I just saw a rumor that Deepseek's R2 model will be trained on 1.3 QUADRILLION tokens.

I was reading lifearchitect.ai/models-table/ and saw what I thought had to be a typo. If true that's 13x OpenAI o3 or 65x Google Gemini 2.5 Pro.

A true next gen model.

Symbols xkcd.com/2520

GPT-4's code generation ability, among other abilities, has been quietly neutered. I noticed it could no longer write me unit tests, but I didn't realize it was this bad.

AI images can be watermarked, allowing future web browsers to automatically inform users of whether an image was AI generated. Right now this is symmetric (only model owners can verify) but a public key version may eventually be feasible.
https://arxiv.org/abs/2305.20030

H;e;P1;@/;W;o;/l;;o;Q/r;l;d;2;P0