7 Comments
User's avatar
Neela 🌶️'s avatar

I have some thoughts.

TALKIE found a strong signal for television in its corpus and extrapolated. It found a contradictory or weak signal for German fascism and defaulted to the optimistic interpretation. That is what typically happens when a language model encounters uncertainty.

It regresses to the mean of its training distribution.

The mean of pre-1931 Western intellectual literature was probably cautiously optimistic about Germany. The model reflected that back. Garbage in, garbage out. I know most tech CEOs think AI models are amazing, but there are always limitations. Just my 2 cents.

I wonder if AI can predict who's going to win the World Cup? hmmmm

Happy Friday, Andrew.

Andrew Sniderman 🕷️'s avatar

Writing this I learned about the temperature (t) parameter. If you dial temperature down to 0 a model becomes practically deterministic — same answer every time. TALKIE is set at .8 which introduces enough variance in response to seem human. Anything over 1.2 becomes gibberish, full of hallucinations.

I brought in my dad’s old typewriter when I showed my HS class TALKIE. Close to the same era. They typed:

There is not enough time to make bread and CO2.

Where are my doll hairs from the nearest applebees?

HS temperature? 2.0

Neela 🌶️'s avatar

.8 is where TALKIE sounded just human enough to be believed. Germany's answer was a confidence-calibration failure. This describes most of the AI deployments I have reviewed in the last three years. I am writing an article about Claude - I will try to finish this week for the following week. You will dig it lol

Happy Monday, Andrew.

Andrew Sniderman 🕷️'s avatar

I dig everything you write; but I am a little Claude pilled :)

Neela 🌶️'s avatar

Oh me too but..... The article may also get me into trouble so I need to get it approved 🫪

Andrew Sniderman 🕷️'s avatar

YOU NEED SOMEONES APPROVAL WHAT IS HAPPENING

Neela 🌶️'s avatar

I know, I know.

Conflict of past interest 😂

Happy Saturday Andrew