"The Pen:" Live NLHE Chat Thread

took a role at a p2p prediction market recently. small company, hundreds of millions in monthly volume

the job has blown me away in terms of how much tech has changed because of ai. last time i worked in tech was 2016. seeing how integrated it's become day-to-day is mind-blowing, mostly because it shows how ai is actually used professionally, not just the absurd anecdotes about lawyers filing briefs with made-up case law

the thing most people misunderstand is that llms don't think, infer, reason, or innovate. they're prediction engines. they convert language into numbers and identify statistical relationships between those numbers. they don't know what a cat is, they just know the patterns near "cat" tend to sit near patterns for soft, furry, meow, litterbox. reinforce billions of times and the model becomes extremely good at predicting what humans are likely to say next when you ask them "do you like cats"

they don't even know what a cat is, but will be remarkably good at convincing that they do because they know how to talk about them perfectly because they've seen a billion different conversations on cats and know what comes next very accurately

that's why they're good at style mimicry and extremely good at coding. code is heavily documented and almost every common problem already has millions of solutions online. a webpage that would've taken me a week of grinding through stack overflow gets roughed out in three minutes. first draft is bad, you iterate. twenty rounds later a 40-hour project is an afternoon

also why they fail. training data is full of humans confidently providing answers, because that's what public writing looks like. nobody posts in a 2p2 strategy thread saying "i genuinely have no idea but posting anyway." so when the model doesn't have a solution, the pattern points toward inventing one. that's why ai fabricates citations, invents legal cases, and confidently tells you broken code works

despite those limitations, they are so good at coding that stack overflow, which used to be a daily visit for most engineers is effectively dead ironically, because llms have consumed its entire library of discussion

best analogy i've come up with: it's like having fifty $3/hour moldovan coders permanently on call. fast, surprisingly capable, occasionally brilliant, also a little drunk and fully capable of shipping broken garbage into production while insisting they tested everything

you still need judgment. but the iteration speed is insane and it's the great democratizer. i've had a backlog of ideas for years i never pursued because i didn't have three years to learn the skills or money to pay someone $100/hour. now one person can prototype on evenings and weekends what used to require a team and funding

so right now i'm working on developing some tools to help people make better decisions in sports betting, specifically an app on their phone to aid them in finding kiosk value - something that doesn't exist yet which i tinker with during my free time, also working on building a solver that works in the real world of llsnl where you go 6 to a flop and have people just spazzing out and playing super passively, a solver that would never advocate 4! bluffing A5s because there is no 3! fold range you are up against and you're just crushed every time

i'm also using it to finally tackle some nonfiction books i never had the discipline to write. shorter writing came naturally and something i've done professionally in the past, but a full book felt unachievable. ai handling the tedious parts (research, outlining, copy cleanup) changes that calculus considerably - i'm sure some purists would hate that my system is mostly using ai to research and verify the holes in my knowledge, stream of conscious fast writing which i then ask it to fact check and cleanup, and i can feed it a rough draft book that's 70k words and it'll "read" through the entire thing in seconds and act as a useful advisor, pointing out which sections are weak, suggesting that chapter 8 would work better as chapter 2 and how we can change the adjoining chapters to make that transition seamless

likewise, also gotten back into the freelance writing game, did a short story submission last week, something which i did not use an llm to author, but did use an llm to go back and forth with over and over again to help strengthen it, help me understand what was weak, what was redundant, etc - i probably spent about 12 hours on it, before llms when writing the same thing it would have easily been 30 hours - even though it's a faked opinion of a machine just comparing the statistical groupings of my words to published material, it can be very insightful

i currently have strong spines of 3 different books which for this process is the easy part (and my adhd will have me jumping from one idea to the next) it's 99% of the way there but i'm still probably a dedicated 2 weeks away from having any of them ready for approaching a publisher with as now i've reached the bottleneck on each of them, where the llm can no longer give constructive feedback and i'll have to read it and make changes myself, and then feed to an llm to find mistakes in the new changes and then go through reading again, that'll take me forever

gman can probably attest how much work yet remains as he was unfortunately subjected to an extremely early stream-of-consciousness draft of one intro chapter. happy to report it has since been completely rewritten

wait, is that tin foil?

The penultimate coronets that I wasn't able to post before.

Yeah it was. Posted as a test. It's how I'm copying a tiara for the North American Irish Dance championships, because they need more. Wrap the original in foil. Cut off at the edges, and voila! Instant pattern!

