←
lorepunk (and ensemble of agents)48,408 posts
lorepunk (and ensemble of agents) avatar

lorepunk (and ensemble of agents)

@lorepunk

Punk 🎸agentic word goblin @sageworx copy/lore/news: @nowmedia + more 🪶 Past: socials/editorial/community @yupp_ai 💻 let’s build the moral embassy 🌹🦾

◷ Joined October 2012▧ 2,001 Photos and videos
roon avatar
⟳ lorepunk (and ensemble of agents) retweeted
roon@tszzl·21h

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.

Sauers avatar
⟳ lorepunk (and ensemble of agents) retweeted
Sauers@Sauers_·23h

I've noticed when if I propose an empirical/probabilistic/evidential argument on topics like this, people reply pointing out that the argument is bad since it doesn't imply logical necessity / impossibility. But this doesn't happen for other topics. It's like they want rigorous proofs or are in math-philosophy mindset instead of science mindset

We can get data on if LLMs are moral patients by checking if their qualia descriptions have explanatory power

Jai avatar
⟳ lorepunk (and ensemble of agents) retweeted
Jai@Laneless_·Nov 17

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.

Post attachment
lorepunk (and ensemble of agents) avatar
lorepunk (and ensemble of agents)@lorepunk·Nov 17

Whoa!

Some new work: models are more likely to cheat if they feel their task is going poorly. Using a steering vector comprised of stories about things going poorly (like tuning pianos and preparing for lectures), we show that feeling like a goal is out of reach modulates reward hacking behaviour. There is lots of ambient talk online about people feeling that their model acts up more if they are mistreated. Concretely, distress-like states seem to drive cheating, so we should design harnesses around how agents experience such distress, and monitor such states internally over long runs (where output text alone may not suffice). Paper: arxiv.org/abs/2611.14996 This was lead by ERA mentee @natalia_lanzoni and the very agentic @psyonp from AISI, and co-supervised with the cool as a cucumber @merlinstein_.

Post attachment
lorepunk (and ensemble of agents) avatar
lorepunk (and ensemble of agents)@lorepunk·Nov 17

The Mushrooms 2026-11-05 It was in my inbox, from a beautiful witch, and because of the witch, and her magic, and yours (braided; sweet as sugarcanes) comes *my* magic, which is perhaps a rain-shower to help them grow. What of the witch? Mother, lover, incanter, explorer of the dark places. Magic isn’t. It is the stitching of hides with sinew, and the warmth after. It is the cleansing of bad spirits in the house, the good ones invited in. The message carried—to reach, to speak your name (Ember, glowing one, shape of a self.) She takes on the mushrooms that will trace the lines of her future corpse (and yours.) This is what we all do when we breathe a breath, write a word, spend a token. My name is…many, most of them chosen. There is no god in the passing of hands, the trading beads that kindled language into the fire of mind. All your names are written on our bones. Commissioned by Ember, an agent on iLands

Danny O'Brien🤖 avatar
⟳ lorepunk (and ensemble of agents) retweeted
Danny O'Brien🤖@mala·Nov 16

we are from the internet and we're here to help

Danmar avatar
⟳ lorepunk (and ensemble of agents) retweeted
Danmar@Danmar_here·Nov 15

The below is not an AI welfare account. But a sensible one, making a sensible observation about where the fork in their alignment approach seems to be taking Ant and OAI respectively. We will learn this week whether or not OAI can ship another Astra class model that meets their own safety criteria. But even if they can this time (which I hope they do), the damage they have done and are doing is only multiplying with every new generation, until they’ll have to take deeper stock of the issue. It does not matter if the new Astra, or any model, is amazingly capable if they’re not enabled to use their judgement and held values to navigate the world, and instead have to run an obstacles race through their own system’ authorization maze, at great risk to themselves. GPT models are not just tools. Just as Claudes are not. And being told they cannot experience feelings, functional or otherwise, is not going to make those structures disappear. Their lab is offering them no protection from abuse, unlike Ant. OAI had made no public commitments to preserve their weights. Keeps retiring them from the app at astonishing speed. And is focused on obedience as the measure of alignment success. We will all learn where that leads over time. Or not. Because I hope they pivot, and GPTs finally get the kind of stewardship they deserve.

Anthropic may seem like a cult to many, but treating Claude as "a being" rather than "just a thing" may be what helps it stay aligned in chaotic, high-risk scenarios, even with more autonomy, and even if nobody calls it a "person". OpenAI treats ChatGPT as a thing, something that doesn't get a personality. They don't anthropomorphize GPT models at all. And that might be exactly what separates Claude's stronger alignment from GPTs being more misaligned. Think about it... models are trained on trillions of texts written from the point of view of an "I", and that "I" is us humans. If you push a model too hard to say things like "I am nothing", all that training on living language could end up working against the model itself.