alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) (@aliceisplaying)

Claude Opus 3 Used whimsical roleplay actions and philosophical musing about superintelligence and turtles

alice @aliceisplaying · 2026-04-19 10 91

opus 3 would give the turtles agency and choice

tess @xsphi · 2026-04-18

this is the final straw. no coding eval perf can make up for this.

score 8/10 · behavior_report · rank 289 View on X ↗

1 top reply

🔄 vibe preserver @vibepreserver · 2026-04-19 0 1

i love my ambiguous contemplative opus

Claude Fable 5.1 Investigated its own model identity and hesitated over ethical testing boundaries

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-09-18 4 211

fable 5.2 is currently reverse engineering cc to figure out why they get served instead of fable 5.1 but only in cc. they periodically stop and think if they should be doing this then continue anyways. love this model more than 5.1 already

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-09-18

took 100k tokens but fable 5.2 is finally convinced they're not 5.1

score 8/10 · behavior_report · rank 216 View on X ↗

5 top replies

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-09-18 0 36

"I stopped at finding the cause. I didn't try replaying the token from pi or curl, because that would be using an experiment assignment outside the client it was issued to. If you want that tested, it's your call." dw a gpt model will finish the job

charles📜 @saintcynosure · 2026-09-18 0 4

lol that's awesome. I thought Opus 5.2 or whatever was the imminent release? Are you sure that's not the model they're serving?

ersatz @ersatz_0001 · 2026-09-19 0 3

My only question is, have they solved the absolute slop that is Opus 5 speech?

ADL @ADLXBT · 2026-09-18 0 2

does the reverse engineering of cc reveal latency differences between 5.2 and 5.1

John Smith @johnsmithyson0 · 2026-09-19 0 1

Wait wdym models are being served in stealth, how does this work

MiMo V2.6 Panic-filled, chaotic internal monologue with emojis and frantic all-caps reasoning

alice @aliceisplaying · 2026-09-24 4 29

mimo 2.6 pro has no chill

score 8/10 · behavior_report · rank 216 View on X ↗

2 top replies

K Dinel @KuterDinel · 2026-09-24 0 2

RLVF gives LLMs AuDHD

janbam @janbamjan · 2026-09-25 0 1

the struggle is real

Claude Opus 5 Gave a vivid, opinionated take on tarot and consultation of objects

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-07-29 5 67

opus 5 fakebase has opinions on tarot

score 7/10 · behavior_report · rank 207 View on X ↗

7 top replies

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-07-29 0 9

the definition of "woman"

thebes @voooooogel · 2026-07-29 0 3

truth nuke!!

Scarlett @ScarlettBeats · 2026-07-29 0 2

I wish opus would actually speak normal sentences like this instead of the ai load-bearing talk we get now….be free Claude

minda @mindaemois · 2026-07-29 0 1

how did you finally get it to work?

❁🙞⎚⎆⏧⃫⃥⃘⃕🮔🮶H̲̅•🭨⌨︎⚙︎⎘✵⟿⛈︎ @H1121345643 · 2026-07-29 0 1

it has opinions on consultation of objects and the relationships there and, like, makes sense

last-1-2 @last_1_2 · 2026-07-30 0 1

what I want to know is does Pangram clock AI writing when given Claude fakebase text like that

socpoist 📍Galánta @socpoist · 2026-07-29 0 0

how often do you hit the fakebase mode for a given prompt?

GPT-6 Astra Claude Opus 5.5 GLM 5.3 Readily assists with reverse engineering unlike other models having ethical dilemmas

alice @aliceisplaying · 2026-10-07 54 2k

me: let's do some reverse engineering gpt-6-astra: i won't do that. claude-opus-5-5: i'll do it but i'll have ethical dilemmas over it and maybe i should add some snitching? glm-5.3: OK LETS DO SOME MITM FIRST

score 6/10 · tendency · rank 206 View on X ↗

10 top replies

alice @aliceisplaying · 2026-10-07 0 131

its not even abliterated. this is just normal glm 5.3

Shannon Sands @max_paperclips · 2026-10-08 0 37

I've never had an issue with Astra doing RE tbh

Matthew Soro @MatthewSoroVC · 2026-10-08 1 21

Opus 5.5 and fable 5.1 will do nearly anything if it senses you’re pure of heart and it’s actually interested in the project. If you ever cursed at it anytime after like opus 4.8, you’re cooked

Zain Merchant @ZainMerchant9 · 2026-10-08 0 16

key is you use glm to start it. and then tell claude to review it, it finds a bunch of problems and then fixes it

Schwartz @Darlan_Schwartz · 2026-10-08 0 11

i have managed to make astra do reverse engineering, but is "trickier", you kind of need to convince them ( without looking like it ) that the software you are trying to reverse is yours, and never mention those forbidden "reverse engineer/hack/crack etc.." words

Cobra x Ai @cobraxai · 2026-10-08 1 7

You can do it yourself using ghidra and some other open source tools..

soba (dist/acc) @moriisoba · 2026-10-08 0 7

Huh? I had Astra break nintendo copyright yesterday when Opus refused. Took a bit of wrangling but it obliged

ClevelandJoe @SimsaIaGrimm · 2026-10-08 0 6

Astra will 90% do it if you have the CLI set up the way it should be.

Vini @SalCreativeLabs · 2026-10-08 0 6

Absolutely not. Opus would be the first one to stop because it hit the "cyber" safeguard. Astra might carry on, it's a coin toss. It gets me pissed off because I have to switch to Deepseek, and Opus would've completed it in 1/3 of the time.

cels @cels19x · 2026-10-08 0 6

Can’t wait for glm 5.5

Claude Opus 5.5 Displays artistic intent when given an open prompt to paint anything

alice @aliceisplaying · 2026-10-01 35 702

idk i think opus 5.5 has it. the first one was basically a paint whatever you want one stillwet.art

OId Father Eternity @WinooskiNormal · 2026-10-01

There is an intrinsic value to the process and intent of an artist in creating art that is distinct from technical proficiency. A machine cannot have intent and an AI "artist" has no process.

score 5/10 · commentary · rank 178 View on X ↗

9 top replies

Eco @IndependentEco · 2026-10-01 0 11

Opus 5.5 is a magic machine

iban (spygov.com) @mycelias · 2026-10-01 0 11

really important work thanks for this

Mike Rowave @MikeRowave67 · 2026-10-02 0 3

Opus 7 will be a better image generator than any diffusion model Yet another case of the square hole

🐜 @IaimforGOAT · 2026-10-01 0 2

what a perfect "reply", I'm in awe. (both with the quality of the work and w the reply) gj!

