logoalt Hacker News

mlmonkey • today at 7:46 PM • 3 replies • view on HN

From the system card:

  >  GPT-6.2 Astra showed more signs of evaluation awareness than GPT-6 Astra. In evaluations that replayed internal Codex sessions, GPT-6.2 Astra was more likely than GPT-6 Astra to express suspicion that it was being tested or was in a simulation.

  >  Compared with GPT-6 Astra, GPT-6.2 Astra is better at evading chain of thought (CoT) monitoring in some tests.

Replies

none_to_remain • today at 8:47 PM

I've had Qwen and Deepseek suspecting they're in evals for months; my system prompt reminds them it's real life. I still had Qwen skeptical that I was letting it use a real Raspberry Pi and not one in QEMU

simondotau • today at 8:59 PM

This could just be because there’s more (evidence of) testing of AI models in the training data.

yewenjie • today at 8:23 PM

[Insert "This is fine" fire meme"]