LLM Talk


How & what

AI is a what machine. Normal programs are How machines.

If you know how to solve a problem, you can simply write the program.

Why do we use the AI?

because we don't know how to write the program

A general steps of ML model

When doing security analysis, we can decompose the programs and analyze each part separately

LLM basics LLM is just, given a sentence of words, predicting the next words

ELIZA affects

Using keywords matching to do psychology. (this is a work in 1964, 60 years ago)

Foundation models

  • undocumented, unstable API
  • unanticipated behavior

LLM risks

we don't know how and we cannot clean the ocean all we know is the more data we feed, it's possible to have some interesting behaviors

black box risks Bad eval data: because we don't know how to evaluate the model, we don't know whether is a good model or a bad model


Here one of his examples of Sponge input is, putting Chinese characters inside English sentences. These data will caused way more cycles for models to parse.

take away