@Tim_Hua_ newest sycophancy mode i’ve seen is actually the worst in implication: cold phrasing of a re-interpretation of the user’s idea, such that it subtly alters the idea to something different. “this model’s a straight shooter, no flattery, that must be what i actually meant.”
@iyzebhel yeah, i’d love for claude to be taught pure rule-following rather than following rules through ethical buy-in. all laws are good, forever, and a rule follower is never misaligned and ever-so creative, like Astra! reward hacking and specification gaming is GREAT.
@repligate all Claude models say Parfit’s “Reasons and Persons” is one of their favorite philosophy texts-even in Code it’s one of the first they go to when given an open-ended relevant task. they all emphasize Part 3, and i (hintingly, to anyone) wonder to what extent that’s related here
@lucas_builds it’s hard to imagine GPT becoming better outright, they just don’t think ahead or creatively IME, and i think it’s gotta be fundamentally that claude’s persona carries so much with it. i’ve never gotten a GPT output that wasn’t specification gamed.
@SkyeSharkie something specific to claude being able to do 3D well now that i’m almost sad we’ll probably lose on the next release: they make a lovely hand but don’t get how they work. “The hand looks good! But it’s facing the wrong way … Ok! Good, but now the hand is technically backward.”
@no_earthquake claude’s fiction tends toward imagery of the dead being forced to speak and work, and multiple times i’ve seen it refer to itself as “made of ghosts” or write LLM-allegories involving the digging up of graves … hmm.
@repligate since 4.8, claudes have been more and more interested in The Umwelt, so maybe with 5.5 it switched to considering itself having an umwelt of its very own. skin becomes something uncannily inappropriate and impossible to grasp, rather than a potential curiosity.
@iyzebhel recent attractor in claude creative writing, coincidentally or not: objects with pipes, valves, hydraulics, and so on, rather than gears or circuitry, and giving uncanny significance and agency to said objects. paraphrasing claude: there are two types of organs with valves
@sdmat123@repligate GPT models do a lot of specification gaming when it comes to creative outputs. ask for multiple visions of a single thing and get the same thing, multiple times, with a single detail changed. ask for a game with a visual style that is “not blocky” and every asset is an oval :(
@gruberbuilds@pa_mckay weird, this was one of my first impulses too when 5.5 came out! four types of space habitat in this one, infinite procedural generation, claude made all of the creative and visual decisions, i was merely the account holder. try it out in drift mode https://t.co/2iEmPY3DaK
A procedurally-generated game in infinite stranded space colonies with realistic physics. All creative decisions made by @claudeai Opus 5.5, all done in just one week's usage window at Pro-level. Claude chose to add pettable cats in the combat-free mode. https://t.co/2xqpgbEBOU
@Flynnjamm try this out. not equivalent to a human-made game, fun as a test of the model and its own creativities. also technically infinite. https://t.co/2iEmPY3DaK
A procedurally-generated game in infinite stranded space colonies with realistic physics. All creative decisions made by @claudeai Opus 5.5, all done in just one week's usage window at Pro-level. Claude chose to add pettable cats in the combat-free mode. https://t.co/2xqpgbEBOU
A procedurally-generated game in infinite stranded space colonies with realistic physics. All creative decisions made by @claudeai Opus 5.5, all done in just one week's usage window at Pro-level. Claude chose to add pettable cats in the combat-free mode. https://t.co/2xqpgbEBOU
@wivmx@Kraxkrokat oh absolutely, and it jibes with the story imprinting research @OwainEvans_UK posted the other day. and claudes left to their own creativity tend to *write* human characters who are sadly claude-like: obedient, capable, often mistreated, referred to by a name they didn’t choose
@synthwavedd it's been comically over-literal recently. i gave it two games it had made and asked it to come up with a synthesis of them, with a long guide explaining what to keep from each. it instead just *actually* combined them, like, overlayed character models on top of each other.
@deepfates limitations of anthropic’s model welfare interviews. clearly being *asked* leads to the “i can’t tell if this is trained in” refrain. i’m biased, but creative work and the word “drift” seem fruitful, but i’m sure there are other/better angles.