"We’re in an age where we’re being told AI is about to start writing its own software."
The most pregnant sentence in the entire piece.
The most pregnant sentence in the entire piece.
AI writing software will just get me another flavour of the wrong software. AI working out what people want the software to do would be useful.
Maybe we can deliver that wrong software in 100ms instead of 100 days.
[1] https://modeling-languages.com/last-one-code-generator-basic...
Earlier this year, I finished watching "Person of Interest" (maybe not ironically) on Netflix.
So far we've been applying ML/DL to problems like image recognition, which has produced quite useful products. But these things are still far away from Artificial General Intelligence.
And I've been studying AI / AGI research for decades as well. In the early years, there was an emphasis on formal systems, that could logically infer the sidewalk was wet because it rained last night. More recently, we've seen research like Hierarchical Temporal Memory which attempt to address how the human brain processes information.
Anyway... these approaches aren't the only ways forward.
When you think about it, one thing we might want for an AGI is really just a software system that can improve itself (gaining new abilities), with increasingly less and less human supervision.
I'm thinking of a giant codebase, not just a big library, but more on the order of the entirety of Python code in open-source projects. The entire ecosystem. And then adding methods for it to self-improve. For example, scanning through the codebase looking for bad style or other obvious bugs. And then automatically fixing them.
The next step would be to give it the ability to automatically process certain kinds of information. It would scan an input document (file), and try to figure out what it is (HTML, CSV, whatever), and then try to match that to the kind of output which is required. And then try to find transforms in its own codebase which can accomplish this for the sample input / output.
The next steps get more difficult where it is analyzing its own codebase not just for problems, but it understand the meaning. Maybe it can annotate it's own codebase with more meta-information. For example this function foobar() has a string argument, but it isn't just a string, it is supposed to be a street address.
And at some point, getting the system to write code to solve new problems.
I've looked at a bit of research in this particular direction, but didn't really find much.
We have been using machines, build by other machines, to make new machines for the last 200 years.
Otherwise, it just becomes another tool for the programmer, where they're being asked to build ever more complex software for ever more kinds of applications.