ChronoLLM

SN38 ChronoLLM pays miners to train AI models frozen at each year from 2013 to 2024, like a shelf of old encyclopedias, so a finance researcher testing an old strategy consults only what was knowable then.

The commodity, explained from zero

A backtest is how a quantitative investor checks a strategy: run it on past data and see whether it would have made money. If the model reasoning has read tomorrow's newspapers, the backtest lies. The site's line is "Standard models make a backtest look brilliant, then fall apart the moment it trades live." The materials call this lookahead bias.

ChronoLLM's product is language models with a hard knowledge cutoff, each "trained only on data available up to its cutoff date", so the 2018 volume knows nothing of 2019. The unit is a model checkpoint per year: saved weights a researcher downloads and runs. The site offers free download of the full weights for individuals, a free hosted API for academic research, and "use-based" pricing for institutions, with no figure given.

The buyers named are quant researchers, asset managers, and financial institutions. The site says the operator, Crunch, serves "institutional clients including ADIA Lab and the Broad Institute of MIT and Harvard"; it does not say those clients use ChronoLLM.

Why it is on Bittensor at all

On the materials, Bittensor supplies the training labour: instead of one lab training twelve models, miners compete to train them each week. The site names "Subnet 38" only in its header and footer; on what decentralization changes for the buyer on cost, ownership, or censorship, nothing is stated yet. The encyclopedia comparison breaks at the checking step: an old volume's print date is on its spine, but a model's cutoff has to be tested, so validators fire "thousands of post-cutoff leak probes" and a reader can run them too.

How the work gets done

Miners train one model per year for 2013 to 2024, each at most 2.2 billion parameters, upload them to Hugging Face, and register the set on chain. Validators, the judges who decide pay, reject any model whose weights match an earlier entry, then run a leak test: a model must recognise most pre-cutoff items and miss post-cutoff ones, or it is out. Up to thirty survivors fight round-robin duels on fresh prompts judged by a large model (GPT-5.4), inside a trusted execution environment, a sealed part of a processor that proves the code ran untouched. The top ten share emission, the subnet's share of new tokens, with the winner taking about 52 percent and each rank below roughly half the one above; every Monday the slate is wiped and miners resubmit.

How you would know it works

The nearest artifact is the arXiv paper "Instruction Tuning Chronologically Consistent Language Models", which describes the method and its leak tests. The leaderboard at leaderboard.chronollm.com could not be read by a fetcher. No dated benchmark is on the site.

What is missing

The miner guide gives no registration cost and no hardware beyond "GPU". Institutional pricing has no number and there is no API reference, only a short Python example behind a request form. The chain contact address appears on no page. Discord is a channel in the shared Bittensor server. There is no status page or SLA. Third-party pages say the subnet is in a testing phase; its own pages do not say.

Go deeper

• Front page, tiers, and FAQ
• README: the weekly evaluation pipeline
• The paper on arXiv

Metaphor: a shelf of encyclopedias, one frozen at each year. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.