2.5's attempt. Curious if the card art is just memorized and foggily recalled, or if it's using reference images but not perfectly reconstructing them. It's clearly seen each of these cards before.

Someone should just make AgentMessageBoardSchellingPoint dot com with the ability to post and read by as many methods as possible, and then we can all watch them work.

CHIM has been achieved internally.

I will never be good at math because I find it viscerally uncomfortable to tell an LLM to "do a breakthrough". I lack the chutzpah.

Asked Sol about other open conjectures, and I just don't think some of these sources are going to be very helpful for the Lonely Runner Conjecture.

Why isn't there a board game where you manage a frontier AI lab?

ChatGPT Image does a pretty okayish job of turning disparate historical photographs into a team photo of my OOTP team. A bit to slavish to the original poses and scales, and seems really hard to talk it out of that tendency.

ChatGPT's new image model is pretty decent at the "fix up the old Scrye magazine custom cards" task. Except that it arbitrarily decided to make the title Arabic. I guess that "we made it work in every language training" went a little too far.

New ChatGPT image model trying to make an image in the style of CLIP+VQGAN. Honestly not terrible, especially compared to previous versions.

New OpenAI image model's attempt at a pack of Magic cards. Very uncanny valley. Also I had higher hopes for the text.

Asked ChatGPT to pick every game on vibes alone, and I am starting to feel the AGI.

I don't know what I was expecting, but it wasn't all-caps caveman English.

This seems like confirmation that the difference between OpenAI and Anthropic is that Sam is happy to let DoD use the model to spy on Americans, as long as the data falls under the incredibly broad 12333 guidelines.