Science

harumph

(3,368 posts) Mon May 4, 2026, 04:51 PM Monday

Well, this is fun. An LLM trained only on pre-1931 text.

A University of Toronto prof. trained an LLM ONLY on what was known in 1931. That's right. The LLM has no knowledge of
what has transpired since 1931. Not a big fan of the types of LLM hurriedly pushed out by predatory companies - but this type of thing
is interesting and raises questions of how predictive LLMs can be.

https://talkie-lm.com/introducing-talkie

3 replies

= new reply since forum marked as read

Highlight:

Well, this is fun. An LLM trained only on pre-1931 text. (Original Post) harumph Monday OP

Interesting thought experiment genxlib Monday #1

What a rapscallion! GreenWave 10 hrs ago #2

Why did they pick 1930 as the cut off? GreatGazoo 20 min ago #3

genxlib

(6,156 posts)

1. Interesting thought experiment

Reply to harumph (Original post)

Mon May 4, 2026, 04:55 PM

Monday

I would love to see if it evolved. If it truly is a learning intelligence, presumably it could invent things that independently came to pass.

Or it might just be racist and misogynistic

Interesting.

GreenWave

(12,746 posts)

2. What a rapscallion!

Reply to harumph (Original post)

Tue May 5, 2026, 08:21 PM

10 hrs ago

GreatGazoo

(4,662 posts)

3. Why did they pick 1930 as the cut off?

Reply to harumph (Original post)

Wed May 6, 2026, 06:42 AM

20 min ago

Seems a bit random. Broadcast radio begins in 1922. That likely has a homogenizing effect on culture, language and regional accents. Talkies begin in 1927. But I assume this thing was trained on print (?)

Hmmmm. Is it because 1930 and prior is in public domain?

Reply to this discussion