[jumploops]

English is terribly imprecise, so it makes sense to use pseudo instructions to improve the bounds/outcome of a language model’s execution.

I do wonder how long hacks like this will be necessary; as it stands, many of these prompting techniques are essentially artificially expanding the input to enhance reasoning ability (increasing tokens, thus increasing chance of success).

[nimbleal]

I hope there are more models trained on more precise inputs going forward. I understand that natural language feels the most futuristic but while it has the lowest barrier to entry it’s not only imprecise but also slow. Visual approaches (for example control nets in stable diffusion, image as input in Chat GPT, though both of these are somewhat bolted on), 2D semi-natural languages all merit further inquiry.

Another (and perhaps the ultimate) possibility is to have some way —- perhaps through simulations —- to directly expose the model to the problem, rather than having a human/natural language intermediary.