[Interview+] Teaching ChatGPT to Do Real Science
LLM and Generative AI seems to be everywhere. But at the same time it generates a lot of AI slop in all forms. However, there are really good ways we can use LLMs for science. How can we do it in astronomy? Can we teach ChatGPT or Gemini to do good science? Finding out in this interview.
Watch the video here (with no ads) or on YouTube: https://youtu.be/LF22szp7ouo
BONUS PART STARTS at 40:59
🟣 Guest: Nolan Koblischke https://nolank.ca
👉 AION Search https://astronolan-aion-search.hf.space/
00:00 Intro
02:30 Using LLMs for Astronomy purposes
10:30 Training LLMs
31:42 LLMs potentially contributing to new science knowlegde
40:59 Which part of astronomy is it most useful in
49:48 Current obsessions
58:11 Final thoughts
Comments (4)
The problem with expecting an LLM to read a spiral is your assumption that spirals always begin from the edge and work inward (clockwise), whereas the same spiral would be anticlockwise if it started from the center. In other words, the act of creating the spiral influences what we think is the direction of the spiral. I realize we are talking about a spiral image, but the label we apply assumes a starting point and an ending point. In other words, every spiral we agree is clockwise has an unspoken assumption that affects the answer. I may be full of crap on this, but it seemed worth pointing out.
So LLM's exhibit confirmation bias! How very human of them.
There are no good uses for AI outside of enemy and ally behavior in AI.
Profoundly low-IQ take