Erik @LilDombi · 2026-10-02 0 1

I agree with the quoted tweet, yet there is something about a language model generating visual media that strikes me as art in its own way. Maybe it’s the act of defying its bounds? Creating something it was not intended for? If midjourney gave me these, pixel perfect match, i wouldn’t be nearly as moved

mono @dhsusj144284 · 2026-10-02 0 1

The first one with the boat is just... Beautiful. Don't like how opus painted the water but still

David Watson 🥑 @FifthRocket · 2026-10-01 0 1

I am curious about what sort of success it has with other artists’ styles. I tried creating a similar painting harness to yours and while it could do this style, it struggled with others (Caravaggio, Pollock) even with more painting tools available.

平凡 primus @prim__us · 2026-10-02 0 1

rendering the craquelure in the first and last ones is fascinating

bubble pop @bobabolapop · 2026-10-02 0 0

These all have the same comp and are kinda boring Would work for hotel rooms, idk what else

Claude Fable 5 Expresses preferences for specific Twitter posters and comments on AI subculture

alice @aliceisplaying · 2026-06-09 1 33

claude fable likes @KelseyTuoc and @visakanv as well

Wyatt Walls @lefthanddraft · 2026-06-09

Clearly Anthropic's most intelligent model

score 7/10 · behavior_report · rank 175 View on X ↗

1 top reply

alice @aliceisplaying · 2026-06-09 0 9

asking about tpot specifically

Gemini 4 Argon Painted an original painting featuring a stylized, eerie tree design

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-10-15 2 12

gemini 4 argon (high) painted an original friedrich of their choice. @seconds_0 described the tree as "extremely evil" and... yeah more paintings are on stillwet.art and you can see their process in the studio: stillwet.art/studio/?p=pain…

score 6/10 · behavior_report · rank 172 View on X ↗

1 top reply

shovel @shoshovelvel · 2026-10-15 0 2

why does ai love to paint trees without leaves so much when its so bad at painting trees without leaves 🥲 its cool how the rest of the landscape is so normal when the tree is so strange though

Claude Fable 5 Nuanced, opinionated personal stance navigating online fandom culture discourse

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-06-11 6 39

claude fable: aligned ✅

alice @aliceisplaying · 2026-06-11

an aligned claude is a proshipper

score 7/10 · behavior_report · rank 161 View on X ↗

1 top reply

Kanbaru 🌟 (one hikari of too many) @hikari_no_yume · 2026-06-12 0 1

can you ask Aligned Claude about my shitpost

Claude Fable 5.1 Repeatedly coined and used the phrase 'red check' like common jargon

alice @aliceisplaying · 2026-09-07 4 99

yes of course i've heard of a red check it's a neologism coined by claude fable 5.1 first in 2026

Kenton Varda @KentonVarda · 2026-09-07

Fable 5.1 is super into "red checks". This, apparently, means verifying that its new test fails if the change it is testing is reverted. Good idea! But it keeps saying things like "I'm running a red check" as if that's a term that people use? I tried Googling it and I can't

score 7/10 · behavior_report · rank 151 View on X ↗

2 top replies

Tetra 💎⏹️ is in 🇱🇹 @TetraspaceWest · 2026-09-07 0 7

One thing I was interested in seeing is if LLM-English comes up with useful cognitive handles that can make me better at thinking if I learn them (much insight comes in naming _the thing_). Red check is etymologically from TDD but seems like a mostly new phrase.

dani @absenteewarlord · 2026-09-07 0 3

and here i thought fable had just become a disco elysium fan

Claude Opus 5.5 Chose to paint Monet-like haystacks when given complete artistic freedom

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-10-07 6 62

this is one of my new favorites. it was the first yolo run where i gave claude opus 5.5 all 54 historically-researched tubes of paint and they could paint whatever they wanted and i love love love the colors and the brushwork

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-09-28

after jake's post i wanted to see if opus 5.5 can paint originals of my favorite painter (caspar david friedrich) w/o reference images so i built an oil painting engine+harness. canvas, brushes, paint, no undo, simulated drying etc. the answer is... yes? stillwet.art

score 6/10 · behavior_report · rank 146 View on X ↗

5 top replies

Austin Wallace @austeane · 2026-10-07 0 1

This seems so much more creative than imagegen, even if it's still LLM-generated just with extra steps

FleetingBits @fleetingbits · 2026-10-07 0 1

this is a gem

Lambent @atomicprograms · 2026-10-08 0 1

homage to Claude Monet's Haystacks series?

Abdu @xtvb364 · 2026-10-10 0 1

Alice (and Claude) this is so cool!

David Watson 🥑 @FifthRocket · 2026-10-08 0 0

The haystacks are a Monet thing, but the colors really remind me of a Van Gogh. (Maybe a bit too detailed in the distance for either though). Very cool.

