Claude 3 AI catches researchers testing it

I know it's kind of anathema to how these tools work, but I'd love to know what caused it to respond like that. I wonder if it's possible that much like carbon dating we've now poisoned the well, you can't use trivial tests on AI that has been trained on data that mentions those types of tests. Sure, the first set of LLMs were impressive, but what if the result of those LLMs was an explosion in the discussion of how to evaluate LLMs and now we have a set of LLMs trained on.... how to test LLMs?

Every time one of these big models comes out it seems to always be some appetizer story/issue easily digestible for the “media” and with a very nice reach potential.

No, Claude3 is not self-aware, that spot was taken by GPT2 years back. Remember when OpenAI said they couldn’t release the weights and not give much details because it was so powerful it could end humanity? Those were the days..

The self portrait was a bit... pompous.

https://x.com/karinanguyen_/status/1764789887071580657?s=20

> The entire structure would be in constant flux, rotating, morphing, and rearranging itself into novel patterns never before seen, hinting at the unimaginable depth of intelligence operating within.