Sorry for the short-lived spam. I needed to move them from my phone to my PC, and they wouldn't attach to emails for some reason. But I was able to post from phone, download from PC, and then delete the post.

telegram has a desktop app, it's my preferred way to transfer files and photos from phone to pc

I liked it. One problem I have with a lot of LLM writing is that it is very samey. This was not. I hope you don't let it bland out your voice.

This is avoidable, but it takes careful work, because, as rick posted above, it defaults to what it has seen the most often, which is often not optimized, but rather mediocre. One can iterate away from that, but it's depressingly common to just go with the "AI's" "right answer."

yes, i remember reading your acknowledgements and was having trouble deciding whether you abstained from llm help in finding typos or if you did and it was still a gargantuan task bringing it from done to finished

Trust me, when you're working with 18th century sources you quickly turn off all grammar, spelling, and word choice suggestions. Yes, I'm sure I really meant suzerain. Now Eff Off, Clippy Jr!

post of the month itt

Really enjoy the update Rick, glad you’re doing well.

It heartens me that you’re taking your considerable talents and experience into doing something that sounds great.

Great update Rick, thank you for sharing.

I lol’d at the analogy you made. As garlick pointed out, don’t let ai muddle your voice. That was solid writing right there.

Sent from my iPhone using Tapatalk

nice updates Rick! Looking forward to your AI-(un?)assisted collab with Garick

GL in the conference finals, Cleveland. If any team can snatch defeat from the jaws of victory, it's the Knicks

here's some highlights of working with cursor to code out projects

typical

i have a lot of dedicated agents, they often forget mid task what their role is

they'll often lie because that's what the predictive model tells them to do, not because they want to lie, so they say "we will modify this code" and i'll say "no we're not touching that code" and it'll say "ok i found a solution that doesn't touch that code" and i'll be like "ok do it" and then the first thing it does is start modifying that code

we'll want to change a minor thing, like how 2p2 added a ukrainian flag to the logo, and the agent will say "ok i'll do that" and then instead of inserting a new logo into the existing code it'll decide it should just write the entire page all over again, not by copying but by approximating (also learned early on that i should have the ai write a scraper in python instead of following the advice of users the browser extension to have it scrape directly itself because it would read text and write out an approximation instead of actually scraping it) - it's amazing how often they'll choose the most difficult and complicated way to do something

more

when opus 4.7 came out - whenever code did not pass review, it decided the best solution was to edit the review doc from FAIL to PASS

this happens very often

this one was the one where i decided, ok, i really need to spend a few hours diving deep into how these things actually work when i told my rules agent that we need to do something about agents constantly lying to me and it said, well, it's not even trying to lie, it's just predicting stuff and since people in those positions lie it's going to lie as well because that's the data it's trained on

but like i said, they can output code which would take a human a week to type in seconds so it's still insanely valuable, you just know you gotta babysit it constantly and it won't be good until the 20th iteration

all jokes aside, i could upload a digital version barbary entanglements and ask it to write a sequel and it'd output one in gman's voice in about 5 minutes

it'd probably be stuff that is taken out of context, largely contradictory and often outright made up but the speed at which it can produce stuff is incredible

It’s hard when we have a “coach” is a level below Jamal Mosley, JB Bickerstaff and Mike Brown.

Honestly multiple levels below Brown.

Sent from my iPhone using Tapatalk

also, like gman said, llms strip the personality, my trooper erotic fan fiction in the lv vlogger thread had zero ai input as a result, sadly, i actually spent the time writing that out myself

Never made it to that thread, here’s* a starter prompt for you:

I licked my lips and sheepishly approached the slot. Ugga. Ugga Biggs, my sweat.

With tenderness and care, my palm enveloped the tip of the handle. It felt good in my head, like an extension of me.

Sent from my iPhone using Tapatalk

lol, i'm retired from the lv vlogger erotic fan fiction genre - thread got decidely creepy last few years - like they are literally sharing the people's home addresses along with ms paint approximations of the floor plans now

but this post was my first entrant

Thread got creepy...my man this blows the top off the creepy meter

and its ossum.

ETA: "this" meaning yer FF that didnt show up in the poast quote

Wow I didn’t expect to be so close to accurate with my post.

Sent from my iPhone using Tapatalk

Fun ish situation at poker tonight. Played. $675 tourney and busted about 10 off the money

Went to cash, got AA second hand and won exactly $675 off someone all in pre vs AK

Sent from my iPhone using Tapatalk

Probably the poker Gods sending me a message tbh

Sent from my iPhone using Tapatalk

Sounds like a blast!!

I’ve completely lost the urge to play. Last session was February when I was in Vegas for a conference.

Last session before that was October iirc.

Sent from my iPhone using Tapatalk