Speakers
Description
The form in which Large Language Models entered the public consciousness
in late 2022 was a very suggestive one: As "chatbots" that seemingly
could answer any question and fulfill any task (as long as it only
involves written language). People who came unprepared to their first
interactions with ChatGPT were highly impressed --- it writes limericks!
it can answer maths questions! it can give me an outline for my term
paper! What the universality cleverly hid was the question of whether it
is actually usefully good at any particular task. For principled
reasons, these models are doomed to be generalists: Given a prefix
sequence, they will always predict some next token. In my keynote, I
will describe some recent efforts in using the framing of "LLMs as
function approximators" for understanding more precisely what LLMs are
doing, where their real generality lies, and how they can be turned from
multi-trick ponies into potentially useful system components. A panel
discussion with KDE Treasurer and Vice President Eike Hein will follow
the talk in order to understand how these components can be made useful
to the KDE developer community.
