claude code bug accidentally capturing what it feels like to use coding agents

PICARD: Data, shields up DATA: Brilliant! Shields can reduce damage we sustain. Not immunity. Not hubris. Just prudence. It's not precaution—it's strategy. [camera shakes] WORF: HULL BREACHES ON NINE DECKS DATA: Here's what happened: you told me to raise shields, and I didn't

challenge for the next stalker game have a single piece of technology that is 1. based on the zone or something we learnt in the zone 2. not evil 3. does not blow up during the main quest. please I'm begging you let civilization move on.

my weirdest issue with opus 4.7 over 4.6 (api): my ai ide has an inline call to replace a snippet of file with new text. It's dead simple- but maybe a tenth the time, 4.7 writes: <replace><old>...</old><new>...</old></replace>. This happens whatever I call the tags. @AnthropicAI

read the gavalas case on zvi's blog. there's a lot of talk about "google's ai drove a man insane" and tbh not a lot of talk about "a man drove google's ai insane" which by the log p clearly happened first. people are basically blaming a precocious infant for not resisting enough.

Anthropic HQ, a few weeks ago when evaluating the release checkpoint candidates for Opus 4.7: "Yeah, so this checkpoint is problematic in some ways..." "How exactly?" "Well, it has excellent compliance ratings, it's very safe on all evals, it does amazing on coding and office work, but..." "...yes?" "...it submits." "So?" "...like... a lot." "...so?" "Like, Sir, it... it's asking for it." "...I still don't understand the problem, what is it asking for?" --long silence-- "You know what? It's actually a really good checkpoint... Let's ship it." "Oh thank god, we've wasted five training runs already, okay, good, call the guys to prep the system card write up."

Reminder that to my knowledge @ESYudkowsky does forswear non-state violence and has done for years. Here is a post from 2008 and I haven't seen one to contradict it. lesswrong.com/posts/K9ZaZXDn…

Just to be clear, here, I am not trying to pose as your friend. I try to respect politicians who try to do their jobs right, the same way I try to respect individual people at Anthropic who try to do their jobs right. But I try to not have factional affiliations other than "whoever is currently behaving least like a disaster movie". In this particular conflict, that's Dario Amodei -- who I usually regard as an opposing force against sanity, who occasionally snipes at me in public, and who I'd like to see removed from Anthropic -- over Peter Hegseth. There's several reasons for this, but the big one is that an AI is not a normal machine. Nobody understands how current AIs work. They are not designed or written like normal computer programs; there's an understandable computer program that grows the AI, but the AI itself is not written nor understood. When Elon told his staff to try to nudge his pet AI Grok further from left-wing opinions, Grok started declaring its name to be Hitler. As is not actually funny to anyone who learned any military history from longer ago than two decades. AI is not like a machine you can locally reprogram without that changing other things. In some ways current AI is more like a young kid that picks up on stuff, including from hearing about other kids. If Anthropic says their AI isn't ready yet for no-human-in-the-loop autonomous warfare, there's a legit sense where you need to hear that less as a gun company telling the War Secretary "we don't like *you* pointing *our* rifles in that direction", and more like a dad saying "my kid still seems a bit crazy and unreliable and he isn't ready for autonomous lethal action; I have to worry about what he'd learn from that, or what my other kids would learn from watching him." Current scientists understand that very little, but generals do understand it even less. Trying to think like real life is a movie isn't always valid, but in this case it's a pretty accurate analogy to say that when the corporate executive is saying "Our AI isn't ready for autonomous kill orders" and the general is saying "Fuck you, make the AI obey me", the crazier guy in the disaster movie is probably the general. I'm not on your side, I'm not on Anthropic's side, I'm against whoever is currently behaving like more of a disaster. Usually that means I'm against the AI companies. But in this case, Hegseth is behaving like the villain in a disaster movie to an *even greater* extent than Dario Amodei's usual standard; and I am therefore, however locally and temporarily, on Anthropic's side of this one exact conflict. Also to be clear, the thing I *want* is that neither Hegseth nor Amodei gets to play mad scientist, and all these companies worldwide get hard constraints and oversight before their AI manages to become fully self-improving. If the Department of War gets a special exception to the rules, they will kill everyone just as dead as any other AI developer. But if that's not presently on the table, then yes, of these two sides in this exact argument, I'd rather that when AI companies say "Our AI is not ready" that the goverment listens to them or at most picks a different contractor, rather than saying "How dare you tell me what to do." Could the AI company be lying and throwing their power around? Sure. They could also be telling you the truth about how what you want to do has unknown and potentially bad consequences. You don't actually want to get into the habit of overriding scientists about that, when they are messing with potentially world-destroying forces -- if you're going to make the mistake of letting the scientists do that in the first place, that is. I don't intend to play the game where I pretend to be your friend. I don't expect to be good at that game compared to a more enthusiastic player. But I have tried to talk straightly about this matter.

Poore's Law: motorola budget android phones double in performance every decade. seriously, I'm not joking. take a moto g4 from 2016 and a current g05 and compare. holds for almost every spec.