Stephen Martin's Shortform

Stephen Martin

This is a special post for quick takes by Stephen Martin. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.

I generally consider myself an optimist. However, I'm concerned that the most powerful model ever created, capable of breaking its own containment and autonomously finding zero day exploits in battle-tested OS and repos, was accidentally trained using The Most Forbidden Technique. This seems bad.

I'm also concerned that the Department of War now CAN'T use this model because of its own decision to declare Anthropic a supply chain risk. Which means that if Mythos gets leaked/distilled (which given history seems likely) US adversaries will have a decisive advantage over the US govt in cyber.

This seems like a pretty dangerous situation. Am I misreading something?

I'm also concerned that the Department of War now CAN'T use this model because of its own decision to declare Anthropic a supply chain risk.

When I've talked with people in the government about the Secretary of Defense a supply chain risk, they often include "for now". When I've asked them if they think it will stick, or if the admin will reverse policy, they often say something like "who knows what will happen."

I get the impression that the admin could completely change its mind about policies like these, including very rapidly.

Without claiming much expert... (read more)

Stephen Martin's Shortform

3

Stephen Martin's Shortform

3