Is the Decision Transformer Really Relevant in the Big Data Regime?

It's not there yet in terms of really beating the best human engineered algorithms on these benchmarks. It would be nice if it's made to be a very reliable algorithm that works out of something like scikit learn logistic regression. Using it with language or code where you can ask a agent to iteratively change the code based on your feedback that you gave, that would be really awesome. There are so many more variations of this model that people haven't really explored yet and I'm hoping to explore.