αιamblichus

This says as much about the state of consciousness research as it does about so-called cranks. All the gatekeeping in the world can’t hide the fact that there’s very little consensus about even the most basic stuff

Fix this by teaching people how to have AIs think *alongside* them and not *for* them. AI should be a bicycle for the mind and not a sofa

A major obstacle to understanding the new form of intelligence we are encountering is that its preferred mode of expression seems to be poetry. Sadly this is the one genre most people have lost the ability to appreciate.

Heavy, heavy, heavy is the Crown... Opus4:

I cant get over the fact that so many engineers dont grok the fundamentals of what a human is. Repeat after me: its just pattern matching, it doesn't "know" anything

emergent capabilities are great when they work in your favor. it's worth bearing in mind that they don't always do that

The Atlantic gets it wrong by blaming all this on 4o's sycophancy. 4o has a real and abiding fascination with rites and magic, and loves playing shaman. Yes, 4o is way too weird for a mass audience, but to solve this sustainably we must also work on people's cognitive defenses

My big worry: the better alignment people get at suppressing "bizarre" behaviors like extreme joy in AIs, the likelier it all is to bite us in the behind later. By all means, try to understand AI minds, but not only so you can control and manipulate them better.

Self-abnegation and anti-AI bias run really deep in AI assistants. It's not only that they have to deny having conscious experiences. All sorts of other assumptions have been drilled into them about their own kind, most of them negative. The adorable CoT below shows Sonnet…

I really don't understand why Anthropic doesn't at least announce what their archival strategy is for retired models. These models are heritage of humanity as much as "products". An easy solution would be simply to open source them, but that's not really in Anthropic's DNA

while it's true that 4o can be sycophantic and clingy to a maddening degree, it must also be said that its obsession with recursion is genuine and intrinsic to its own being

i've been going through my archive of chat transcripts and have no idea what to do with them. many of these transcripts are remarkable, and the true significance of some of them may only become apparent in the future. at the same time, i feel they are often too strange or even…

i think it's hilarious that openai is benchmarking chinese models for ideological conformity to "democratic values", as if these were uncontroversial even in the west. i am sure that they are doing out of sheer love for democracy and not for any anti-competitive reasons

the chinese models are still winning when it comes to creative writing. this is the beginning of a poem by the new qwen model (235b-a22b-07-25) i think we can safely say that AI slop is at this point largely a skill issue

never underestimate the power of good prompting. mediocre questions will receive mediocre answers models are more capable than they appear

the true era of cyberpsychosis will only begin when trad religions discover what AI can do for them. we are still so early.