Brainwagon

I used an AI assistant to build a parametric 3D design tool for a rolling storage cart that holds IRIS bins, generating cut lists and iterating on geometry until the design was correct.

I created a Formula 1 prediction game for my partner and me, then added an AI opponent named BOT-tas.

I used an AI model to audit the subjective scoring in my prediction game and found several inconsistencies that, when corrected, result in a tie instead of a one-point lead.

I was looking for a compact pen which I could carry around easily, and that would match the size of the small 4.5"x3.25" Caliber Composition notebook that I'm trying to get in the habit of using.

I used Claude Code to help me process images from my Seestar S50 of a supernova in NGC 7331, performing astrometric calibration and photometry to measure its brightness, and found it consistent with a Type Ia supernova.

I gave Claude Code a raw FITS image from my Seestar S50 telescope with minimal instructions, and it wrote a Python script using numpy, OpenCV, and astropy to flatten the background, denoise, stretch the brightness, and even remove a satellite trail, produ…

I gave Claude Code two FITS files of Pluto I'd taken a year apart and asked it to find Pluto and estimate its magnitude.

I found an old video of the moon from 2020 and used Claude to apply lucky imaging, stacking the best frames to recover detail.

A Perry Daniels video about the "Grail Method" of keeping one notebook prompts some thinking about notebooking, the two very different approaches to it I saw from David DiFrancesco and Tom Duff, and where 25 years of blogging succeeds and falls short as a…

I learned about Plastic Fantastic, a project with two 3D printed telescope designs for six-inch mirrors: a standard Newtonian and a binocular version with two six-inch mirrors.

I compare my use of local LLMs to how I use my 3D printers, explaining how they help me quickly prototype code for personal projects and test ideas like evaluating a telescope mirror without committing to long‑term effort.

I searched for Frederick Veio's spectrohelioscope book, compared my copy to an online PDF, scanned its bibliography, and used an AI agent to verify which references are freely available online, finding many via NASA ADS and ETH Library while noting others…

I built a local voice assistant named trashbot that runs on my PC GPU, handling timers, time, and weather without sending any data to the cloud, using ASR, LLM, and TTS models with low latency and barge‑in capability.

I explore a Martin Gardner puzzle where you guess a card from a skewed deck using yes‑no questions, showing that the optimal average number of questions is exactly three via a Huffman code.

I solved a month's set of variant Sudoku puzzles by hand, then used an inexpensive Deep Seek V4 Flash AI model to solve one of them, showing how it extracted the puzzle from an image, built a Python backtracking solver, corrected missed constraints, and p…

I realized my current HP Victus 15L with an RTX 4060 (8 GB VRAM) holds me back from running larger AI models, so I looked at what it would cost to buy a comparable new machine or one with more VRAM, finding prices around 1.5× to 2× what I paid two years a…

I visited the estate sale of the now‑bankrupt Tippett Studio after surviving layoffs at Pixar, hoping to take home a piece of its history.

I used a Python script that combines text and reference images to generate AI images, then attempted to create a Star Wars-themed portrait of myself using various models.

I was expecting a mixture of experts model based upon Qwen 3.8 similar to their 3.6 model (35B parameters, with 3B active).

I tested an old 2014 AMD A4 machine to see if it could run Ollama with a small model, having Claude handle the setup and prompts.

I set up an HP EliteDesk 800 G2 Mini as a Proxmox homelab host, ran Ollama in an LXC container, and measured about 5.8 tokens per second—roughly 2.3× faster than my older AMD A4 box—while adding almost no extra cost since the machine is already on 24/7.

I wanted to run a useful mixture-of-experts model on my HP Victus laptop, and after trying the compact colibri engine I decided to build my own lightweight inference engine for a model that nearly fits my hardware.

I walked through someone else's test of running the dense Qwen 3.8 27B model on an 8GB RTX 4060 by offloading half the work to the CPU, which yielded roughly 4‑5 tokens per second and highlighted where a mixture‑of‑experts approach could improve speed.

I experimented with a newly released free model on openrouter.ai, asking it (via a Gemini‑generated prompt) to code a complete Pac‑Man clone using only HTML, CSS, and vanilla JavaScript.

I ran a 35‑billion‑parameter mixture‑of‑experts model on my modest desktop using a hybrid of GPU VRAM and system RAM, thanks to a new model engine that activates just a few experts per token.

I analyzed a video showing AI agents operating within a persistent, game-like world, and I realized the real appeal isn't the premise, but the technical stack.

I use a combination of paid AI subscriptions, like Anthropic and Google, for intensive coding tasks, but I also enjoy supplementing my workflow with free, local models running on Ollama.

I wrote a simple script to test the true cost of generating images using various AI models hosted on OpenRouter, since the listed price and the actual bill are often different.

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.

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 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 built a command-line tool called 'igor' that acts like a deep research assistant for me.

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.

I updated my blog with several minor technical improvements, including fixing broken permalinks so that old links still work correctly.

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.

Have you noticed this about AI writing, both in prose and scripts for "informative" videos?

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.

If you need a plumb bob for a DIY project, check out my new design that I printed!

You won't believe the journey of this old telescope!

Are you curious about whether AI models can truly write production-ready software?

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.

Do you find modern media saturation overwhelming?

I'm taking a much-needed break from developing my self-taught checkers AI because I've hit a frustrating wall.

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.

If you're curious about how AI learns to make complex strategic guesses without needing perfect rules, you need to read this.

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.

I finally reveal what my month of deep research achieved!

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 ran my second-opinion relabeling test at full scale, and you need to see how surprising—and complicated—the results are.

I finally have the results of my deep-search 'second opinion' experiment, and the answer landed precisely on a statistical tipping point.