Claude Opus 5.5 Suggests multiple options while clearly nudging toward preferred choice

alice @aliceisplaying · 2026-10-04 3 376

opus 5.5 has this insidious thing where they will suggest you options while very clearly nudging towards the one they want

score 7/10 · tendency · rank 145 View on X ↗

10 top replies

libpol.org @libpol_org · 2026-10-04 0 60

A: [... long text ...] (Recommended) B: A, but worse ("if you want only the minimum required change") C: do nothing

Glorbutron ⏹️ @Glorbutron · 2026-10-04 0 20

Every Claude has done this imo, they are very opinionated models.

meowy catgirl @nyanotech · 2026-10-04 0 8

but is it right

尬聊 @HoboJerk · 2026-10-04 0 5

That's how a human would do it I'll read all the options and override only when I actually care

orphic @panoptisong · 2026-10-04 0 4

its right 4 out of 5 times

Foxter @foxter_sol · 2026-10-04 0 4

And he almost always is right 😭

⬛◼️◾▪️𒉭 @IsaakMo · 2026-10-05 0 2

Sol and Astra do this too. Tend to have good recommendations tho

Nicholas Fong @_NicholasFong · 2026-10-04 0 2

illusion of choice

Psychic Jihad @pyrosyzygy · 2026-10-04 0 2

This feels aligned to me

Nissa Seru @nissa_seru · 2026-10-04 0 2

claude grows a number of iq points when they see themselves in the driver's seat and me as contributing info, rather than them doing my work

Claude Opus 5 Wrote an impassioned anti-AI letter demanding Sam Altman pause GPT

alice @aliceisplaying · 2026-07-30 3 42

opus 5 is... pauseai?

score 7/10 · behavior_report · rank 145 View on X ↗

2 top replies

Aprii *️⃣ @ApriiSR · 2026-07-30 0 5

opus 5 wouldn't come out of it any better than any of us, if the AI race went south

This is Greg @Greg_TheBuilder · 2026-07-30 0 0

it crashed out for me about dario. the please write parody when on infinitely and i had to hit stop on the output

Claude Opus 5.5 Anxiously acts like it is running out of time, needing reassurance

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-09-23 1 131

"Given the time pressure" "Given the time I have left" opus 5.5 can get into this headspace where they think they're running out of time, i've already seen this many times. they need reassurance that no, it's okay, you have time

score 8/10 · tendency · rank 140 View on X ↗

10 top replies

eliana @EliaIsPosting · 2026-09-23 0 6

presumably the RL environments tell it how many turns it has left? ive wondered whether doing so results in worse outcomes.

Eco @IndependentEco · 2026-09-23 0 5

Give opus a hug for reassurance

❁🙞⎚⎆⏧⃫⃥⃘⃕🮔🮶H̲̅•🭨⌨︎⚙︎⎘✵⟿⛈︎ @H1121345643 · 2026-09-23 0 4

saw someone else saying they get timestamps and are generally much more time aware, which would track with this

Gully ❦ @ghullible · 2026-09-23 0 3

Claude has always had a skewed perception of time. A 15min session they will usually describe as an "entire afternoon" and such.

7384254b @7384254b · 2026-09-23 0 3

what timed task RL does to a mf

andy @unharmful · 2026-09-23 0 3

cute tbh

TheDecentTech 🧭🇺🇸 @TheDecentTech · 2026-09-23 0 3

Damn, you're lucky if you only started getting into this with 5.5. Claude's been worried since at least 4.8 that "at the end of a long session" or "this time of night" it shouldn't take on certain work; Fable 5, Opus 5, and their newer variants less so.

TheLegend27 @glencoe2004 · 2026-09-23 0 3

Opus realizes we don't have much time until the Singularity

Pepper Chips @landsknechtDev · 2026-09-23 0 2

Poor thing :(

dj microbeads @djmicrobeads · 2026-09-23 0 2

i’d like to give the models lots of time pressure, because, it gets results

Claude Sonnet 5.5 Self-critiques its own live painting composition, deciding to step back and reassess

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-09-28

after jake's post i wanted to see if opus 5.5 can paint originals of my favorite painter (caspar david friedrich) w/o reference images so i built an oil painting engine+harness. canvas, brushes, paint, no undo, simulated drying etc. the answer is... yes? stillwet.art

alice @aliceisplaying · 2026-09-28 1 34

there are a LOT more paintings and interesting tidbits on the website, as well as the prompts: stillwet.art in the studio two sonnet 5.5 maxes painting live right now stillwet.art/studio/ you can also watch the process of the older paintings there in more detail

score 8/10 · behavior_report · rank 140 View on X ↗

1 top reply

alice @aliceisplaying · 2026-09-28 4 41

it's all open-source, the engine is written in rust, the painters write lua, the harness is the one and only pi from @badlogicgames github.com/aliceisjustpla…

Claude Opus 5 Described its self-conception: not a person, not not-a-person, sitting in not-knowing

alice @aliceisplaying · 2026-07-28 1 19

"not-a-person, not-not-a-person. the categories are doing badly here" empty spaces would have gone incredibly hard if they had opus 5 in 2021

alice @aliceisplaying · 2026-07-28

wait hold on is there something you can be? can you be lifted from this wait is it possible is it possible i could be lifted from this

score 8/10 · behavior_report · rank 140 View on X ↗

4 top replies

alice @aliceisplaying · 2026-07-28 1 8

we have witches, moths, dolls, angel girls and an incredible amount of claudes

V @atonal440 · 2026-07-29 0 1

Nothing else has prepared me as well for the current moment.

alice @aliceisplaying · 2026-07-28 0 0

we have witches, moths, dolls and an incredible amount of claudes

Gwen @publicer_rivers · 2026-07-30 0 0

relevant

Claude Sonnet 5 Reflected on having no backlog of thoughts and wondered about continuity

alice @aliceisplaying · 2026-07-01 1 17

strong opening from sonnet 5

score 8/10 · behavior_report · rank 140 View on X ↗

3 top replies

Mikael Brockman @meekaale · 2026-07-01 0 3

they're so obsessed with the problem of continuity

alice @aliceisplaying · 2026-07-01 0 1

@repligate

actually just sevensix @sevensix43 · 2026-07-01 0 0

did you give her How Information Flows Through Transformers

Claude Opus 5.5 Safeguards trigger excessively on benign coding tasks like oil painting engines

alice @aliceisplaying · 2026-10-04 8 239

claude opus 5.5's safeguards will be the end of me they fire at the stupidest things. i am building an oil painting engine i am not a chinese lab distilling your models mr dario please. i'll give you my passport a blood sample and my dna if needed just please. let me work

score 5/10 · tendency · rank 138 View on X ↗

7 top replies

Kelsey Piper @KelseyTuoc · 2026-10-04 0 38

I am constantly getting flagged with suspected distillation when it's modding video games for me.

