> that test reasoning and explanation in a non-automated way > early experiments that take a much more interactive approach > we're exploring strategies like

> that test reasoning and explanation in a non-automated way > early experiments that take a much more interactive approach > we're exploring strategies like peer grading Are you guys working on any genetic / learning algorithm implementations of AI sufficient enough to evaluate responses to arbitrary prompts? It sounds both like something that should be feasible with modern tech, especially when time to response can be large (a student can wait hours for an answer to a complex essay response) and the scale can be arbitrary (as many server clusters as "its worth" yet also feasible (we can see similar scale literary processing in other spaces). I can imagine such an algorithmic implementation where you take some common query given to students across hundreds of schools, with all the teacher graded responses, and maybe after a few years it can get within an accuracy confidence threshold to start grading the prompt, and after another few years providing critique on that grade. That would be the most revolutionary educational tech since the Internet itself, even if it were pigeonholed in several years studying the same single problem for now. Sounds awesome either way. You can even use that peer grading tech to build up a library of problem / solutions to build your causal relations table from. I just hope you guys have someone in a back room working on this, because to me it seems as real world usable computational linguistics happen more often, there arise a ton of useful applications even in the minimum case.