good morning to another day on the X app

Agents are great at building agents

First for me: chat generated a diagram unprompted to explain a technique to me. Super helpful direction even if imperfect.

As models get smarter does your ability to inspire become more important?

If anthropic uses ai to evaluate ai safety why can’t I do it too

The Super Bowl is cool but have you tried testing different ai models to see if they are deceptive in complex multi-turn conversations?

“Doing ai research” … “Grok what do you think about what ChatGPT just said?”

Treating Codex like a high-performing team mate works so well. Especially for ai alignment research.

so have any of you published on arxiv before? 🤔

Currently on v1.6 of this thread: do llm’s might llm’s deceive under pressure? If so why? Having fun, learning things (mostly research mistakes). Codex5.3 is really good.

ai is so good at ai

Codex: Here is our work plan for the next 12 days. Also Codex: I am now proceeding with the full implementation. (Done 10 minutes later)

now i get anxiety when my fleet of agents is sitting idle like i am falling behind

Given global demographic trends, maybe a new agent-to-agent ai economy is the easiest way to keep demand growing? [… imagining for a second your p(doom) is low]

The worst part of ai generated replies is the generic chat gpt beats. It leaks through no matter the writing persona. “This isn’t X it’s Y!” … “This is big. Not just X. But Y.” … “This is X. But the real question is Y. That’s what matters.”

Caught myself talking about ai as if it were a species. Upon reflection that seemed, accurate?

Kinda guy who can’t walk his dog without giving codex a huge task first.

codex just absolutely wild these days