Elektrine lite

← Feed

@tao@mathstodon.xyz

Post #3707022

2026-07-10 01:29 UTC

It only occurred to me recently, though, that with modern AI agents it should be a routine matter to migrate this nearly 30-year-old system of files and data to a much more modern and easily maintainable framework. And indeed, after a day of (rather painlessly) working with such an agent, much of this data is now migrated to https://teorth.github.io/tao-web/, with the pages now generated from several YAML files serving as the ground truth, as opposed to each page being separately maintained by hand. In the process multiple inconsistencies with the old human-maintained web sites were discovered. Modern AI still has some tendency to hallucinate, of course, so it is possible that some new errors were also introduced; but upon my inspection it certainly does seem that the error rate is now lower than it was before, and more importantly large-scale corrections are significantly easier to implement. While web page maintainance is perhaps one of the least glamorous or exciting aspects of academic workflow, this type of tedious, routine task seems particularly well suited to modern platforms (such as Github), as well as automated tools (including both modern AI and traditional deterministic scripts). (2/2)

Replies (2)

  • @tao@mathstodon.xyz 2026-07-10 05:02

    With a few more hours of work, the agent was also able to revive two dozen of my old Java applets, written in 1998-2000, porting them from the (now deprecated) Java 1.0 to modern Javascript: https://teorth.github.io/tao-web/applets.html . The agent even added some new features, for instance colorizing my old black-and-white Besicovitch set applet with some prettier color gradients (and sliders): https://teorth.github.io/tao-web/apps/besicovitch-sets.html . The process was quite smooth, with only a very small number of bugs generated. Was quite a good feeling to see my very early attempts at machine-assisted mathematics (way before the technology was there yet) come back to life!

    Open ##4296001

  • @tao@mathstodon.xyz "Modern AI still has some tendency to hallucinate" A better term would be "to confabulate" (https://en.wikipedia.org/wiki/Confabulation) rather than "to hallucinate", but the horse has bolted on this one.

    Open ##4296003