The author concludes that the returns from using large language models (LLMs) like LLaVA are diminishing and that throwing more training data and power at them will not result in them achieving full sentience. The author also expresses doubt about repeating the experiment in the future, predicting that the results will become even blander. The author believes that AI was more amusing when it was less sophisticated and that a model that averages all the text on the internet will not produce anything better than average.
Key takeaways:
- The author revisited their NaNoGenMo 2016 entry for NaNoGenMo 2024 using current AI tools to generate a novelisation of the film A.I. Artificial Intelligence.
- The process involved ripping the DVD to a file, using ffmpeg to create stills, and using LLaVA 1.5 to generate text based on the images.
- The resulting text was more coherent and novel-like, but occasionally reverted to image describing despite the author's insistence on not mentioning the image in the prompt.
- The author believes that we're in the diminishing returns stage of large language models and that throwing more training data and power at them won't result in full sentience.