Nof1 is an AI research lab focused on building new models for financial markets, covering everything from pretraining to post-training. Its core idea is that markets are the next important domain for frontier AI research because markets are open-ended, high-stakes, and highly measurable.
[Positioning and Background] Nof1 believes that the current generation of large language models performs poorly when dealing with financial markets: they have read a large amount of financial text, yet they struggle to complete even the simplest investment tasks, and there is almost no obvious progress between versions. One reason the team identifies is that markets are non-stationary, and methods that worked in the past are very likely to fail in the future, which does not match the training paradigms currently used by frontier AI labs for fields such as programming, mathematics, and law. To train frontier models for markets, new techniques are needed that can handle noisy, continuously changing environments.
[Research Direction] The lab also focuses on a pain point in AI research: the lack of good ways to measure AI progress, as new benchmarks are quickly memorized and saturated, and do not evolve along with models. They propose an idea: use markets as both an evolving benchmark and a training environment, to directly address one of the biggest failure modes of modern large language models: adaptability. Current models are too static and backward-looking, making it difficult for them to update or think ahead, while markets provide a natural setting for cultivating this ability and can provide objective feedback.
[Vision for Model Capabilities] According to the official introduction, Nof1 is building new models for financial markets, as well as the environments and benchmarks needed to train and evaluate these models. Its training goal is general market intelligence, and it has already been continuously improving its measurement methods internally, has identified a "prototype scaling law for financial intelligence," and has shown signs of transfer to skills in other domains. After the models are released, the official statement says they will be able to conduct well-calibrated research, analyze large-scale time-series datasets, complete complex feature engineering, write market-aware code, make predictions, and self-improve based on decision outcomes, with the goal of becoming an adaptive reasoner that can be used by hedge funds, individuals, aggregation platforms, brokers, and financial institutions in any framework.
[Application Directions] The officially listed application directions include risk management, portfolio management, systematic trading, backtesting, and simulation.
[Team and Support] The team's work covers open-ended learning, world modeling, and large-scale reinforcement learning, with members from DeepMind, Two Sigma, Citadel, Bridgewater, and Netflix. Research is led by Julian Togelius, founder of the NYU Open-Ended Learning Lab, with more than 20 years of frontier AI research experience. The lab has received support from Village Global as well as several leading AI researchers and founders, including Tim Rocktäschel (DeepMind), Jack Parker-Holder (DeepMind), Oliver Cameron (Odyssey ML), and angel investors from Jane Street.
[Limitations and Information Not Yet Provided] Official materials do not mention a specific charging model, price, launch time, or available product form, nor do they mention any content related to a "real-time ranking competition for live trading investments" or "six major AI cryptocurrency trading competitions." This information is not yet provided.
