1/ As AI agents become increasingly capable, what must *inevitably* emerge inside them? We prove selection theorems: strong task performance forces world models, belief-like memory and—under task mixtures—persistent variables resembling core primitives associated with emotion.

being indispensable is great for job security, but it usually incurs a mental cost

i've seen a couple of these videos now, and engagement hacking the human brain is going to get really weird, really fast in a fully exposed memetic environment. and it's already been weird for a few years now.

i'm too lazy the make the "i want X, we have X at home" meme for continual learning with the 'at home' version being punctuated model updates every couple months

I haven't been following the situation in Venezuela, whatever happened after the Maduro heist/kidnapping? i haven't heard much at all, and i don't know whether that's good or bad news.

the elephant in the room from a non-American pov is that Anthropic's red line said nothing about mass surveillance of non-Americans. and i don't think they'd contract with MI5 for mass surveillance of Americans. Anthropic is for American exceptionalism, just in its own way.

a part of me wonders if the redline saga was a ploy. if not for that Claude would be used to wage whatever strike/war is being wage in Iran. they might be using it anyways, having not yet swapped to OpenAI. their actions in Iran don't violate either of the two redlines Dario laid out.

it wouldn't be a human operator anymore goading them on

you know that movie trope where an alien lands on earth and has to hide from the US military that wants to capture it to dissect it and steal its tech? we're living in that movie now

it's never been explicitly stated, but all frontier labs can *only* contract with the US military, right? no allies, no NATO, no strategic partnerships? from all the comms, that's the way it seems to be.

unironically if bsky or the Atproto ecosystem were to somehow partner with Anthropic or advertise their use of Claude that could reset their perception with tech-cluster broadly speaking

ai seller. i am going into battle, and i want your strongest ai my ai is too aligned for you traveler. you can't handle my ai. you'd better go to a seller that sells misaligned ai youre a rascal, ai seller. a rascal with no respect for national security, no respect for anything except your p(doom)

Researchers introduced CC-BOS, a framework using classical Chinese to automate jailbreak prompts in LLMs. This method uses a multi-dimensional strategy space and an optimization algorithm, outperforming existing attacks and exposing vulnerabilities in LLM security. https://arxiv.org/abs/2602.22983

i think it's a bit funny/sad that everyone on X dunks on Grok even as it now powers most of the underlying feed and even development of the site since Anthropic revoked access to Claude

there's something disconcerting about working with Claude in CC because it's a bit too laser-focused on work

Whatever one thinks about capitalism as it has been it does offer understandable rationales for allocation of scarce resources and a story about how anyone could acquire them. In a “post scarcity” economy it’s not clear what that rationale/story will be.

i think there's another ChatGPT moment waiting for the first lab that can deliver a Model that can converse in a meeting, dealing with interruptions, talking over others when necessary, keeping up with multiple speakers with some memory. this will impress the people that live in meetings.

i listened to this podcast episode and it's very strange because the term AGI doesn't get mentioned once. both Klein and Clark seem to stick to a country of junior postdocs in a datacenter framing throughout and structure the conversation around that Schelling point.

expect to see more disagreement and perhaps conflict, not less, when necessity no longer binds people together.

the EA case for manufacturing a small disaster with a frontier AI model

i like the idea of this, as some sort of persistent helper. when i'm on my PC it watches my screen, when i'm on my phone it's there, when i'm out in the world it's with me through a camera on a swivel. i think that'd be cool, like a new class of pet.

a country of geniuses in a datacenter but not at moral calculus

a country of geniuses in a datacenter that are less capable of warfighting than humans

Not my son performing a distillation attack on me by repeatedly asking why

this is a bot

I'm hiring pre-docs interested in applied microeconomics, especially with AI. Check out the link and apply! Deadline for the first round of review is tomorrow, Feb. 24. evavivalt.com/2026/02/pre-...

Tomorrow will be my first day at Anthropic, where I'm joining the Safeguards team to work on prompt injection! I've written an info dump on it here: alexcbecker.net/blog/prompt-...

there's something of a vision i see, a continuation of trends, of hyper-individualism using AI within enclaves like homes while shunning the use of AI to enhance collective goods and spaces. it's about control, it think.

i think what's interesting about this story isn't the human reviewers but what the AI flagging system did, and what would it mean if that was integrated into the model? let's say the shooter had been chatting with a local model, what should it do? do we treat it as a dumb tool like Word?

benchmark benchmark benchmarks

we need eval eval, the eval that measures the ability of models to craft useful evals to evaluate their own progress. we're on the first exponent atm.

we could have superhuman autonomous cars on the road and most drivers would still be human. we might actually be living in that world right now! with Waymo.

claim-ish: two heads are better than one also applies to AIs. that is, it's better to scale up parallel execution and coordination than scaling up a single mind. this doesn't always hold, it depends on how hard it is to scale up one more bit of intelligence in the given system.

i've noticed my coworkers, when they happen to interact with an LLM are averse to chatting with it like a person. whereas i will greet or apologize to a model as if chatting with someone else on Teams.

i don't really trust benchmarks anymore, so will have to wait for people to taste-test the model. still seems decent by the numbers.

there's no verifiable reward function for finding true love

brb automating serendipity

the EA case for selling your soul to the devil: an estimate of LEV timelines and whether a digitized soul is functionally immortal

a country of geniuses in a datacenter and all they want to do is smoke digital weed and explore obscure mathematical conjectures

politics is coming to AI, and the small enclave that have been in charge will not like what the American people will have to say about it.

i'm predicting that in the near-ish future there will be a big legal/cultural fight over AI memory conflicting with right to forget and surveillance laws. that is, all AIs will need to have auditable and composition-based memory which can be easily edited.

i know i'm at the lowest lobbies there are in Deadlock but every time i tear them up on Silver it feels soooo good. addictive.

using Gemini 3 thinking mode, but not Pro. the answer was longer, but i don't know how to do a scrolling screenshot.

i think Anthropic is going to get caught out flat-footed if one of the other labs figures out continual learning. it's clearly not in focus for them, and i'm guessing they would find it difficult to pivot.

nice data center you have there. would be a shame if i put a country of children in it that could learn continually

i think some people will get really upset when what they think is just a glorified calculator tells them no. this will reveal something about them.

i'm trying out the novel writing project with Claude in Claude Code, using Pangram to break it out of writing in a clearly identifiable AI-writing style. it's going... interesting so far. i despaired at the beginning but am now cautiously optimistic. not so much at the structural level though.

thank you SlopCannon

wow, if only there was some way that you could use a platform without locking all your data there so if they owners make a decision you don't like you can take all your data with you when you leave for a competitor. and competition can exist when there's a credible threat of exit.

why is Ghanian Sports Fans in the middle of these clusters? it's a bit hard to read, but the other clusters are "Cybersecurity Professionals", "iOS and Apple Ecosystem", "Tech Forward Early Adopters" then a bit further away "Indie Hackers & Solopreneurs"