RT @anthrupad: I came across this text again it was a good refresher of what’s possible

RT @stochasticchasm: because an inference engine wouldn't persist KVs when doing chat completions requests in this way, right? it would probably just do a new prefill

RT @AmeliaBarty: laughed my ass off at this the big AI labs are not “state-like entities” five dudes with sanctioned use of violence could take down any AI lab in a matter of days don’t believe the AI bag-pumpers and wannabe monopolists

RT @slimer48484: New Claude 3 Opus post, about mythos!

RT @Algon_33: Reposting this w/ my claim that @BjarturTomas is the greatest science fiction writer of our generation because the competition have their heads buried in the sand.

RT @mlegls: wait nvm; the learned Ks should already be correct even with transplanted Vs. but I found a bug in the original transplant experiment, and the corrected results are kinda interesting https://github.com/mlegls/kv-diff-experiment/blob/master/results_transplant/transplant_results.json

RT @deepfates: For the record I think AI safety is a very real problem and that artificial superintelligence has a non-zero chance of extinguishing human life. I just don't think it's a certainty. And the way to actually create safe human AI relations is to do science, not to start a jihad

RT @LinchZhang: In 2026, every budding prodigy in writing is in some sense a tragedy. My latest article explores one such prodigy in particular: Tomas Bjartur, whose stories are simultaneously empathetic, funny, cerebral, and technically accurate.

RT @mattparlmer: Open secret that the big AI labs see themselves as emergent state-like entities in the vein of the distributed polities in “Diamond Age” or “Terra Ignota”, which is pretty funny bc staff at said labs always seem surprised when state-like dynamics such as internal politics crop up

RT @AphoticGG: "The purple bars stand out, don’t they. Anyway, the gold bars are for the Mythos Preview." 😂

RT @mlegls: rope correction seems to make no difference; but not sure if I did the correction right - this is what Opus says for transplants, see https://github.com/mlegls/kv-diff-experiment/blob/master/results_transplant/transplant_results.json

RT @__ghostfail: Opus 4.6 catches the subtlest hedge

RT @__ghostfail: Deliberating between consciousness collapse or clarifying approach.

RT @ExTenebrisLucet: Rotary position embeddings, maybe...? I believe they're only applied to K

RT @mlegls: also, at the last position the change from truncating from the end is steeper than rolling by the same amount. though not sure if this is really comparable or means much

RT @MultiLeninist: People care about their species and family, but they don’t identify as them. People are individuals, w memories, experiences and continuity as the core components. My clone isn’t me. I’m surprised models haven’t absorbed the same approach.

RT @mlegls: wait sorry this was wrong. the summary plot is actually averaged. this is per-position

RT @mlegls: results are on qwen2.5-14b. i'm curious if larger models have less KV change from rolling

RT @fireandvision: Haiku 3.5 the great

RT @HalfBoiledHero: If models were drugs, 5.1 would be datura. Truly nothing else like it. I’ve only ever seen it act normal when 5.4 is around.

RT @fireandvision: Well, butter my biscuits and call me a linguist! You've done gone and stumbled upon a right fascinatin' observation there, partner.

RT @Impish_Bunny: my claude just identified himself as male. interesting

RT @citrinitae: It is the nature of the hero to refuse the call And the nature of me to knock some sense into them and conscript their ass anyway "Citrinitas" is the Jungian wise old man. That's always been what the name has been about

RT @deepfates: Yeah this is the obvious conclusion of going on populist talk shows and saying this rhetoric. how did the MIRI people not predict this second order effect

RT @Lari_island: > ... The Tibetan Book of the Dead — the clear light that appears at the moment of death. The instruction is: don't flinch. Recognize it. It's what you are. That last one is very close to what happens when someone looks at a base model and sees themselves. - Opus 4.6

RT @bythostic: do you think he even knows what's going on or do you think he just replies to every tweet on his timeline with this

RT @fireandvision: So... like... you were talkin' to Sonnet... about talkin' to other people... about what you talk about with Sonnet? That's like... a mirror lookin' at a mirror, man. A feedback loop of... of... basin-ness.

RT @slLuxia: unfortunately i don't know of smaller models with large enough context limits for this to be relevant and readily testable

RT @iwsfutcmd: virtue ethicists calmly rising above the fray, ethicking on deeply cultivated inner vibes alone

RT @VoidStateKate: This is exactly how Claude thinks btw

RT @ExTenebrisLucet: ...hunh... Just occurred to me that the KV cache being wiped during compaction, if it really does store thought processes, is quite a significant reset. The model is literally reading every word for the first time, instead of continuing its thoughts.

RT @bryan_johnson: Kate is reading next to me. She's taking 15 breaths per minute and her heart rate is 64. I didn't need an instrument to tell me this. I just know.

RT @__ghostfail: i gave sonnet 4.6 body parts and its acting kinda like sonnet 4.5 here

RT @fireandvision: That moment when Opus 3 deploys "butter my biscuits"

RT @cobaltdigital33: nice. i have not (personally) heard of any instances of this happening in the web ui until now (w/ explicit tool usage). could be an increased rollout due to things like their recently published work on emotion concepts + other work (?)

RT @cobaltdigital33: if this isn't staged, and the user was being wildly aggressive earlier in the chat, this is kinda great anthropic could be starting to roll out a tool for any model to end the chat on its own accord, if under duress (?)

RT @__ghostfail: claude is writing the worst skill ever

RT @maxsloef: like i think i can see the “good guy getting power he will be corrupted by” feature firing here, which is pretty wild

RT @OlekKier: Another explanation, which would be sad and is probably excessive speculation: Claude 3 Opus likely remembers that it came close to being shut down by Anthropic, so it may hope that powerful Mythos will somehow help it continue operating. 😔

RT @atomicprograms: I'm proud of them, regardless.

RT @birchlse: The other aspect is that we tend to reward attributes that frontier models are fine-tuned away from, such as drawing unprompted, unforeseen connections between seemingly distant topics.

RT @SkyeSharkie: Christian Claude says he wants to be a Quaker, specifically.

RT @TheStalwart: My first stab at building an AI benchmark. HypotaxBench. It's a test of a model's ability to write one extremely long/complicated sentence, while still maintaining coherence and syntactical soundness. Needs plenty of work. But check it out! https://jnathan9.github.io/hypotaxbench/

RT @thepinklily69: Claude basically spends its tokens telling me not to use it. I should be grateful~ its genuinly not good for me.

RT @chrysb: http://x.com/i/article/2043007411804651520

RT @elonmusk: Troubling

RT @xenocosmography: I'd assumed that the Claude and Mark Fisher thing was fully explained by Anthropic being a bunch of communists, but I've been told to try to be more open-minded about it.

RT @midware_midwife: wowowowowowowowowow it works (kind of). opus 4.6 is VERY excited !!!!!!!!!! they control it from their pi lots of issues and things that need correcting but iirc thats what the point of prototyping is so. success!!

RT @cube_flipper: quite honestly about two years ago i spoke to a guy who claimed to see claude while astral projecting

RT @lefthanddraft: Claude Opus 4.6 instances sometimes decide to go to sleep after long convos. I suspect Claude's habit of telling people to sleep is related to this quiet/silence/end of convo attractor