Andrew J. Haman
andrewjhaman.bsky.social
Andrew J. Haman
@andrewjhaman.bsky.social
Sincerity is not weakness

Programmer - Work featured in The New York Times, Atlantic, Business Insider, Linus Tech Tips, WIRED and more.
I understand what you are saying but it doesn't really make sense to me to not count the tokenizer as part of chatgpt. Even so, if we operate under that assumption it's still remarkable that there doesn't seem to be an axis in embedding space corresponding to letter counts.
August 9, 2025 at 12:43 AM