thomas

We will soon be passing the point of no return beyond which an AI seeking to ace a Turing test will no longer need to appear as smart as a human, but rather as dumb as a human

PhD in aviation ∩ climate science | Aviator, diver, ETH maxi, nexialist | Host of /diving /ethfinance /eth-staker /geopolitics /science /singapore /thomas

(as in, introduce the occasional and deliberate hesitations, inaccuracies, and outright mistakes to remain plausibly human)

like when google showcased their ai assistant that could call places for you (that went nowhere i'm pretty sure?) but it had programmed uhhhh and ummmms to help sound more human

I feel these are easy ticks to incorporate. While unique combinations of rare yet familiar words and sentence constructions are harder for an LLM to fake.

Hardest to fake: ppl with an ear for emergent slang that has not stuck yet but will eventually stick

reminder to use broken grammar. 

the LLMs won't.

agree

https://warpcast.com/cbxm/0x51096d41

my fave: the fake keyboard clicks when a live call bot pretends to be looking something up