Frontier AI researchers give drop-in remote workers a 1-3 year runway
As seen on the 24/7 Wall St. homepage on September 11, 2026.
AI researchers debate how close we are to recursive self-improvement
- Panel steelmans failure: Moravec's paradox, no continual learning, a persistent sim-to-real gap blocking generalization
- Schulman on distillation from Claude creating an open-weight monoculture, and Chinese labs closing the gap via router proxy data
- Cursor already runs online RL off user acceptance signals every five hours
- Rapid-fire: 1-3 years to drop-in AI remote workers, 2 years to 10x researcher uplift, 3-10 years to beating top experts
Frontier researchers, including Thinking Machines chief scientist John Schulman, put drop-in AI remote workers at one to three years away, which is the timeline underwriting the entire AI capex trade. Their bear case is technical: RL generalizes task length rather than reasoning across domains, even with task completion time doubling every three months per Edgebench.
Continue ReadingShow less
The episode of the Dwarkesh Podcast brings together John Schulman, chief scientist at Thinking Machines and a co-founder of OpenAI who led the RLHF work behind ChatGPT, alongside Beren Millidge, CTO of Zypra, and Charlie O'Neill, head of model training at Base 10. The rapid-fire timeline section puts drop-in AI remote workers one to three years out, 10x researcher uplift at two years, and beating top human experts somewhere between three and ten years.
The panel's bear case is technical rather than political. Moravec's paradox, the absence of continual learning in current systems, and a persistent sim-to-real gap that blocks generalization are all raised as reasons recursive self-improvement could stall. The concern is that reinforcement learning is generalizing task length rather than reasoning across domains, even as task completion time doubles every three months per Edgebench.
Sponsored
_________________________________
What's Your Number...?
Here's a question most people 5y from retirement can't answer: at your current savings rate, how much do you need, and how long will it actually last? A good advisor can put a date on that in a single meeting. SmartAsset's free quiz matches you with up to three fiduciary advisors serving your area, so you can get YOUR retirement number now (sponsor)
__________________________________________
Schulman raises a specific structural risk on the open-weight side: distillation from Claude could be creating a monoculture among open-weight models, concentrating capability in a single lineage. He also flags Chinese labs closing the gap through router proxy data as a dynamic worth tracking separately from the headline model benchmarks.
One concrete data point grounding the near-term timeline: Cursor already runs online reinforcement learning off user acceptance signals every five hours. That kind of closed-loop feedback at production cadence is evidence that the infrastructure for automated AI improvement is already operating in the wild.