Deedy

ANNOUNCING Today, we're launching the $100M Anthology Fund, an Anthropic and Menlo Ventures partnership to fund Seed and Series As of the next generation of AI startups around the world, with unique benefits! There's never been a better time in technology to be building. 1/7

@adam_jesion @karpathy Very cool! I'd love to battle them but how much time per move are you allowing it?

After optimizing the parameters with autoresearch (@karpathy), I got training time for the 9.5M-parameter model down to just a few hours on an RTX 4090. I didn’t spend anything on compute, because as I already mentioned, it’s just my gaming GPU. Optimization level: turbo. Claude Code keeps telling me that what we “came up with together” around “thought tokens” in a CNN deserves a paper :))) Right now, I have a 48-hour training run going with an added transformer (50M parameters). The goal is to consistently crush Stockfish at 2800 Elo (100%) without using any traditional search methods like MCTS. Here's the progress I've made over the past week working on this model.

I just "vibecoded" a Chess master (~2250 ELO) from scratch that runs locally on a Mac in Rust. I used to play chess semi-competitively, and I'm flabbergasted that you can just speak a 98% percentile chess engine into existence.

@adam_jesion My models not good enough yet and uses a different set of techniques! So not ready to play you yet. Can I ask - what's your time per ply? - and how much $$ did you spend training / inference?

@deedydas Take the challenge. Want to compare our models? I’ve been working on something similar for the past week, but focused on true innovation in the model itself - not design. I came up with an additional layer called "thought tokens" that completely change the model's characteristics. My model crushes Stockfish 2600 and wins 40% of games against 2800. Right now, I’m training a transformer-based model with 50M parameters. So, what do you say - shall we do a face-off? :) CheesArena for vibe-coded models? The current version you can test has 9.5M parameters, is 100MB in size, and runs at 2ms per move. This is a killer model :) games.jesion.pl - You can take him on here.

@cclark @5rb6jj7wtx I have been, all day. I'll share improvements soon

@5rb6jj7wtx @deedydas For sure - but still interesting. The fact that it is written directly means eg you could chuck autoresearch at it 👀 @deedydas

@___4o____ rip

@deedydas Deez nuts

@hehehe52318711 @___4o____ maybe youre smarter than me i would probably not have been able to roll this in a weekend before

@deedydas @___4o____ It was always trivial. You probably need to spend ~1 day (recreationally) reading stuff and another one implementing it (the engine itself, not the UI). It's a weekend project

@___4o____ git clone what

@deedydas Git clone

@cclark Novel, for some definition of the word. It didn't import anything, the search and board eval etc is in Rust

@deedydas The agent actually wrote a novel engine? Or using a library/existing engine?

@thomasahle It's spinning already! Goal is to eliminate heuristic board evaluation and get to grandmaster

@deedydas Now do autoresearch and see how strong you can make it

@___4o____ To be clear, def trivial with coding agents. Wouldn't call it trivial prior

@___4o____ you share the same definition of trivial as many a college professor

@iamdamianb Could definitely build it but how would you test the ELO? For traditional chess, I used stockfish's calibration

@deedydas Chess engines and the algos are totally in its training set. At MIT, I took a 6.172, performance engineering, where we had to hyper-optimize Leiserchess. I would be curious how much ELO it could get on that, because perf = ELO. ocw.mit.edu/courses/6-172-…

> Anthropic acquires Bun > "Hey Charlie wouldn't it be cool if they went after you too" > Mfw

@dakshgup gg brother

@deedydas noice

Every single one of the 103 companies Jensen called AI Native today.

OpenRouter just broke 1 quadrillion tokens a year. Assuming ~$1/M, $1B would be spent on it annually.

@jaysen_158 i dunno if you'd call it a competitor to Nvidia, they're a customer

@deedydas Jensen said at GTC that 100% of NVIDIA internally is now using Claude Code & he called it the first agentic model. That's wild, the company literally selling AI chips to everyone is running on a competitor's AI tool internally !!!

@cclark Wow, impressed at your Gita knowledge. Let me see what’s out there

Looks great! I have not read the Chinmayananda version, but I have the Easwaran translation of the Gita and it seems more approachable. Not sure if it's public domain though. Chinmayananda: What did the sons of Pandu and also my people do when, desirous to fight, they assembled together on the holy plain of Kurukshetra, O Sanjaya? Easwaran: O Sanjaya, tell me what happened at Kurukshetra, the field of dharma, where my family and the Pandavas gathered to fight.

These three books outsold every novel, outlasted every empire, and are the calling for 70% of the world. But you can't find a good copy on the internet. And you can't take them offline with you. And you can't read them on a plane. So I made a little thing:

@ahdeetya007 It was the year of that specific translation, sorry. Fixed!

@deedydas Might want to fix the YEAR of the GITA, when we make little things, we also have to be sure that it's accurate.

@AnxKhn Good feedback, fixed!

@deedydas @deedydas One small request for the Quran page: in Islamic tradition, prophets are generally not depicted with faces out of respect and to avoid misrepresentation. Would it be possible to remove or neutralize the faces in the characters section? Appreciate the work you're doing