ChronoLLM

The first LLM conscious of time. Live on Bittensor Subnet 38.

ChronoLLM-16B pre-training is almost done and was built on Bittensor, by Subnet 38.

2 big updates:
1. ChronoLLM will release the largest Point-In-Time LLM out there, a 16B MoE.
2. Early vintages beat every point-in-time model at every cutoff through 2020. ChronoLLM should own the frontier for LLMs without lookahead.

ChronoLLM-16B is 8x bigger than its predecessor and scores 22% higher.

Our closest comparable is AQR's point-in-time model, built by a $100B asset manager. ChronoLLM-16B has 4x the parameters, 8x larger context window and scores 12.5% higher on the Open LLM benchmark (22% on the 2013 cutoff). For scale, 10% on this benchmark is the gap between a 7B and a 13B Llama: a full model-size tier. And it stays flat across every cutoff from 2013 to 2022. That’s what no lookahead looks like to us.

ChronoLLM-16B is also the first point-in-time model that can reason. Until now, every other point-in-time model scored near random on reasoning benchmarks like MMLU and GSM8K. ChronoLLM doesn't.

We’re here to stay.

You trust a pilot because of the flight record or the hours in the simulator, and predictive AI has neither until the simulator is built honestly, which is the problem ChronoLLM exists to solve.

Watch Jean's (@herelle_jean) full talk from @ExploitSummit in Montreal.