So it's partially AI, sure, but it's also because you're getting the summaries from what is an extremely stupid AI compared to the "good stuff" that's out there.

On the compute side, you can't load most individual experts into even high-end consumer-grade cards (e.g. RTX 5090) because the weights are simply too big.

On the latency side, you need to ship the activations and wait for the somewhat underpowered cards to actually do the matrix multiplications, then send the results back. This means each token takes hundreds of milliseconds or even more, which is borderline unusable.

Finally if you're running a P2P network, the only way to verify that peers aren't lying to you is to run duplicate calculations - I don't see how SHA-256 or signed model state help here unless I'm missing something (What are you calculating the hash of? What exactly is signing attesting to?).

This is one of the few places where borrowing ideas from cryptocurrencies actually makes sense - you could imagine a system where peers place bonds and forfeit them if they lie about calculations. You can look at projects like Bittensor and others for how these kind of things are currently implemented.

Yes, they say it will be a few milliseconds. This service is squarely targeted at algo trading firms.

I went an ungrad school that was top-5 in engineering. But my experience - and in the experience of other people I've talked to - formal undergrad education was, and always has been, a farce. At best, you learn through working on projects that are meaningful to you and learn "how to be an adult" (and later, you learn how to manage the enormous financial debt you acquired). But more typically, it's pure credentialism - no one cares what your grades were, only what school you graduated from.

The amount of actual learning that goes on from classes is minimal, but somehow we can't shift the overton window away from this silly game of grades that don't measure anything meaningful.

After graduating, I've was asked about my grades exactly twice in my life -- once when I applied to a master's program, and at one job interview (the company had a policy of asking about GPA for anyone who graduated less than 10 years ago).

I'm pro-education but anti-school, and all this nonsense makes me this way even more.

This doesn't make sense to me. "A clearer signal of leverage" implies an objective way to measure software engineering output, which has been the white whale of engineering management for the last 50 years.

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.

Comments URL: https://news.ycombinator.com/item?id=48088819

Points: 14

# Comments: 3

The premise is just false. The parable might be true when comparing, say, lower class vs lower-middle-class, or lower-middle-class to middle class. But the difference between upper class and middle class is not "spending less money." It's a vastly different net worth that comes from inheritance, building / running businesses, investments, etc.

The boots theory focuses on the costs, but the real difference comes from the income & net worth

The language is not a feature!

I'm a C++ programmer, and even I don't care what language the applications I use are written in.

"Fast" is a feature, not that it's in C++

A transformer-based (but not LLM) chess model that plays like a human. The site right now is very rudimentary - no saving games, reviewing games, etc., just playing.

It uses three models: * A move model for what move to make * A clock model for how long to 'think' (inference takes milliseconds, the thinking time is just emulated based on the output of the clock model) * A winner model that predicts the likelihood of each game outcome (white win / black win / draw). If you've seen eval bars when watching chess games online, this isn't quite the same. It's a percentage based outcome, rather than number of centipawns advantage that the usual eval bars use.

Right now it has a model trained on 1700-1800 rating level games from Lichess. You can turn it up and down past that, but I'm working on training models on a wide variety of other rating ranges.

If you're really into computer chess, this is similar to MAIA, but with some extra models and very slightly higher move prediction accuracy compared to the published results of the MAIA-2 paper