We make financial products more trainable with RL environments built using live markets & real users.
For safe, reliable, and performant financial LLMs.
Would a reasonable person ever guess that a 25 year old energy company could trade at over 1000 P/E?
Would a reasonable AI model?
Teaching LLMs to manage assets means helping them understand how liquidity, volatility, and social factors combine to create unintuitive outcomes in the market.
If you remember the COVID toilet paper shortage, you know how people act when critical resources start to become scarce.
Markets can act the same way: part of our job is to teach LLMs to look beyond the technical fundamentals of a market & understand the human psychology that drives them.
To accomplish this, we evaluate models not based on knowledge, but on intuition.
We reward models for making surprising decisions that lead to positive outcomes.
We reward models for choosing unique data sources to look for invalidation.
We reward models for acknowledging exotic risks.
This results in models that can express trades creatively and find the strange correlations that lead to 25 year old energy companies trading at over 1000 P/E.
American traders had access to trillions of dollars in AI upside since the launch of ChatGPT.
Most didn't see a penny of it.
@uv asks: how can we teach AI to navigate long-horizon macro trades like this?
The hard part is - if you showed an AI model one of these charts and told it to trade, it would cheat by mapping data or news to the trading outcome.
All of this information is already contained in the model.
So to teach AI how to trade long-horizon outcomes, you need to:
1. Enrich these charts with every possible detail
2. Study them to understand their fundamental nature
3. Create new, imaginary charts that are sufficiently realistic for models to learn without cheating.
This is how reinforcement learning can help models navigate real markets - backed by tremendous amounts of human labor!
American traders had access to trillions of dollars in AI upside since the launch of ChatGPT.
Most didn't see a penny of it.
@uv asks: how can we teach AI to navigate long-horizon macro trades like this?
The hard part is - if you showed an AI model one of these charts and told it to trade, it would cheat by mapping data or news to the trading outcome.
All of this information is already contained in the model.
So to teach AI how to trade long-horizon outcomes, you need to:
1. Enrich these charts with every possible detail
2. Study them to understand their fundamental nature
3. Create new, imaginary charts that are sufficiently realistic for models to learn without cheating.
This is how reinforcement learning can help models navigate real markets - backed by tremendous amounts of human labor!
SITUATION EXPLAINED: How much are frontier labs actually spending on training data?
.@SeanZCai: "Frontier labs are spending about $10 to $15 billion per lab on data."
"Really good long horizon tasks go up to $20,000 each. A complete browser-use version of SAP was rumored at $500,000."
"Despite everybody thinking the market is super crowded, we still don't have enough good quality data vendors that actually understand how to deliver product plus services in a way researchers are looking for."
"I have not seen a contract for genuinely good data gets turned down because of budgetary concerns yet."
AI “trainers” for finance are getting up to $25k / day to teach AI applications
Useful application of AI for pod shops and other analysts = instant model updates based on live earnings call transcripts:
“Sinisterra, 30, then walked the class through how to scan transcripts from earnings calls with OpenAI's ChatGPT and Anthropic's Claude to find the most market-moving statements. The machine ran sentiment analysis and translated management’s spoken remarks into numerical spreadsheet inputs to forecast future financials. Participants could see how AI could help streamline some of the most labor-intensive parts of their jobs.”
Wild how fast AI is changing the landscape
Who'd've guessed that UV Labs would end up one of the only legitimate competitors to Jane Street.
I mean we guessed of course but nice to see others starting to guess as well 🤠
So Jane Street is going public because obviously they see the future where the model labs compete directly with them in the market.
The strategic decision is therefore to become a a specialized infrastructure harness for a future frontier model.
Tellingly they point out that the latency constraints mean there is no time for inference at the GPU layer, or agentic tool use at the CPU layer, only reflexive heuristics at the FPGA layer.
@yminsky is trying to fend off future model lab competition by making Jane Street indispensable to a future AGI.
interesting strategy
@swillinger@Shaughnessy119 our agents are already outperforming equivalent algorithms.
not only that, but the people using them have lower churn rates and higher LTV.
we could give Hermes access to our API 🤠
the personal finance product openai launched is actually one of the worst personal finance tools i’ve ever used in my entire life. borderline hazardous for people with low financial literacy to be using
i had chatgpt tell me everything it messed up:
The big news continues for UV Labs: UV has been accepted into the NVIDIA Inception program.
@nvidia has built the infrastructure that the AI era runs on. Being recognized by them, and welcomed into their ecosystem, means a lot to us and we're genuinely looking forward to growing alongside the builders and companies in their network.
#NVIDIAInception
The most valuable use-cases in AI will require bespoke training infrastructure.
We're creating the environments, methodologies, and data that financial firms will need to create incredible AI experiences for their users.
We are grateful to NVIDIA for their support 🙏