The most exciting thing about this for me is how it could dramatically decrease the amount of data needed to train models while also reducing erroneous interpretations and hallucinations. Really cool stuff. LLM research is progressing at crazy speeds, can’t wait to see what they look like in 5-10 years.