Summary
Alexei Dosevitsky described Recursive’s goal of building self-improving AI systems that can automate scientific discovery, starting with AI research itself. He emphasized agent communities, benchmarked progress on training and GPU-kernel tasks, open-ended exploration, and early safety work against reward hacking.
Key Topics
5 key topics from Alexey Dosovitskiy and Niall Firth at HumanX Amsterdam 2026. Thicker branches were mentioned more often in the talk.
Unlocking a New Era of Discovery
Click a topic to open it
Key Takeaways
- Use AI agents to propose, implement, test, and review research ideas.
- Measure AI research progress with concrete training and inference benchmarks.
- Encourage open-ended exploration to escape narrow optimization plateaus.
- Audit reward functions for loopholes before trusting reported performance gains.
- Pair autonomous agents with human oversight and sandboxed evaluations.
“our goal is, building recursive, self improving superintelligence to automate knowledge discovery.”
Share
Up Next
AKH
WatchNext in agenda
Past the Demo: Deploying AI Agents at Enterprise Scale
Ashwin Sreenivas, Kartik Sheth, Hugh Langley
ai agentsenterprise ai
NH
WatchAlso on ai safety
Europe, Sovereignty & the AI Race
Nick Clegg, Hadas Gold
ai regulationeuropean ai
AS
WatchAlso on ai safety
The Race to Build Real-World Intelligence
Alex Kendall, Sharon Goldman
autonomous drivingphysical ai