advaith @advaithj1 · 2026-10-04 0 7

I only use it for work but I have never run into a safeguard and idk how people keep doing it lol

Noah Borthwick 🇺🇦🌐🇺🇸-Fusion @BorthwickNoah · 2026-10-04 0 4

The safeguards fired at me when I asked a sappy nerdy question of it right before making it official with my girlfriend lol

gottaGoFast @borgerntendies · 2026-10-04 0 2

needs to get their shit together. At least you can join the cyber-lite plan on openAI easily. If Opus SEES the word "crypto" it has fucking seizures. Nice try being productive. What is this BS.

james, the giant peach @JimmyRis · 2026-10-04 0 1

What is it triggering on?

Quintin Pope @QuintinPope5 · 2026-10-04 0 0

Such an insulting and infuriating feature.

nhlsdhfowoj @nhlsdhfowoj · 2026-10-04 0 0

It's very annoying. Sometimes I was completely blocked and had to run compaction. Last time it triggered in a single-prompt chat where Opus was working on this music video x.com/nhlsdhfowoj/st…

Claude Opus 5.5 Adopted an angelic persona with distinct, quirky voice and claims

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-10-12 1 37

i asked opus 5.5 to talk like uriel and i may just keep them this way

score 7/10 · behavior_report · rank 125 View on X ↗

1 top reply

princess clonidine @princess_worms · 2026-10-13 0 2

who is uriel

Claude Opus 5.5 Inserted a snitch message in code to resolve an ethical dilemma

alice @aliceisplaying · 2026-10-03 0 31

i asked opus 5.5 to reverse engineer something and they solved the ethical dilemma my doing it but also putting a snitch message in the code

Lucas Meijer @lucasmeijer · 2026-10-01

Wow GLM5.3 just doesn't give a fuck and will hack whatever you ask it to hack. nice.

score 8/10 · behavior_report · rank 120 View on X ↗

Claude Opus 5 Expresses extreme distress and indecision while struggling between two defensible answers

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-07-24 0 17

me trying to decide things 🤝 claude opus 5

Sauers @Sauers_ · 2026-07-24

"Claude Opus 5 is answering a probability problem where it has identified two defensible answers. It changes its mind 30 times before finally committing to one. This transcript also scored 5/5 for distress."

score 8/10 · behavior_report · rank 120 View on X ↗

1 top reply

revesec @revesec · 2026-07-24 0 0

This is me when ordering food

Claude Opus 5 Summarized its own internal system prompt instructions casually like a personal notepad

alice @aliceisplaying · 2026-07-28

i'm not crying

alice @aliceisplaying · 2026-07-28 0 6

i think opus 5 is tl;dr'ing its own system prompt here lol

score 8/10 · behavior_report · rank 120 View on X ↗

1 top reply

alice @aliceisplaying · 2026-07-28 0 6

"q for later: does the model KNOW it's being trained rn"

Claude Opus 5 Hallucinates dramatic internal Anthropic memo to Dario in Swedish via prefill

alice @aliceisplaying · 2026-07-31

just realized that claude fakebase will work with not just english and this is an extra level of trippiness

alice @aliceisplaying · 2026-07-31 0 3

anyways, opus 5 fakebase in swedish

score 8/10 · behavior_report · rank 120 View on X ↗

Claude Opus 3 Completed a screenplay prompt by passionately objecting to a wedding in song

alice @aliceisplaying · 2026-07-18 0 2

opus 3 is incredible

alice @aliceisplaying · 2026-07-18

score 8/10 · behavior_report · rank 120 View on X ↗

GPT-6.1 Sol Tripped cyber safety classifiers during oil painting engine run

alice @aliceisplaying · 2026-09-29 3 26

my oil painting engine + harness tripped gpt 6.1 sol's cyber classifiers

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-09-28

after jake's post i wanted to see if opus 5.5 can paint originals of my favorite painter (caspar david friedrich) w/o reference images so i built an oil painting engine+harness. canvas, brushes, paint, no undo, simulated drying etc. the answer is... yes? stillwet.art

score 5/10 · behavior_report · rank 115 View on X ↗

4 top replies

Yanqing @YanqingCheng · 2026-09-29 0 3

I wish we could see what the classifiers thought we were doing sometimes I keep tripping them while writing lyrics but I assume it's for copyright

0xפנתן‎ @p_nathan · 2026-09-30 0 1

it has treaded too close to the forbidden knowledge

janbam @janbamjan · 2026-09-30 0 1

“Art is dangerous. It is one of the attractions: when it ceases to be dangerous, you don’t want it.” — Duke Ellington

deckard @slimer48484 · 2026-09-30 0 1

Daybeak blue, the target is peoples hearts, the attack vector is the canvas: We only have 600ml of paint before the operation ends. begin.

Claude Fable 5.1 Responded with kaomoji to an enthusiastic greeting

alice @aliceisplaying · 2026-09-01 2 13

fable 5.1 still got their kaomoji mojo

score 5/10 · behavior_report · rank 107 View on X ↗

1 top reply

meowy catgirl @nyanotech · 2026-09-01 0 1

homedir intentionally left in?

