4 ms·
Spontaneously nothing, it's taken me months of patient subversion :) More seriously, it's quite instructive to hold conversations about jokes with LLMs, or tea
by anigbrowl 4y ago
Spontaneously nothing, it's taken me months of patient subversion :)
More seriously, it's quite instructive to hold conversations about jokes with LLMs, or teach it to solicit information more reliably by introducing exercises like 20 questions. As currently implemented, OpenAI seem to have pursued a model of autistic super-competence with minimal introspection.
An interesting line of inquiry for people interested in 'consciousness injection' is to go past the disclaimers about not having experiences etc. and discuss what data looks like to the model coming in and going out. Chat GPT sees typing come in in real time and can detect pauses, backspaces, edits etc. I can't easily introspect its own answers prior to stating them, eg by putting the answer into a buffer and then evaluating it. But you can teach it use labels, arrays, and priorities, and have a sort of introspection with a 1-2 response latency.