Smart Software and 100 Lessons

December 6, 2021

100 Lessons from 1 Year of AI Research” breaks down into learning 8.3 “things” a month or 0.27 “things” a day. What’s interesting is that the list suggests that this learning pace is not cumulative; that is, learning does not appear to slow down. If anything, the list suggests that some “insights” cannot be learned, possibly may never be learned; for instance, lesson 24: “Ensure a strong mastery of foundations.” Yes, master.

Here’s another example. Lesson 9:

Try to work towards significant innovations instead of delta improvements that generate only little or negligible insights.

Does this imply that today’s smart software lacks a broader vision? Incrementalism implies that modifications are situational and their cumulative or system wide implications are not known or understood?

Keep in mind that the essay contains 99 other lessons, and as I worked through the list, three points struck me:

  1. Smart software is single-person centric; that is, what a laborer in the vineyards of artificial intelligence is doing takes place in a bubble. What sets AI apart is that the work product can affect the entire vineyard and maybe the wine industry itself. The best part is that no one knows that this is happening.
  2. Cutting corners and being “cute” play a part — maybe the major role — in smart software development.
  3. Join an AI / ML cabal. There is safety and work if one is part of the in crowd.

Pretty interesting. Now how about a list from someone who has been pitching biased algorithms for smart ad sales for three years. What’s that list look like? Maybe no entries or just one: Do what’s needed to get a bonus and promoted. By the way, I try to learn one “thing” per day. Here’s an example: Dr. Timnit Gebru has quite a bit to teach the AI crowd.

Stephen E Arnold, December 6, 2021

Comments

Comments are closed.

  • Archives

  • Recent Posts

  • Meta