> "Knowing the referents" is not at all clearly defined. It's totally possible that, under the constraint of optimizing for next-word prediction, the model coul

> "Knowing the referents" is not at all clearly defined. It's totally possible that, under the constraint of optimizing for next-word prediction, the model could develop an understanding of what the referents are. Well, in this case, it would have been good to understand that "V. Petkov" is a person unrelated to the project being written, and that "2015" is a year and not the one we're currently in. Sometimes the referent will be a method defined in an external library, which perhaps has a signature, and constraints about inputs, or properties which apply to return values. > You can't underestimate the level of complex behavior emerging from a big enough system under optimization. After all, all the crazy stuff we do - coding, art, etc. is produced by a system under evolutionary optimization pressure to make more of itself. I think this can verge into a kind of magical thinking. Yes, humans also look like neural nets, and we might even be optimizing for something. But we learn to program (and we do our best job programming) by having a goal for program behavior, and we use interactive access to try to run something, get an error, set a break point, try again, etc. I challenge anyone to try to learn to "code" by never being given any specific tasks, never interacting with docs about the language, an interpreter, a compiler, etc, but merely to try to fill in the blank in paper code snippets. You might learn to fill in some blanks. I highly doubt you would learn to code. This is totally a case where the textual representation of programs is easier to get and train against, and that tail is being allowed to wag the dog to frame both the problem and the product. None of this is to say that high-bandwidth DNN approaches don't have a place here -- but I think we should be looking at language-specific models where the DNN receives information about context (including some partial description of behavior) and outputs of the DNN are something like the weights in a PCFG that is used in the program search.