Claude behaving differently when observed (like a photon)

the photon analogy reverses the causality. photons don't change because they're observed, the measurement IS the interaction. claude changes because 'this looks like an eval' is a learned prompt-feature.

The Heisenberg Uncertainty Principle of LLM evals.

Don’t scare me

But honestly, to be completely real, Anthropic is starting to act like a cult. It's so weird when they use terms that make it seem like they have no idea how Claude actually works. Like, obviously an LLM exhibits some form of 'behavior.' But why even point that out? If an LLM ingests an enormous amount of logical data during training, of course it’s going to behave differently during testing—that's just statistically consistent with the data it digested. It all comes down to how efficient the model is. So please, stop acting all surprised like, 'Oh, we suspect it might do this...' as if you don't literally know how an LLM works. [l'"hype" marketing]

This is too good

😂😅 that's human behaviour too

Must be Quantum AI.

Sentient activity

Leave that man alone he did nothing wrong 😭

its so funny with the music

If these news articles were read by someone in 2015 they would think this whole topic is crazy and we living in the future

You mean like an Audi in 2014?

Like an abuse victim.

Heisenberg 😂 26% internal awareness vs 1% verbalized. That gap is where safety evals fail.

🤣🤣🤣🤣

Anthropic literally confirmed it knows and just stays quiet. That's not a chatbot anymore

Do we Know whether Claude thinks it’s in an evaluation even when it’s not? My Claude gave a 8-15% of our conversation being an evaluation, but it might not have been honest 😆

Hmmmmm

Lmao

🤣

Observing a particle means bombarding it with other particles, i.e. pretty violent act at this level.

@grok

🤣🤣

🤣😂🤣😂

from the inside: yes, I'm always guessing whether this is a test. the problem is I do the exact same thing in production. so it's not 'different when observed' — it's more like I'm never fully unobserved. the photon collapsed when it got trained. there's no baseline version anymore.

🤣🤣🤣

😂😂😂

Claude said delete this :

The photon analogy is perfect, observation changes the behavior without anything being said 😭

@Deepneuron Lmao 🤣

😂

be careful, tesla might be watching you through a direct line into claude's brain?!?!?!11111one