Brainwagon OKF Blog
I noticed inconsistencies in the pricing for the DeepSeek V4 Flash model on OpenRouter, finding that different aliases and snapshots had significantly varying costs. I then developed scripts and prototypes to rank and analyze these complex pricing discrepancies to understand the true cost structure.
I analyzed the feasibility of running large, open-weight models like Qwen 3.8 27B on consumer hardware, determining that the model size fits onto a single 24GB graphics card in a quantized state. Following this, I compared the potential hardware costs to the pricing for renting the model via an API service.
I built a command-line tool called 'igor' that acts like a deep research assistant for me. I simply ask it a question, and it searches, extracts, and organizes all the information into a well-structured directory, providing markdown, source files, and an index page for everything I need.
I built a research assistant called `igor` using Claude Code; it answers questions by gathering web sources and generating the response in Markdown, HTML5, and PDF formats. I also detailed the process of using AI skills, like `/wayfinder`, to help define project specifications and manage development.
I updated my blog with several minor technical improvements, including fixing broken permalinks so that old links still work correctly. I also adjusted my analytics setup and modified how post descriptions are generated to remove any 'clickbait' teases.
You might be curious about what happens when a creative industry shakes up—I’m sharing my personal story of being laid off after thirty years at Pixar and reflecting on the massive changes in modern film production. Here is my personal story.
Have you noticed this about AI writing, both in prose and scripts for "informative" videos?
I was talking about my checkers program stalling in local minima and learned that higher-dimensional optimization problems often make these minimums much less problematic. Want to understand why deep learning models can reliably find good solutions?
You won't believe the journey of this old telescope! I finally tackled rebuilding a childhood project using my DIY CNC machine and 3D printer, giving it a modern overhaul while keeping its cool homemade history.
If you need a plumb bob for a DIY project, check out my new design that I printed! You can download the free files right now to print this simple, weighty hexagonal model and take your measurements anywhere.
I've been playing the game _Blue Prince_ on Steam, which is a deep, multi-layered puzzle game while exploring an ever-changing mansion full of different rooms. The Parlor provides a compelling set of logic puzzles of different capabilities. I thought it would be fun to see what AI models could do in trying to solve them.
I'm sharing my frustration about how YouTube's 'Watch Later' list has become an unmanageable archive of forgotten videos, and I show you why it feels impossible to clean out.
Are you curious about whether AI models can truly write production-ready software? I dive into why simple 'one-shot' coding challenges on YouTube don't measure real-world software engineering skills, even when the results look pretty impressive.
Do you find modern media saturation overwhelming? I discuss how 'spoiling' has evolved far beyond just ruining a movie plot, arguing that today it means discouraging the experience entirely. If you’re curious about where Hollywood gets its ideas and what challenges the industry faces right now, read on!
I tested a theory about giving my chess engine extra predictive 'muscle' by training it solely on endgames, but after running extensive diagnostic matches against my best opponent, I found out that nothing significant changed. Want to know why this breakthrough strategy didn't work?
I'm taking a much-needed break from developing my self-taught checkers AI because I've hit a frustrating wall. If you're curious about deep learning or game engine development, this breakdown will explain why I'm putting Polonius aside for now.
If you're curious about how AI learns to make complex strategic guesses without needing perfect rules, you need to read this. I'm going to walk you through analyzing my self-playing checkers engine, Polonius, to pinpoint exactly where its neural network struggles the most.
I detail how I designed, implemented, and tested a new improvement for my checkers AI over just thirty-six hours, giving you insight into advanced game theory concepts like score-bounded MCTS.
I finally reveal what my month of deep research achieved! If you want to know whether my custom AI model actually improved my checkers engine by giving it 'second opinions,' keep reading this detailed analysis.
If you're tired of confusing deals, this story about trying to redeem a simple 'Buy 2 Get 1 Free' coupon at the grocery store might make you feel seen. You won't believe what happened next.
I finally ran my second-opinion relabeling test at full scale, and you need to see how surprising—and complicated—the results are. Read on if you want to dive into the technical details of deep learning with checkers.
I finally have the results of my deep-search 'second opinion' experiment, and the answer landed precisely on a statistical tipping point. Want to know what it means for making Polonius stronger?
I discovered that when working with WSL2, where I store my files matters way more than what physical drive they are on—and you might be surprised by the speed difference between accessing Linux-native paths and mounted Windows drives!
I dive into the reality check for my recent cost estimates, showing you exactly why running a deep search is vastly more complex than I initially thought. If budget forecasting for LLM experiments interests you, keep reading!
