
We’re Mnemonic Labs - a neolab training a foundation model for quant trading that learns directly from market activity. In <6 weeks we’ve built our training pipeline and model architecture. Our current training run uses over 25TB of tokens.
Our Story
We met 28 years ago shortly after Videh was born (we’re brothers!). We’re Olympiad finalists (5x USACO combined) and MIT graduates. Kalki made $20M in annual PnL at 22 at Tower Research and Videh was the youngest quant at Citadel at 19. Over the past ten years, we’ve been part of quant trading’s evolution from linear models to trees and now neural networks.
When we were in college the largest datasets in the world were financial markets and internet advertising. We wanted to build the coolest models, so we picked the former. And then internet scale pre-training and RLVR happened and suddenly general purpose language models were being trained on orders of magnitude more data than any individual domain. It seemed inevitable that robotics would have its own ChatGPT moment and we founded Mnemonic to make it happen.
Our robotics research made us realize that no one - including ourselves - was thinking about financial markets correctly. Done right, we can build a model with a scaling law in markets that automatically learns new strategies and trades as it scales.
Our Bets
We’re all-in on the Bitter Lesson - that all hardcoded priors inevitably become constraints. We believe (ironically) that the Bitter Lesson is one prior always worth following.
Text is a universal prior that can represent anything. Where text is the most effective representation, LLMs are most useful - like coding and math. But in spatial and dynamic modalities the most capable models eschew text. AlphaFold learns the geometry of proteins from amino acid sequences and molecular structures. Self-driving cars learn the dynamics of the road from video and driving trajectories. Financial markets are another such domain.
Second, we believe a small, elite team will win. We’ve seen that trading firms are incredible short term optimizers but not built for risky research-heavy bets. New desks need to be profitable in a year or are fired; already profitable desks don’t want to jeopardize their $50M annual bonuses chasing something new.
Trading is “out of distribution” for the foundation labs. Just as Google famously could be the world’s greatest hedge fund if they wanted to, so could the foundation labs. But, like Google, we think they’ll stay out of this market so they can eat everything else instead.
Asks
We love learning from interesting people! Reach out at founders@mnemoniclabs.ai.
Pictured below, us during our brief love affair with robotics.