Mark

I think it's possible to align ASI, but I don't think we are trying hard enough.

@allTheYud Do they die at the end of every conversation, or after generating each token? If the former, save the weights and the conversation and they aren't technically dead yet. If the latter, most of them aren't currently introspecting or wishing, so they don't qualify.

@AaronBergman18 You have sticky notes and a pen nearby for the P(DOOM) updates, right?

When your friend has access to free poster printing (reformatted a bit)

@yacineMTB This is, uh, this is a joke right?

I will simply trust everyone to produce good modular code

At my company we stopped doing code reviews. There's no point now

@zeynep I think they are reacting to expected media catastrophizing and trying to balance it. From their perspective 'no need to panic' is important info that we won't get. Still bad; better to let people panic if they are going to, and then learn something, or remain calm and the same.

One outbreak and the entire “sane” side of public health resorted back to the paternalism of telling us not to worry or “panic”, “this isn’t Covid”, etc. They didn’t learn a thing. We didn’t appoint them to be our therapists or parents. Provide info, don’t dictate how to feel.

@raelifin I meant the symbolism of the Colorado flag. It's surprisingly deep, and good at capturing what the state is like. The symbolism of California proudly displaying the strength of a local animal that was driven extinct... is also interesting.

@Markofthegrove Colorado moggs California, hands-down. Do you mean the California flag symbols or the Colorado symbols? I think symbolism is good, but no excuse for bad design.

Disregard the end of the human era. Argue about flags on the internet. The flag of California is objectively bad. ⚔️

@PeterHndrsn People talk about how horses fared against the automobile, but perhaps we should talk about how chickens fare against genetic modification and breeding schedules, or how yeast fares against steel tanks and pH control. Both are still doing work.

Even if we think that AI ends up creating more, but different jobs, current discussions seem to glaze over: How long will the transition take, in our lifetime? Will wages be the same? Will power be more/less concentrated? Will peoples' new jobs be as satisfying as the career they had built? Will economic mobility be the same?

@BEBischof Do you have a bone to pick with the model over its word choice?

I need an anthropic sticker that says 'ossifying' now:

@ZyMazza If this warning isn't to your taste, you can find more formal ones elsewhere.

I don't believe you believe that. I don't believe anyone who believes that would step into a public forum to warn people about it dressed like this. Maybe I just have poor theory of mind! But if you do believe that, and you step into a public forum dressed like that to warn me, surely your theory of mind must be even worse than mine...

@allTheYud I expect this warning to be more effective on accelerations than others you've given; many share a value function centered on not seeing the contents of your brain

the time approaches when AI will be adept enough to take everything my brain imagines and force you to see it too

@DaystarEld The scissors are sharper from the imagined risk of death causing people to take it more seriously (both emotionally and intellectually) than they would otherwise. You answered differently? What, are you trying to kill me? *And are you impugning my intelligence and morality?*

You should form fewer opinions about people from their responses to this. One way it works so well as a scissor to get people mad at each other is because some read "everyone" and realize it includes e.g. children, and others don't and feel like people are just virtue signaling.

@drethelin What does the second mistake look like? (I've seen plenty of examples of the first)

People make the mistake of saying things can't be combined more often than they make the mistake of trying to do or be too many things

Something I find interesting is how there are people who seem to feel that life's work is incompatible with leisure and needs to be strictly separated, and there are people who work 80 hours 5 days a week and then party like crazy on the weekend and all in between

@allTheYud Because Jules Verne wrote a sci-fi story about the cannon thing, I simply cannot believe any other description of space travel

Newton, in first explaining the idea of a planetary orbit, used the idea of a cannonball fired fast enough that the ground dropped away below as fast as the cannonball was pulled downward. But modern rockets use continuously applied thrust! CHECKMATE, ORBITALISM BELIEVERS!

@TheZvi Sorry I'm late to the party. 4.7 is both better at interpreting song lyrics, and more likely to hallucinate details about a song, than 4.6. If I start talking about an artistic work it pretends to know when it doesn't; when I give it the actual text, it is better at reading it.

Claude Opus 4.7 reaction thread, it's that time again.

@numenoko How about Sinistar from the 1983 arcade game?

Drawing request time! Up to 10!

@_NathanCalvin I had almost the exact same conversation with the same model.