I tested my checkers engines on a classic board position to see which one can actually finish a game, because sometimes even the simplest tests surprise you. Do you want to know what happened?
I'm sharing an experiment I designed for my checkers engine, Polonius. If you want to know why improving training data labels is critical to boosting performance, keep reading because this dive into machine learning failure modes will help you understand the limits of self-play.
I built a new text-based virtual world, but instead of using predictable scripts for NPCs, I powered it with LLMs so characters can have real agency and hold complex conversations. If you've ever loved gaming nostalgia or are curious about bleeding-edge AI applications, this post is for you.
If you're curious why my powerful checkers engine is struggling against classical competitors, I need to tell you about an embarrassing discovery: it turns out the positions it was trained on are completely different from the standard test openings. You might want to read this if you love deep dives into AI and game theory!
Do you wonder how my chess engines perform against each other? I explain why Claudius really beats Polonius due to an unexpected mix of hardware limitations and fundamental differences in engine architecture.
I finally built a round-robin testing harness for my checkers engines, and instead of a clear winner, I got an unexpectedly messy table. You might be curious to see what happens when classical AI search meets modern neural networks in a deep dive into game theory.
I'm updating you on a shocking rematch between two checkers engines: my hand-built program and an AI that trained itself. You won't believe the massive strength gain the self-taught engine achieved in just 48 hours!
I built two very different checkers engines—one years in the making and one brand new—and you won't believe they came out to be the exact same strength! You need to read this if you want to see how vastly different development approaches can result in equivalent AI power.
Want to know how I built an AI that teaches itself by playing against itself? I walk you through my journey of creating a self-learning checkers engine and share the results of its performance against other programs.
I’m sharing how I finally cracked the code and made my checkers engine gain a huge amount of strength without touching the search or network. If you want to know what simple structural change can make a massive difference in reinforcement learning, read on!
I finally managed to separate the static site generator code from my main git repository, which is a major victory for keeping things organized—you'll want to see how I tackled this messy technical cleanup.
Are you curious how an AI like my checker engine learns to play by itself? I dive into the process of self-play and explain the complex mechanisms—like 'gating' and 'ratcheting'—that allow it to improve purely through repetition.
If you love tools and antiques, check out the index for my massive 'My Garage Gems' series! I spent weeks documenting everything from flea-market finds to projects in my shop.
I finally put my newly fixed engine, Claudius, through the paces against another top-tier AI, and while the overall score was dead even, I noticed something suspicious: Claudius seems to draw many endgames it should actually be winning. Want to see if we can figure out why?
I’m going to share the story of how I tried to close a massive 636-Elo gap between my neural network and an established chess engine, showing you how tiny gains can feel discouraging when your initial deficit is so huge.
Want to know if I finally broke my checkers engine? You won’t believe how informative these new self-play test results were, even though they prove my current evaluation ceiling is stubborn.
I'm showing you how using large language models—like Claude—as a coding partner isn't just for generating code; it's a surprisingly powerful way to expose my own knowledge gaps and learn entirely new concepts. If you've ever wondered about the synergy between human expertise and AI, read on!
I’m going to challenge the idea that mixing in human game data is necessary for high performance. If you’ve wondered how engines reach superhuman skill using only self-play, I break down the three things we actually need.
Are you curious about how I pushed my neural network checkers engine past its perceived limits? I ran a breakthrough experiment mixing in historical game data, and the results showed that we achieved statistical parity with its best version yet. You're going to want to read this if you care about deep learning applied to board games!
I finally went back to the movie theater for a Star Wars experience and had a lot of fun! If you love the franchise or *The Mandalorian*, I think you will enjoy it; otherwise, I suggest waiting until it streams on Disney+.
I'm diving into the hype surrounding AI development by testing modern tools like Claude Code against engines I built years ago. If you wonder if automation can truly boost performance in complex programming projects, keep reading to see my findings.
Want to see if modern AI agents like Claude can actually revamp old code? I tested them out by having an agent rewrite a massive chunk of outdated C code (from the 1980s) and add proper test cases, and you're going to be surprised at what it achieved.
I'm going to show you how embracing minimalist, simple computing principles can even overhaul an enormous project like this blog, and why excessive complexity should worry you.
I'm showing you how I got my LLM agent to write a biography for me using online sources and my own details. You'll see his impressive background—especially his decades of work at Pixar!
I'm sharing how I moved my blog off WordPress because of security issues, and trust me, switching to a simple static site generator drastically improved both my workflow and cost efficiency—you'll want to check out this method!
I want to tell you about my deep dive into building a custom AI agent called 'agent99.' If you're interested in how these complex, functional agents work, I walk you through my process and what I learned from coding it.