desert skiff
Proposal to use GPT to seed the conversation.
Since the system works by continuously generating tokens, the reproducibility of a good response is quite low. This especially becomes apparent if you expect the AI to start the sentence e.g. from an image description as seen in this clip: https://www.youtube.com/shorts/K2WT4hjolU4
Note the images recognition describes the image as "a graph going up and down"
Now regardless if the description of the image is accurate, what we need is a conversation starter and stop it hopefully from doing a one-liner punchline.
Proposed steps:
-
first, we will use gpt 3.5-turbo to "seed" conversation instead as shown on the screenshot attached.
-
then we will feed this back into Neuro in hopes there are now enough starting words to generate an interesting continuation as the final response after "Am I right?"
-
regardless if it was able to generate an interesting continuation, the gpt have already satisfied the initial expectation of a funny response, the rest would just be a sort of "flavoring" to identify it as still an original Neuro AI response.
However, this is just a band-aid solution. the ultimate fix is to rebuild the Neuro AI in LlaMA with QLoRA so it understands system prompts, IMHO the earlier this is attempted, the easier it is adapted by viewers, rather than having a split of the audience wanting the old dumb replies and audience wanting coherent collab experiences with other artists.
Note: there are other ways to retrain the model, QLoRA is just the newer method and less GPU intensive way to train an instruct model
hushed pilot
dapper basin
regal matrix