Trusted Loops Become AI’s New Moat — 2026-10-12

The cleanest starting point is the a16z show, which released the race to recursive self-improving AI on October 12th.

Martin Casado spoke with MakerMaker.ai co-founders, Dhaval Ajoda and Owen He, about recursive self-improvement, RSI, and importantly, tried to separate the buzzword from the real thing.

Their distinction matters.

A lot of what gets called recursive self-improvement today is really AI helping automate pieces of AI development.

Writing code, generating tests, tuning prompts, evaluating outputs, designing better harnesses.

That is powerful, but it is not yet the science fiction version where a system broadly understands itself, improves itself, and compounds without outside help.

The a16z framing was more disciplined.

True RSI would mean a closed loop where the system being improved is also doing the improving across models, harnesses, evaluations, and deployment machinery.

That phrase, a closed loop, is the key to the day because when you look across the October 12th data, closed loops are everywhere.

Models are not just trained once and shipped.

Agents are learning how to choose tools.

Enterprises are trying to turn AI into repeatable workflows.

Robotics companies are paying humans to generate physical world training data.

Pharma companies are discovering that the loop between drug behavior, cost, and real world outcome is messier than the clinical trial headline.

And the central question becomes, who controls the loop? Who validates it? And who gets harmed if it starts optimizing the wrong thing? On X on October 12th, Aaron Levy pointed to research on agents choosing their own harnesses and wrote that this must be the eventual architecture of agents.

The cleanest starting point is the a16z show, which released the race to recursive self-improving AI on October 12th.

Martin Casado spoke with MakerMaker.ai co-founders, Dhaval Ajoda and Owen He, about recursive self-improvement, RSI, and importantly, tried to separate the buzzword from the real thing.

Their distinction matters.

A lot of what gets called recursive self-improvement today is really AI helping automate pieces of AI development.

Writing code, generating tests, tuning prompts, evaluating outputs, designing better harnesses.

That is powerful, but it is not yet the science fiction version where a system broadly understands itself, improves itself, and compounds without outside help.

The a16z framing was more disciplined.

True RSI would mean a closed loop where the system being improved is also doing the improving across models, harnesses, evaluations, and deployment machinery.

That phrase, a closed loop, is the key to the day because when you look across the October 12th data, closed loops are everywhere.

Models are not just trained once and shipped.

Agents are learning how to choose tools.

Enterprises are trying to turn AI into repeatable workflows.

Robotics companies are paying humans to generate physical world training data.

Pharma companies are discovering that the loop between drug behavior, cost, and real world outcome is messier than the clinical trial headline.

And the central question becomes, who controls the loop? Who validates it? And who gets harmed if it starts optimizing the wrong thing? On X on October 12th, Aaron Levy pointed to research on agents choosing their own harnesses and wrote that this must be the eventual architecture of agents.