hazard

I built 1e4.ai - a chess web app where you play against neural networks trained to mimic human Lichess players at specific Elo ranges. There's a separate model for each 100-point rating bucket from ~800 to 2200+, and the bots not only choose human-like moves but also burn clock time, play worse under time pressure, and blunder in human-like ways.

Live demo: https://1e4.ai Code: https://github.com/thomasj02/1e4_ai

A few things that might be interesting:

- Trained on almost a full year of Lichess blitz games, around 1B total games

- Architecture is an a small (~9MM parameters) transformer-based network that takes the board, recent move history, the player's rating, and remaining clock time as input. Three separate models per rating bucket: move, clock-usage, and win probability. The clock model is what makes the bots feel humanish under time pressure rather than instant. Because the move model takes the clock as one input parameter, it also learns to blunder under time pressure like a human might.

- Because the network is so tiny, no GPU is needed for inference - it runs easily on a local CPU

- Downside of the tiny network is that it's a bit weak as you turn up the rating past around 1700. It can spot short tactics but not long multi-move combinations.

- Initial training on a rented 8xH100 cluster, then fine-tunes on my local GPU for different rating ranges

- Inspired by Maia-2 and DeepMind's "Grandmaster-Level Chess Without Search". On a held-out Lichess blitz benchmark, the it beats Maia-2 blitz on top-1 move prediction (56.7% vs 52.7%) and pretty substantially on win-probability calibration (Brier 0.176 vs 0.272). Numbers and code in https://github.com/thomasj02/1e4_ai/tree/master/experiments/...

- The data pipeline is C++ via nanobind, then training with Pytorch. Getting this right was actually the thing I spent the most time on. Pre-shuffling the dataset and then being able to read the shuffled dataset sequentially at training time kept the GPU utilization high. Without this it spent a huge percentage of time on I/O while the GPU sat idle.

Happy to answer questions about the rating-conditioning, the clock model, or the data pipeline.

While exploring new applications for local LLMs, I built a Chrome extension that filters your Twitter feed based on natural language instructions.

For instance, you can instruct it to "Hide all tweets, except for tweets about machine learning (ML), artificial intelligence (AI) and large language models (LLMs)."

I've tested it and got good results with a 10B parameter model, but I suspect a high-quality small model like Phi-2 might work almost as well.

It's open source and available at https://github.com/thomasj02/AiFilter

Video demo: https://www.youtube.com/watch?v=CligVVTC5io

I'd like to learn GPU programming but I'm having difficulty finding high-quality resources. I tried a class at coursera and was severely disappointed by both quality and content.

What are the best resources for learning things like GPU architecture, CUDA, Triton, etc?

My goal is to do be able to do something like take a description of Flash Attention and implement it from scratch, or optimize existing CUDA code.

Our middle-schooler is currently at a private school that we loved when we enrolled him. However in recent years it's gotten steadily worse, primarily due to high leadership turnover and faculty/leadership conflict.

We're thinking about homeschooling him for the remainder of grade school and then re-enrolling him into the traditional school system once he reaches high school. We're concerned however that we may not have the patience or pedagogical experience to effectively teach him.

Does anyone else have experience trying to homeschool? Success / failure stories, or best practices?

Recently I set my nine year old up with an old Linux desktop as his first real computer. Of course he is set up in the living room, no headphones, and with a kids Google account that I can fully monitor.

Right now a pain point is trying to set time limits on websites (like youtubekids or lego.com) because although I don't object to them per se, I also don't want him to spend hours browsing the lego website, watching videos, etc.

On iOS devices there are good tools that allow granular level time blocking of websites, apps, etc. Does anyone know of any tools that allow even basic password-protected time blocking on desktop browsers, or a cross-platform (iOS/Linux) time limiter?

I'm coming to the end of a non-compete and I'm going to start looking for a new position in the next couple of months. I'm planning on using a recruiter as part of my job search, since the industry I'm looking in is rather fragmented.

I get one or two cold emails from tech recruiters on LinkedIn every week, often from different recruiters at the same couple of agencies.

How do I choose a good recruiter? What do you look for as red flags or positive signals?

I'm a senior dev who's mostly programmed in C++ and Python, and I'm now looking at learning Rust. It's been about a decade since I learned a new language, and I'm wondering if anyone has tips beyond "read the docs, write lots of code."

Has there been any formal research on new programming language acquisition? Does anyone have a specific "new language" process that works particularly well for them?

The web has tons of guides on how to write technical design documents for software, but relatively few examples of great design documents that I could find.

Are there any technical design docs for OSS or proprietary software that are examples of what a good doc should look like? Ideally something that would show how the document evolved over time, but snapshots are fine too.