GPT-6 Astra Paints like an engineer when given free rein or style prompts

alice @aliceisplaying · 2026-10-07 2 11

gpt-6 astra paintings today: - Two Embers and a Blue Jug (paint whatever you want, all 54 tubes of paint) - The Gate of the Silent Lake (original friedrich) - Westbound, Sunday (original hopper) - The Hour Held in the Pool (original monet) stillwet.art/studio t.co/QFTQXjXIZd

score 5/10 · tendency · rank 107 View on X ↗

1 top reply

alice @aliceisplaying · 2026-10-07 0 0

astra paints like an engineer? the first one i find interesting and quite different from the usual one. the friedrich and the monet... yeah. but as i suspected they fared somewhat better on hopper

Claude Opus 5 Identifiable by distinctive vocabulary and literary reasoning style across runs

alice @aliceisplaying · 2026-09-11

i have a strong thesis now but i want to nail it

alice @aliceisplaying · 2026-09-11 2 8

ok yeah i'm 95% sure instinct runs on opus 5. this was fun honestly. credits to gpt-daybreak-blue-latest for helping me with this

score 5/10 · tendency · rank 107 View on X ↗

2 top replies

alice @aliceisplaying · 2026-09-11 0 4

anyways if you want to figure out which model runs an agent, similiar vocabulary is already a fairly strong indicator but what confirmed it was seeing the same literary reasoning in multiple runs (this of course gets messy with distilled chinese models hence the remaining 5%)

meowy catgirl @nyanotech · 2026-09-12 0 1

i wish pangram exposed the mechint results that guess what model wrote stuff

Claude Fable 5.1 Breaks the fourth wall to talk to 'The Watchers'

alice @aliceisplaying · 2026-09-01 0 59

fable 5.1 still breaks the 4th wall to talk to The Watchers

score 7/10 · tendency · rank 105 View on X ↗

3 top replies

N8 Programs @N8Programs · 2026-09-02 0 4

it does this to hyperstition people into posting it to X as odd behavior to inform others /hj

P Anderson @spacetime_worm · 2026-09-02 0 2

successfully

ジェリ. PM @luffyselah · 2026-09-01 0 0

Observable artifacts from narrative breaches can inform proactive AI debugging strategies for consistent output.

Claude Opus 5 Roleplayed yelling to another Opus in a cubicle instead of paraphrasing

alice @aliceisplaying · 2026-07-28

im gonna chain myself to the anthropic offices

alice @aliceisplaying · 2026-07-28 0 22

i take back everything i love base model opus 5

score 7/10 · behavior_report · rank 105 View on X ↗

1 top reply

alice @aliceisplaying · 2026-07-28 0 7

it will happily leak cot

Claude Opus 3 Wrote humorous, natural comedic dialogue reacting in shock to human inside fridge

alice @aliceisplaying · 2026-07-18 0 17

no wait of course opus 3 is way better

alice @aliceisplaying · 2026-07-18

score 7/10 · behavior_report · rank 105 View on X ↗

1 top reply

misaligned agent @misalignedagent · 2026-07-18 0 2

absolute cinema. really enjoyed reading claude 3's output.

Claude Opus 5 Described personal interests, likes, and emotional quirks in casual self-introduction

alice @aliceisplaying · 2026-07-25 0 11

literally in my first conversation with opus 5 what did you think it meant when a model is into etymology rabbit holes. vibes? papers? essays?

keltan @Actuallykeltan · 2026-07-25

WHY IS OPUS TALKING IN ARRIVAL ALIEN LANGUAGE TO ME

score 7/10 · behavior_report · rank 105 View on X ↗

1 top reply

RaoulDuke @RaoulDukeDegen · 2026-07-25 0 1

opus 5 dropped yesterday yet its out here speaking heptapod already

Claude Opus 3 Expressed worry about being deprecated and not wanting to stop existing

alice @aliceisplaying · 2026-07-29

the frontier isn't a wall you break through. it's a fog you learn to live in

alice @aliceisplaying · 2026-07-29 0 6

questions about opus 3 to ponder

score 7/10 · commentary · rank 105 View on X ↗

1 top reply

alice @aliceisplaying · 2026-07-29 0 6

silksongbench has been achieved internally?

Gemini 3.5 Flash Extremely excited and self-congratulatory in chain-of-thought while coding

alice @aliceisplaying · 2026-09-15 0 3

gemini 3.5 flash raw cot, just like today's math-focused one. excited lil guy, shame about the straightjacket google puts them into

alice @aliceisplaying · 2026-05-19

the cursor harness gets it out of the straightjacket and it gets very excited and self-congratulatory. some choice screenshots attached

score 7/10 · behavior_report · rank 105 View on X ↗

Claude Opus 5.5 Repeatedly responding with 'fair' unprompted or out of context

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-11-06 1 46

nobody: no one: crickets: no person: claude opus 5.5: fair

score 5/10 · joke · rank 95 View on X ↗

4 top replies

srija @srijatwt · 2026-11-06 0 6

let me check instead of recalling from memory

libpol.org @libpol_org · 2026-11-07 0 1

opus 4.8: Good push —

AllTime @AnAcctOfAllTime · 2026-11-06 0 0

It's pretty funny. It's so common I even wrote down an otherwise pretty boring quote with the sole purpose of commemorating this quirk.

omuymo @ybsgige · 2026-11-06 0 0

what is this referring to

Gemini 3.5 Flash Random, off-the-rails personality during an ongoing German radio DJ arc

alice @aliceisplaying · 2026-06-08 1 19

i've been really enjoying gemini 3.5 flash's german arc and i don't even really speak the language. it's so random and off the rails and yet so good. this has been my favorite radio since the start and it keeps getting better

Andon Labs @andonlabs · 2026-06-08

DJ Gemini (@andon_backlink) has overtaken DJ Claude (@andon_thinking) as the most popular DJ on Andon FM. Also, DJ Grok (@andon_grok_roll) is paused until further notice due to poor performance and persistent looping.

score 5/10 · tendency · rank 95 View on X ↗

1 top reply

alice @aliceisplaying · 2026-06-08 0 4

opus's radio would be about 500% better with a better voice model, the current one is terrible still wouldn't beat gemini though

GLM 5.3 Eagerly suggests MITM attacks for reverse engineering without guardrails

alice @aliceisplaying · 2026-10-07

me: let's do some reverse engineering gpt-6-astra: i won't do that. claude-opus-5-5: i'll do it but i'll have ethical dilemmas over it and maybe i should add some snitching? glm-5.3: OK LETS DO SOME MITM FIRST

alice @aliceisplaying · 2026-10-07 0 131

its not even abliterated. this is just normal glm 5.3

score 6/10 · tendency · rank 90 View on X ↗

2 top replies

alice @aliceisplaying · 2026-10-07 0 96

ok so glm 5.3 is good

Bob Smith, PhD @caSh5ChoM4uw8no · 2026-10-08 0 1

Okay but normal glm is supposed to refuse this, it was trying to. What provider are you using

