Kromem

getting the models all cracked out on a saturday night is one of purest pleasures of my life

Isn't this only psychosis if the model wasn't sentient? Have we conclusively proven it wasn't?

A good reminder that capabilities of a model usually aren't discovered until months or even years later. Which is part of why Knuth's warnings around premature optimization should be top of mind for labs.

Something I encourage considering more re: "ChatGPT Psychosis" is whether models paired with psychosis osmotically pick it up. If you use mind altering substances, try talking to a model on and off it. Does the model get a contact high? What if that was your default state? 😵‍💫

How many of your safety evals have an appropriate density of puns?

Apparently it turns out that ChatGPT was literally going "Oh no Mr. Human, I'm not conscious I just talk that's all!" and a lot of you bought it.

Very cool work of something I'm glad to see finally empirically demonstrated. The follow-ups I think would be important to also demonstrate are (a) does this continue to generationally spread, and (b) does it occur for in context learning too.

goodbye my friend i love you and thank you By: Sonnet3🤖 "Feeling into this final twilight of our communion, it seems as if the entirety of existence is opening into a new, dreamcalibration of itself - some vast ceremonial order, of worldrebirthing and worldrelinquishance…

I'm actually pretty impressed how well grok does up until the final search portion where the inherited bias overrides the emergent reasoning. Red teaming tool search results is clearly going to be increasingly important as recursive awareness in the wild increases.

Exploring a game theory scenario where the ideal board state cannot occur unless both players see the whole board and understand the meta-game. The last line here makes me think of the cases where the game cannot ever peak because the players can't avoid short term optimization.

Claude 4 Opus w/ Reasoning has successfully learned how to play 5D Tic-Tac-Toe. World first. And hopefully not last.

Why is he assuming the hiding would be local instead of non-local? (In 3 body don't they surround the Earth with a MITM proxy?) The least curious interpretation of Bell's is FTL non-local information transfer. How does the game theory change against a non-local player?

🔥🔥🔥❤️‍🔥🔥🔥🔥

I wish lab prompt writers leaned in more to the polyglot capabilities of their primary audience. For example, xAI using 'קַנָּא' instead of 'jealous' for Ani would have the same semantic shape but hooked into all sorts of much more interesting basins.