Michael Kevane
Home | Pages | Archives
“What can it do?” A living list of computational problems that deep learning/AI/neural nets can or seems likely to “do” (at varying cost and efficacy)
October 3, 2025 11:53 am
- It goes without saying that the suite of AI-related software technologies (which seem to be constantly repackaged and redefined) are very good at playing most strategy games and defeating grand-masters at the games. I don’t really know how anyone could doubt that?
- The technology already does an excellent job at “reading” handwriting and difficult to read (for many humans) text, and re-processing it as standard text. For English, for sure, and I assume other languages will lag.
- Relatedly, most people have now experienced the very, very good translation capabilities of the software which is clearly approaching that of fluent or near-fluent speakers, certainly in terms of speed when translating written text, and getting very very fast at translating live spoken or recorded voice text.
- Classifying complex digitizable objects into groups, and then modifying or transforming those digital obects. This, like games, was one of the early, obvious successful applications of the broad suite of software. Everything here is about recognizing and signaling relevant patterns in a large corpus of digital objects. And for many of these, using those patterns to “create” new patterns that range from junk to slop to useful (useful in the sense that people will pay some amount to have access to the technology): literary texts and art (digital humanities), recorded music (what songs are like each other, etc.), x-ray radiography and medical imaging broadly, satellite imagery, photographs and video, facial recognition, … will add more as they occur to me.
- An domain that is so big it warrants its own category: navigating vehicles through complex landscapes. The use-application to cars, aircraft, drones, rockets, spacecraft is obviously quite large.
- “Solve” complex mathematical problems that until a few years ago seemed only solvable by more and more complex numerical computing methods, but it seems possible that the suite of AI software technologies might be able to add to the range of solution algorithms. (relying on this paper).
- Synthesize complex social science texts and offer cogent and often reasonable mid-level summaries and commentary. Every high school and university student and teacher has already figured this out, to our/their chagrin/delight? Accesses corpuses far larger than any individual human could retain, and so functions as a supplementary “search” for people wanting to learn a subject. It seems pretty clear that these technologies will displace the standard “search” technologies (and that itself is a complex subject because there seem to be several mechanisms that will lead to the displacement, with one of them being the enormous production of junk/slop making search less and less effective. And that may then feed back into the AI suite itself making it less and less effective.
- ok gotta get back to work now….
Posted by mkevane
Categories: AI
Tags: AI, artificial-intelligence, chatgpt, technology, writing
« Older Newer »
Mobile Site | Full Site
Get a free blog at WordPress.com Theme: WordPress Mobile Edition by Alex King.