2 Comments

Graphics and memes on point as always, Devansh

Expand full comment

Devansh -

Well done! I read the original paper (twice now), and the approach is a better demonstration one of the evolutionary algorithms in continuously self-updating ML models.

Two comments/questions:

1. Why are so few GPTs and GANs designed to use continuously updated training? High fitness in a specific (niche) environment has been known to lead to the extinction of species for over 100 years. The fossil record is replete with examples of this.

2. I suggest anyone who wants to know how humans came to use novel evolutionary algorithms (speaking teleologically, of course) read the terrific book “Sapiens: A Brief History of Humankind” by Yuval Noah Harari (see: https://www.amazon.com/Sapiens-Humankind-Yuval-Noah-Harari-ebook/dp/B00ICN066A/ref=sr_1_1?crid=1BYQJAPW4MQUD&keywords=sapiens+a+brief+history+of+mankind&qid=1701171557&s=digital-text&sprefix=Sapie%2Cdigital-text%2C91&sr=1-1

Homo Sapiens was but one of 17 species in the genus Homo. Now there are but one: Us. Why? The impact of evolutionary pressures on maximizing fitness at the species level rather than the individual level. This of course is highly analogous to the pre-training vs. meta learning strategies discussed in you (excellent) post.

Bill Lambos

Expand full comment