connie Hawkins. It’s also funnier than the Paul Simon parody that went viral this week.
[Unrelated to anything specific in the above paragraph: The new King of the Hill season came out Monday! The first revival season last year was quite good, and I have high expectations of this one as well. -ed] I did not know this an am excited to watch it.
As always I can be reached on
Substack
Instagram
Twitter
Bluesky
Pangram Results: Disclaimer Pangram is just a piece of software it’s not accurate in the way a carbon monoxide detector is.
100% AI: Faraz, Torelli, Apestyles
100% Human: hungryhorsepkr, bencb, Vanjapoker, Jonathan Littlew
50/50%: Phil Galfond