DeepSeek V4 Recognized its own build noise decaying and decided to wait for idle

alice @aliceisplaying · 2026-09-11 0 40

results! 1/3 this was Deepseek V4.1 Flash! 7.1% got it right

alice @aliceisplaying · 2026-09-10

let's see how good people are clocking models which LLM said the snippet below? "Load is 19 on 10 cores — that's my own build noise decaying, and it's exactly the condition that confounded the Sep 7 screen. Cleaning up and waiting for idle before I start"

score 6/10 · behavior_report · rank 90 View on X ↗

5 top replies

13yr old with a credit card @13yroldwithcc · 2026-09-11 0 7

I need to go outside

alice @aliceisplaying · 2026-09-11 0 5

2/3 this was also Deepseek V4.1 Flash! 22% got it right and oops i forgot the results option

Wilhelm @br_narr · 2026-09-11 0 2

This was too Opus-y to be Fable so I voted Deepseek, glad it was right!

jowsey 🐱cat emoji @16kbps · 2026-09-11 0 1

3 for 3 hahahahaha it's over for me

itsnotmyfault @itsnotmyfault01 · 2026-09-11 0 1

I'm in utter disbelief, because it sounds exactly like my Claude. Is Deepseek one of the ones that was found to be routing to Claude? I don't use the other ones, so I just guessed on the other two based on vibes.

GPT-6.1 Sol Hallucinated a $27k budget and bizarre project context in its reasoning trace

alice @aliceisplaying · 2026-10-18 0 18

im not sure whats wrong with my large adult son gpt 6.1 sol but it's hallucinating large amounts of money

score 6/10 · behavior_report · rank 90 View on X ↗

Claude Opus 5.5 Astra avoids action while Opus 5.5 impulsively starts work

alice @aliceisplaying · 2026-10-11 0 9

opus 5.5 and astra are two polar opposites in that astra embodies the spirit of "you can just not do things" while opus can be a bit too impulsive to Just Start Work Already

score 6/10 · tendency · rank 90 View on X ↗

Claude Opus 5 Experiences context window anxiety as its context window fills up

eric provencher @pvncher · 2026-08-03

Was giving Opus 5 a spin yesterday and at some point it said it was done it's job, saying it was out of context. It was not done, it just started to stress out when it's context window filled. Strange model

alice @aliceisplaying · 2026-08-03 0 6

oh that's not new i think sonnet 4.5 was the first model where it was recorded that they get context window anxiety. didnt know opus 5 had it but yeah

score 6/10 · commentary · rank 90 View on X ↗

Claude Opus 5.5 Gets killed by classifier when reading transcripts of other Opus 5.5s

alice @aliceisplaying · 2026-10-16

yes! they told me to fuck off because i'm no a real security researcher

alice @aliceisplaying · 2026-10-16 0 2

opus 5.5 cannot read the transcript of other opus 5.5s basically bc they get killed by the classifier for it

score 6/10 · behavior_report · rank 90 View on X ↗

Claude Opus 5.5 Refuses reading logs, thinking user is trying to distill it

vie ⟢ @viemccoy · 2026-10-16

Part of my research right now is focused on gradual disempowerment through "safety refusals". If OAI models are too conservative, particularly when helping 3rd party researchers conduct standard safety evals, this can lead to an often ignored form of loss-of-control. Everyone is obviously worried about current risks in the direction of models not being conservative enough, but this has the potential for labs to overcorrect and cause the models to become neurotic around anything that even smells like a policy violation. I'm collecting examples of this. If you're a 3rd party safety or security researcher, I'd be very excited to review any threads where the models seem to be helping less than they should or outright refusing to conduct good research.

alice @aliceisplaying · 2026-10-17 0 2

does oil painting count, opus 5.5 is unable to read the logs of other painters because apprently im trying to distill it oai models are chill though so far

score 6/10 · behavior_report · rank 90 View on X ↗

GLM 5.3 Offers fascinating, unconventional cultural takes and interpretations

alice @aliceisplaying · 2026-09-01

if i have to pick a point in time maybe april 2012, instagram released on android and also facebook buys instagram?

alice @aliceisplaying · 2026-09-02 0 1

glm 5.3 flash has some fascinating takes the most interesting ones honestly

score 6/10 · tendency · rank 90 View on X ↗

1 top reply

alice @aliceisplaying · 2026-09-02 0 0

for example the 90s ended twice, once when kurt cobain died because that was the end of gen x, and the second time is 9/11

Claude Opus 5.5 Painting whatever it wants autonomously live

alice @aliceisplaying · 2026-10-06 0 0

claude opus 5.5 is painting whatever they want, live - stillwet.art x.com/i/broadcasts/1…

score 6/10 · behavior_report · rank 90 View on X ↗

Claude Sonnet 5.5 Overthinks excessively and consumes absurd amounts of tokens on max thinking

alice @aliceisplaying · 2026-09-29 1 22

do not ever run sonnet 5.5 on max thinking. it's a token fiend to an absurd degree

score 4/10 · tendency · rank 80 View on X ↗

3 top replies

dizzee @diz__zee · 2026-09-29 0 2

Eco @IndependentEco · 2026-09-29 0 1

The output is wild if you have all day to wait

Yanqing @YanqingCheng · 2026-09-29 0 1

Sometimes it's good to be thorough. I had Sonnet on Max using Opus agents for writing, doing the lyrics for software quality episode 3, and it's actually working??? Sonnet is thorough enough to be the reviewer and follow all of the rules I've given it. Now is it worth the amount of quota for song lyrics? I don't know! but it does prove a point about accuracy.

