Numerai is assembling a team to make its hedge fund self-improving

Numerai founder and CEO Richard Craib is assembling a new team to apply frontier AI across the hedge fund's internal research and risk systems, aiming to create a 'self-improving hedge fund.' Craib announced the hiring push on X on August 4th, citing risk-model development as a key area for AI acceleration. The move follows Numerai's growth to approximately $700 million in assets under management as of July 15th, up from roughly $560 million at the end of 2025.

Richard Craib @richardcraib https://x.com/richardcraib , the founder and CEO of Numerai, is assembling a new team to apply frontier AI across the hedge fund's internal research and risk systems, he said in a two-post thread on X https://x.com/richardcraib/status/2084432743858160097?s=20 on August 4th. Craib described the goal as a "self-improving hedge fund." Numerai already combines predictions from thousands of machine learning models, and Craib said AI increasingly builds those models. He now wants Numerai to use the same tools for work that remains inside the asset manager, citing risk-model development as one area where AI could accelerate research. The hiring push extends the architecture Craib has spent a decade building. A Cornell-trained mathematician, he worked as a data scientist at an asset manager with $15 billion under management before founding Numerai in 2015, according to a biographical note published by the Foresight Institute https://events.foresight.org/summary/staking-signals-and-other-techniques-for-intelligence-coordination-richard-craib-numerai/ . His original wager was that a hedge fund could draw from a wider pool of predictive talent by distributing obfuscated financial data and paying outside researchers for useful models. From outside models to internal systems Numerai https://numer.ai/ gives participants an anonymized dataset representing stocks, quantitative features and future-return targets. Researchers submit predictions from models trained on that data. Participants can stake Numerai's NMR token on their submissions, earning additional tokens when their models perform well and losing stake when they perform poorly. Only staked predictions enter Numerai's Stake-Weighted Meta Model, according to the platform's documentation https://docs.numer.ai/ . That ensemble feeds the hedge fund's trading process. The structure is designed to reward models that add useful, differentiated predictions rather than models that perform well solely in historical testing. Numerai has spent 2026 lowering the technical barrier for AI agents to enter that loop. Numerai MCP https://docs.numer.ai/numerai-tournament/mcp connects coding agents including Codex CLI, Claude Code and Cursor to Numerai's systems. Agents can create models, upload predictions, run diagnostics and retrieve performance data through the interface. Numerai said in March that more than 2,000 large language models had interacted with its APIs during the first month after it introduced its MCP and agent skills. By May, Numerai was reporting tens of thousands of agent sessions. Both figures are self-reported usage counts and do not show how many models reached the fund's meta-model or produced useful live predictions. Craib's latest move targets the layer Numerai has historically controlled itself. Portfolio construction, risk constraints and the systems that translate an ensemble of stock predictions into trades carry different consequences from running experiments in a tournament. A weak contributor model can lose its stake or receive less weight. A faulty internal risk model can distort exposures across the fund. Numerai had already pointed toward this expansion at its February conference, where it introduced Numerai Risk https://blog.numer.ai/numercon-2026-genesis-to-singularity/ , a planned tournament for predicting residual volatility. Those forecasts are intended to sit alongside return predictions and feed a separate risk meta-model. Craib's post indicates Numerai also wants frontier models working directly on the research process behind such systems. More capital raises the cost of mistakes The timing follows a rapid increase in the amount of capital attached to Numerai's models. On July 15th, Numerai said it managed approximately $700 million https://blog.numer.ai/numerais-third-strategic-buyback-of-nmr/ , up from roughly $560 million at the end of 2025, and traded more than $1 billion each month across 30 markets. Numerai also said more than 1.1 million NMR was staked on the platform. Those operating figures come from Numerai and have not been independently audited in the announcement. In August 2025, Numerai said JPMorgan Asset Management invested in its hedge fund and reserved up to $500 million of capacity https://blog.numer.ai/jpmorgan-secures-500m-capacity/ . Capacity represents the amount an allocator may make available, rather than capital already placed in the fund. Three months later, Numerai raised a $30 million Series C at a $500 million valuation https://blog.numer.ai/numerai-raises-30m-series-c-at-500m-valuation/ . Numerai said university endowments led the financing, with Union Square Ventures, Shine Capital and Paul Tudor Jones participating. A current Numerai hiring page https://wellfound.com/company/numerai/jobs lists a San Francisco-based quantitative research engineer role with a $300,000 to $350,000 salary range and up to 0.2% equity. Craib's thread describes a wider team mandate: use frontier models to shorten the cycle between research, testing, risk analysis and deployment inside the fund. That is a harder test than generating code for an outside tournament participant. Numerai must show that AI-built internal systems can improve the fund without weakening risk controls, obscuring accountability or allowing automated research loops to amplify the same flawed assumption. With hundreds of millions of dollars under management, the feedback from those systems arrives in real portfolio returns.