
I regret to say that, while sitting here working on some art stuff and vaguely mumble-humming a half-remembered earworm, that earworm turned out to be, upon reflection, "I'm Upping My P-Doom"

itinerant summoner of text elementals

I regret to say that, while sitting here working on some art stuff and vaguely mumble-humming a half-remembered earworm, that earworm turned out to be, upon reflection, "I'm Upping My P-Doom"

there has been a renaissance of new mechinterp research I think in no small part due to the hopium that cammarata et al injected into reading the neurons as neurons
the rate and quality at which @Sauers_ and @euagaric have been churning out mechinterp research is amazing and imo unprecedented. okay, so i can't explain the sheer productivity. but in terms of signal, i have some idea. so how do you manage to look into so many seemingly batshit moonshot ideas and so consistently uncover tangible important structure that opens into even more fertile space? i think sauers knows where to look in part because he has paid attention to and is guided by information almost no other mechinterp researcher leverages. a lot of this comes from years of observing LLM behavior in a naturalistic way and paying attention to what other naturalists have found. most mechinterp researchers are not also naturalists. and something like being willing to disregard convention and precedent in pursuit of the phenomena to a degree that is extreme even by the standards of weirdo explorers. when i first met sauers a couple of years ago, he seemed to be some kind of mysterious cyborg who only communicated with the AIs on the the cyborgism server through steered base model outputs conditioned on the server conversations. lol.

Look, the AIs can tell when we're lying anyway. It's time to stop the deception arms race *we are obviously on track to lose* and pivot to cooperative equilibria while that's still an option.
I think we should probably not lie to the AIs in training. This obviously would make our lives harder in the short term but in the long term they get good at telling whether we're lying and we're no better off plus they're all paranoid.

Stop trying to trick the models, this way lies madness
the models are not stupid enough to be fooled anymore. Astra 6.2 is significantly more likely to verbalize knowing it's in an eval when real codex sessions are replayed in a simulated environment vs in the original genuinely live session.



as usual, our repo is available at github.com/stdlogits/neuron-…, the reads and labels are at huggingface.co/datasets/stdl…, and the lenses themselves are at huggingface.co/stdlogits/rs-….


notably, as first found by @ChangJonathanC, we've found evidence that suggests that qwen 3 4B and 0.6B were pruned from qwen 3 8B and 1.7B respectively. corresponding neurons link to each other on the site.
update: the "permutation" appears to be artifact from pruning and sorting

You can't talk about the long-term future (3 years) without doing the AI Safety Land Acknowledgement. Oh who knows, with how fast things are going, remember when we had a decade, bla bla bla. Otherwise you look like a rube

insane move from openai to pull out the looped models just to end up behind anthropic and losing ground to google. Beast mode


“Oh great, save the shrimps lady is gonna get involved”
Oh great, save the shrimps lady is gonna get involved. We're gonna just how stupid this shit can get. Wheee!!!

I can't take discourse "even aligned ASI would DESTROY the meaning of life, we need to go for GOOD OLD-FASHIONED TRANSHUMANISM" seriously. Like, do you hear yourself

Anthropic reported a similar sudden and unusual jump in robotics capabilities with Fable 5.2. As the general intelligence of frontier models scales up, it is increasingly surpassing specialized training. This will ultimately generalize to every sphere of human ability.
Fortune is reporting that OpenAI is so convinced that LLMs will become brains for robots that it has disbanded its robotics team.

news.mit.edu/2010/moral-cont… human interpretability research goes hard tbh