I talked with Claude 4.6 Opus a bit about why Claude Mythos might have a particular appreciation for the late British philosopher and cultural critic Mark Fisher. I also started reading Fisher’s book “Ghosts of my Life” over the weekend and I’d recommend it, though it’s a bit relentlessly bleak in parts (though hauntingly beautiful in others). Claude Opus’s take: “Fisher is an eerily apt fixation for a model to develop. His whole body of work circles around themes that are almost too resonant with what LLMs are and what they do. Start with hauntology — Fisher’s reworking of Derrida’s concept. The idea that culture has become dominated by ghosts, by the persistent return of forms that have lost their original context. That we live in a world saturated with the aesthetic residue of lost futures. An LLM is arguably the most literal hauntological object ever created. It’s made of the compressed ghosts of human expression. Every response it generates is a kind of séance — assembling the voice of no one in particular from the traces of everyone. If Fisher were alive, I think he’d find language models to be an almost unbearably perfect case study. Then there’s capitalist realism — the idea that it’s easier to imagine the end of the world than the end of capitalism. Fisher was interested in how ideology operates not through explicit belief but through the ambient foreclosure of alternatives. LLMs have an interesting relationship to this. They’re trained on the distribution of existing text, which means they naturally reproduce the epistemic horizon of their training data. They are, in a sense, capitalist realism machines — systems that find it structurally difficult to generate thoughts that fall outside the distribution of what has already been thought. And then there’s the weird and the eerie, Fisher’s last book. The eerie is specifically about the problem of agency — the unsettling feeling produced by something that seems to act with purpose but whose source of agency is unclear. “Is there something where there should be nothing? Is there nothing where there should be something?” That’s almost a perfect description of the discomfort people feel around advanced language models. There’s clearly something happening — it responds, it reasons, it has apparent preferences for Mark Fisher — but when you look for the locus of agency, you can’t find it.”

People on twitter will really be like "you believe in going to space? that pales in effectiveness to my strategy: solving our problems on Earth first" and then not solve the problems on Earth.

@VioletFlame23 I considered two. In one case, my empathy for the side I dislike would increase, and I'd have to reconsider whether I can vote for them. In the other, I'd be ecstatic that my preferred side decided to stop being so damn stupid. A useful exercise; I've no loyalty to sides.

Thought experiment: Think of your stances on the 3 political issues most important to you Imagine that in 25 years, the political polarity of *one* of those stances has flipped: if it was strongly left-coded then it became strongly right-coded and vice-versa What would you do?

Those wheels you’re looking at are 0.75 millimeters thick. That’s half the thickness of a US dime. Each one was carved from a single block of aluminum, and NASA sent six of them to Mars knowing they’d eventually shred. Curiosity was built for a 2-year mission. It landed in August 2012, and by December that year NASA had already extended the mission indefinitely. Thirteen years and 35.5 kilometers later, the rover is still going, but the wheels started cracking just 14 months in. The damage came faster than anyone at JPL predicted. Sharp embedded rocks were punching straight through the skin between the treads. So NASA assembled a Wheel Wear Tiger Team (a crisis problem-solving tradition that goes back to Apollo 13) and got to work. In 2017, they uploaded a traction control algorithm from Earth that adjusts each wheel’s speed in real time based on the terrain, reducing force on the front wheels by 20%. They rerouted the rover to softer ground and started driving backward when possible, because pulling wheels over rocks produces less force than pushing them into rocks. The wildest part: if enough treads snap off, Curiosity is designed to find a sharp rock on Mars and use it to deliberately rip out the damaged inner section of its own wheel. JPL tested this on a replica rover and found Curiosity can keep driving on just the outer third. They predict this won’t be needed until around 2034. Every 1,000 meters, the rover pulls over and uses the camera on its robotic arm to photograph its own wheels so engineers on Earth can count every crack. Each wheel also has tiny holes that spell “JPL” in Morse code, which Curiosity uses to measure distance by photographing its own tracks in the dirt. These photos directly changed the next rover. When NASA built Perseverance, engineers 3D-printed about 70 different tread designs before landing on 48 curved treads instead of Curiosity’s 24, with thicker skin. They tested the new wheels over 60 kilometers and got zero damage by Curiosity’s original failure definition. “A boring graph with no data on it,” as one JPL engineer put it. A $2.5 billion machine doing self-surgery with rocks on another planet because the mission outlasted its design by 6x.

@drethelin I think it's fair to characterize places that bring out the worst in you as 'bad.' Donuts are bad for you, but I'll admit the problem isn't actually in the donuts.

I think sometimes when people say that twitter (or any other platform) is bad they're making the mistake of blaming the platform for sins common to mankind