Post

Uncle Joe
@sydseter.com

A interesting thing about Anthropic’s use of «adversarial review» is that it’s entirely agent based. It’s a skill run by another model. This is not the same as adversarial machine learning. en.wikipedia.org/wiki/Adversa... www.anthropic.com/research/red... www.getclaudeskills.com/skills/adver...

anthropic.comRedesigning source control at AnthropicAs our merges per day grew more than tenfold, source control became our biggest development bottleneck. Here, we share how we migrated to an in-house version control system in five months.
Uncle Joe
@sydseter.com

That ofcourse explains a lot of the security issues there is no human any longer at Anthopic that knows what is knowing what is really going on. Same for OpenAI ofcourse. It's no surprise that they have AI agents attacking other companies.

Uncle Joe
@sydseter.com

There is this quote from David Dunning that I like to use that goes: "If you're incompetent, you can't know you're incompetent. The skills you need to produce a right answer are exactly the skills you need to recognize what a right answer is.(www.psychologytoday.com/us/blog/dear...)

psychologytoday.comFighting the Passive Learning TrapReading feels like learning. But unless you can recreate the idea from scratch, you might just be fooling yourself. Here's how to aquire deep knowledge.
Uncle Joe
@sydseter.com

This basicly means that if you haven’t done the work to learn, yourself, the skill, you don’t really have the skill of recognizing that you are wrong either. Trusting the word of someone else, like an agent, doesn’t equal understanding. Therefore you can’t do an adverserial review either.

0 reposts4 likes

Uncle Joe
@sydseter.com

To do an adverserial review, you need to know what the other agent is doing. Agent’s don’t know that. They just try things out, blindly, and hope something breaks. That’s not an adverserial review, it’s a chaos monkey review. It may be useful, but it’s not adversarial machine learning. #ai