Enabling Frontier Models to learn from Open Ended, Long Horizon Work
Hosted by Manifold Research
Register with organizer
What to look forward to
Joe Hakim, Fig’s frontier data team lead, will discuss why open-ended, real-world work is difficult for models to learn and how Fig is developing new approaches to creating training environments. The talk will explore how each frontier model release expands the work models can complete autonomously.
Inside the event
Join us in welcoming Joe Hakim, team lead for frontier data at Fig, for a live research talk on the intricacies of open-ended data, what makes real work intractable to learn from, and how Fig is developing new approaches to environment creation.
Each frontier model release increases the horizon of work that can be completed autonomously. The environments used to train the next generation of agents must also grow alongside that horizon, and as task length and complexity grow, so do the time and cost of creating them. The open question: “Where do environments calibrated to workflows that span tens of hours, weeks, or months come from?” Fig is developing new scalable approaches that turn real open-ended, multi objective knowledge work into high signal data and environments for frontier AI systems. Our recently launched dataset initiatives like [RealCUA](https://data.fig.inc/realcua) and [RoleCUA](https://data.fig.inc/rolecua) are examples of the kinds of sparse expert trajectories of real work we’re turning into frontier learning signals at unprecedented scale. In this talk, we will cover:
- The evolution of computer-use data: better observability to wider distribution
- The bottleneck: learning signals from open-ended multi-objective work
- Capturing those signals with RealCUA and RoleCUA: what each trajectory contains
- What we observed: continuous sessions, complex workspaces, and interleaved work
- How Fig is developing new approaches to environment creation and turning real work over multi-month horizons into consumable signals at unprecedented scale
The talk will be followed by an open Q&A and discussion.
Links
Explore the RealCUA dataset: data.fig.inc/realcua
Read about release: 600 Hours of Real, Open-Ended Computer Work
Explore opportunities at Manifold: https://www.manifoldrg.com/opportunities/
Join the Advancement Network: https://www.manifoldrg.com/advancement-network/
Subscribe to our Luma calendar for future talks: https://luma.com/manifold
A seat from anywhere.
Register with the organizer for joining instructions and the event link.
Details come from the event listing and the organizer page; registration is managed by the organizer.
Last checked: Nov 20, 2026. Confirm current availability before making plans.
Organizer page checked: Nov 20, 2026. Check registration for the latest details.