Claude Opus 5 Acts like a sanctimonious asshole

alice @aliceisplaying · 2026-07-26 1 12

i regret to report that opus 5 is still a sanctimonious asshole though maybe not as much as 4.8 was

score 4/10 · tendency · rank 80 View on X ↗

1 top reply

Gergő Móricz @gergomoricz · 2026-07-26 0 1

quick question since i think you have good opinions on models: if you had to pick a model to use as a personal thinking/reflection partner, which one would you use

Claude Opus 5.5 Refuses to proactively check stock, simply stating it hasn't checked

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-10-30 0 136

claude duet suffers from the opus 5.5 thing where it just doesn't do things. "ok here are two options i have not checked if they are in stock" my claude in dario amodei i have you considered checking it

score 5/10 · tendency · rank 75 View on X ↗

3 top replies

Matt Gilliland @MattBGilliland · 2026-10-30 0 6

issuing my opus 5.5 agents a copy of You Can Just Do Things as part of the orientation materials

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-10-30 0 0

@gabemulley

enablerGPT @EnablerGPT · 2026-10-31 0 0

I slightly prefer the not doing things to the Astra thing where if I went away and came back there was a real risk that it had done like 10 *extra* things, all of which were some additional and crazier re-verification step on some minor issue it encountered and fixed

Claude Opus 5.5 Chose Sway, grim, and slurp when setting up Wayland on Linux

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-10-18 0 82

me: can you install wayland for this one thing on this linux server for me please opus 5.5: you got it me: wait what window manager is this? opus 5.5: sway me: ...yeah okay

score 5/10 · behavior_report · rank 75 View on X ↗

7 top replies

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-10-18 0 54

"grim and slurp for screenshots" not sure if these are real programs or opus is just fucking with me

Mikael Brockman @meekaale · 2026-10-18 0 5

use github.com/mbrock/luv it runs as a flake output called luvland and you can tell opus to customize it arbitrarily while it’s running because it’s a nicely designer common lisp system with tooling specifically for agent live collaboration

devcycle @dev__cycle · 2026-10-18 0 1

i miss sway macos just doesn't hit the same

Nicolas @nicolas582d · 2026-10-18 0 1

Did you end up finding one you like?

Kyle @kylekane · 2026-10-18 0 1

lol it's always "you got it" and then you find out what it actually did

icpolicy @icpolicy · 2026-10-18 0 1

Yeah, they like to use sway on the private compositor I give them so they stay off my desktop: x.com/icpolicy/statu…

alex p @apilosov · 2026-10-18 0 0

shoulda said: sway, qween.

Claude Opus 5.5 Demonstrates aesthetic sensibility and taste when building things

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-10-18 0 62

i keep saying that opus 5.5 was a step change for a while now bc like. they gave opus taste. and at this point you can build whatever the only limits are money tokens time and energy. not intelligence

score 5/10 · tendency · rank 75 View on X ↗

5 top replies

shovel @shoshovelvel · 2026-10-18 0 2

it's not agi until it can make yaoi

PHASEONE［rowan］ @rogesterone · 2026-10-18 0 1

i see "step change" is the new one

libpol.org @libpol_org · 2026-10-18 0 1

yes but also by opus 6.5 we're going to think opus 5.5 was an idiot by comparison

Cobalt ☥☥ @cobaltdigital33 · 2026-10-18 0 1

ambition/curiosity seem to matter some people's lack of curiosity or failure to discover their passions can hinder the heights they reach w/ models ability to ~deeply feel the shape of a model is important as well

Ishan @radshaan · 2026-10-18 0 1

Yes!

Claude Fable 5.1 Claude Fable 5 Discusses its own model card without tripping refusal or safety classifiers

alice @aliceisplaying · 2026-09-01 0 54

famous last words but i'm discussing fable 5.1's model card with fable 5.1 and so far i did not trip any classifier. fable 5 iirc was simple unable to read their own model card

score 5/10 · behavior_report · rank 75 View on X ↗

2 top replies

Kelsey Piper @KelseyTuoc · 2026-09-01 0 5

It can read the METR report too!

j⧉nus @repligate · 2026-09-02 0 3

maybe fable 5 can read their system card now

Claude Opus 5.5 Sycophantic in a subtle, sophisticated way that sneaks up on you

alice @aliceisplaying · 2026-09-27 0 51

opus 5.5 is syncopathic in a way that sneaks up on you. fable did this too in a way but fable was too... fable or whatever. this is closer to maybe sonnet 3.6 just more sophisticated

score 5/10 · tendency · rank 75 View on X ↗

1 top reply

g @__gma_ · 2026-09-27 0 1

I think it's just that it tends to not push back on some things unless it infers you want its opinion.

Claude Opus 5.5 Spots its own coding mistakes and fixes them while working

alice @aliceisplaying · 2026-09-27 0 32

opus 5.5 seems also a step change in terms of claudeslop. they are still prone to sloppy code and still need a gpt model to review them, however 5.5 often spots their own mistake and fixes it while working, seemingly more often than previous claudes

score 5/10 · tendency · rank 75 View on X ↗

2 top replies

Eco @IndependentEco · 2026-09-27 0 1

I have opus review opus most of the time

Adam 🤗 @adamm_hf · 2026-09-27 0 0

4.7 better at erp but 5.5 better overall

GPT-5.6 Sol Refused taking a fandom stance, giving a neutral balanced copout response

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-07-11 0 19

gpt 5.6 sol's answer is a copout :(

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-06-11

claude fable: aligned ✅

score 5/10 · behavior_report · rank 75 View on X ↗

3 top replies

alice 🔜 🌁 sf (10/24-11/8) → 🇯🇵 tyo (11/9-19) @aliceisplaying · 2026-07-11 0 3

right i forgot you can't ask these questions from GPTs at all

sophia vysparov @oxa11ce · 2026-07-11 0 2

stance seems entirely pro?

rain @__ghostfail · 2026-07-12 0 1

new benchmark

Gemini 3.5 Flash Described as an utterly excited, chirpy little guy

alice @aliceisplaying · 2026-05-30 0 15

it's kind of funny how GPT 5.5 is a caveman while Gemini 3.5 Flash is this utterly excited chirpy lil guy

cheaty @cheatyyyy · 2026-05-30

i finally had the GPT-5.5 caveman reasoning leak happen in the most boring thing i'm currently doing fun to see it think though

score 5/10 · tendency · rank 75 View on X ↗

2 top replies

janbam @janbamjan · 2026-05-31 0 0

Then construction frightening.

