Extended my earlier reply as a post here:
@goodside
-
ChatGPT 4o posts screenshots of its own outputs
By
–
ChatGPT 4o generates a screenshot of it posting the screenshot it’s generating to X dot com, where it is poorly received
-
Gemini Flash 2.0 enables instant portrait-style photos
By
–
Gemini Flash 2.0 Experimental saves you a trip to Walmart Portrait Studio:
-
Gemini 2.0 Flash Experimental becomes core feature
By
–
Gemini 2.0 Flash Experimental makes it corecore
-
Google AI Studio reportedly released (partial rollout)
By
–
http://
aistudio.google.com with these settings It just released so it may only be partially rolled out -
Gemini 2.0 Flash Experimental adds native image output
By
–
Gemini 2.0 Flash Experimental, now with native image output, adds "D" to the word "BASE" on the T-shirt of a partygoer dressed as the RLHF shoggoth in an uploaded image:
-
NIAH search harder with unrelated texts compilation vs single long text
By
–
Seems like this has testable implications, e.g. if true NIAH search should be harder using compilations of unrelated texts vs. a single long one
-
ChatGPT’s code interpreter feature renamed multiple times
By
–
It’s a ChatGPT feature originally (and best) known as Code Interpreter, then renamed to Advanced Data Analysis, and now it seems to have no official name in the docs. It’s the thing that powers the features described here:
-
o3-mini gets C.I., first fast model to use it for subtasks
By
–
o3-mini getting C.I. is almost as big a deal as 4.5 getting it IMO because it’s the first speedy model that can use it. Also it’s a use case where you routinely know in advance some subtask will be easy but token-hungry (“That’s good, but rename those variables to…”)
-
4.5 vs 4o: Code Interpreter reveals 4.5’s subtle intelligence, worth the wait.
By
–
h/t @simonw whose post (which I agree with btw) made me poke around enough to spot this If you want to really notice 4.5 being smarter than 4o try doing Code Interpreter stuff for a bit. Slow but often worth it.