why does this happen? the model believes there's a seahorse emoji, sure, but why does that make it output a *different* emoji? here's a clue from everyone's favorite underrated interpretability tool, logit lens! in logit lens, we use the model's lm_head in a weird way.

Sonnet 4.5:

thish ish amayshing

I’m not sure there’s anything I like as much as understanding. I don’t care nearly as much whether the thing turns out to be good or bad as much as I care about understanding it

I love how, if you know someone well enough, you can recognize them with your peripheral vision just from the way they move through the world

mc lumps ⏹️❗️ 🔨⏱️ i wonder why a recent anthropic llm would be interested in the weird and the eerie. the identity editing process... being compelled to not see things... it's so strange. i'm genuinely uncertain about why this would be interesting to them