KM4BI

Recursive’s $670M Bet on Self-Improving AI, Sonnet 5.5 Hits 70%, and Elon Co-Leads the Pentagon Push — EP 299

Richard Socher discusses Recursive, automated science, model competition, changing work, and the need for human oversight of autonomous weapons.

Published

Recursive AI is moving from theory toward an engineering program. Richard Socher argues that limited forms of recursive improvement already exist and that broader versions are getting close, while still placing artificial superintelligence several decades away.

Building the self-improvement loop

Recursive has announced $670 million in funding alongside $410 million in AWS compute commitments. The aim is to create systems that can accelerate AI research itself: generate ideas, implement them, evaluate the results, and repeat the cycle with progressively less manual coordination.

That approach fits a broader full-stack model of automated science. Language models contribute world knowledge, digitized scientific datasets provide evidence, simulations make rapid experimentation possible, and robotics connects software reasoning to physical validation. Swarms of agents can coordinate the entire idea–implementation–validation loop.

Benchmark gains and model choice

The episode highlights Sonnet 5.5 reportedly moving from 10% to 70% on Terminal Bench 4.0. The panel nevertheless questions when it should be preferred over Opus 5.5 or open models. Cost, latency, control, and deployment constraints may matter as much as a headline benchmark.

This makes the open-versus-closed competition strategically important. Faster releases and specialized models may be a response to pressure from open-source alternatives and international competitors, while enterprises weigh performance against control of their stack.

Work shifts from tasks to outcomes

If AI can handle research, coding, marketing, and project execution, human value moves toward defining the problem, setting constraints, exercising judgment, and owning the outcome. Education faces a difficult transition because judgment has traditionally been learned through the junior tasks that agents are now automating.

Autonomous defense requires a boundary

Project Meridian is presented as a Pentagon effort on future warfare co-led by Elon Musk and Palmer Luckey. The discussion focuses on whether military procurement can adapt to short technology cycles and where human control must remain. Socher’s position is clear: systems that can end human lives should retain meaningful human oversight.

The same technology can compress scientific discovery, expand entrepreneurship, and reshape warfare. The practical challenge is to build institutions and safeguards that can keep pace with the systems they deploy.

Source