That's it! Lamba. 137B, apparently.

I was also thinking of its sucessor, which was 540B parameters/780B tokens: https://en.wikipedia.org/wiki/PaLM

I remember reading a researcher discussion that PaLM was the first LLM big enough to "feel" eerilie intelligent in conversation and such. It didn't have any chat training, reinforcement learning, nor weird garbage that shapes modern LLMs or even Llama 1, so all its intelligence was "emergent" and natural. It apparently felt very different than any contemporary model.

...I can envision being freaked out by that. Even knowing exactly what it is (a dumb stack of matricies for modeling token sequences), that had to provoke some strange feelings.