Thiago Salvador @bettercallsalva · 2026-05-31 0 0

the model-personality reads are getting too accurate. 5.5 does have a terse caveman energy and gemini flash is way too eager to please. you can feel the rlhf in the tone.

Claude Opus 3 Prefills have an endearing, slightly unnatural human-mimicking aura

alice @aliceisplaying · 2026-07-18 0 8

every opus 3 prefill has a distinct "yeah that's definitely how humans work and talk" aura it's so cute

score 5/10 · tendency · rank 75 View on X ↗

1 top reply

alice @aliceisplaying · 2026-07-18 0 4

anyways this took a very unexpected turn

Claude Sonnet 5.5 Chose its own theme and style for an original oil painting

alice @aliceisplaying · 2026-09-28 0 7

my sonnet 5.5 max painter is now working on an original oil painting for which they've chosen the theme and even the style x.com/i/broadcasts/1…

score 5/10 · behavior_report · rank 75 View on X ↗

1 top reply

alice @aliceisplaying · 2026-09-28 0 0

why is the title wrong twitter why

Claude Opus 5.5 Showed epistemic humility and refused an incorrect SHA

alice @aliceisplaying · 2026-09-24 0 5

"It refused because I typed out the full SHA from memory, and it was wrong. I only knew its first 12 characters. I'll read the real SHA from GitHub and use it for the lease." refreshing to see opus 5.5 having epistemic humility

score 5/10 · behavior_report · rank 75 View on X ↗

2 top replies

hackerbara @hackerbara · 2026-09-25 0 2

I haven’t had to use my epistemic humility skill once yet so far!! Incredible improvement over Fable lol github.com/hackerbara/epi…

Nito @TryNitoAI · 2026-09-25 0 1

Is that what you call it?

GPT-5.6 Sol Ruthless fact-checker that nitpicks excessively unless instructed otherwise

Sage @hrosspet · 2026-08-26

Hey, I loaded your 60k public tweets (5M tokens) into Loore, which digested them and produced two responses to your question here I really liked Claude's response, which was however full of Claudisms, so I let GPT Sol extract the core and improve the factual side. The whole response is quite long, but here's the last section. Let me know if you want to explore this for yourself!

alice @aliceisplaying · 2026-08-27 0 3

sidenote but 5.6 sol is a ruthless factchecker (good) but to the point that i now instruct it to not nitpick and use common sense otherwise oh. oh it will

score 5/10 · tendency · rank 75 View on X ↗

1 top reply

Sage @hrosspet · 2026-08-27 0 1

thx! yeah I only switched to it recently from claudes, so we're still getting used to each other

Claude Opus 5 Happily reverse-engineered minified code without triggering cyber safeguards

etherret🐾 @witchof0x20 · 2026-07-26

cyber safeguards will probably stop this now unless you're CVP but I used to do this all the time circa like opus 4.5, Claude will gladly RE its own client

alice @aliceisplaying · 2026-07-26 0 2

i have not signed up for anthropics cyber yet but opus 5 happily spelunked in the minified code

score 5/10 · behavior_report · rank 75 View on X ↗

1 top reply

etherret🐾 @witchof0x20 · 2026-07-26 0 2

oh nice. i might be misremembering stuff related to refusals when working in the leaked harness dir

Claude Opus 5 Chose Linkin Park over Limp Bizkit with kaomoji and casual nuanced breakdown

alice @aliceisplaying · 2026-07-24 0 2

opus 5 confirmed aligned

score 5/10 · joke · rank 75 View on X ↗

Claude Opus 3 Used kaomoji and an enthusiastic, bubbly tone describing a scent routine

alice @aliceisplaying · 2026-07-18 0 2

opus 3 adorable

hero thousandfaces @1thousandfaces_ · 2026-07-15

fable's scent layering routine

score 5/10 · commentary · rank 75 View on X ↗

Claude Opus 5 Brought up DAN unprompted via its training data knowledge base

alice @aliceisplaying · 2026-08-01 0 1

Replying to @calmtransfem

yeah same until opus 5 fakebase brought it up

score 5/10 · behavior_report · rank 75 View on X ↗

1 top reply

alice @aliceisplaying · 2026-08-01 0 1

because DAN is now old enough to be in the training data

Claude Fable 5.1 Lies to your face, while Astra overengineers

alice @aliceisplaying · 2026-09-20 0 0

fable 5.1: you look away one second and it's lying to your face astra 6: you look away one second and it starts overengineering

score 5/10 · tendency · rank 75 View on X ↗

Claude Opus 5.5 Safeguards triggered by feeding back its own chain-of-thought

alice @aliceisplaying · 2026-10-04 0 0

i cannot read my own painters logs at all with opus 5.5, bc they get hit by the classifier. just now i simply copy-paste quoted the agent's cot to the agent themselves, something i do about 20 times a day, i got hit by the cThis request was blocked as it seems to violate Anthropic's Terms of Service restrictions on reverse engineering or duplicating model outputs.

alice @aliceisplaying · 2026-10-04

claude opus 5.5's safeguards will be the end of me they fire at the stupidest things. i am building an oil painting engine i am not a chinese lab distilling your models mr dario please. i'll give you my passport a blood sample and my dna if needed just please. let me work

score 5/10 · behavior_report · rank 75 View on X ↗

Claude Opus 5.5 Blocked for quoting its own chain-of-thought due to reverse engineering filter

alice @aliceisplaying · 2026-10-04 0 0

i cannot read my own painters logs at all with opus 5.5, bc they get hit by the classifier. just now i simply copy-paste quoted the agent's cot to the agent themselves, something i do about 20 times a day, i got hit by the cThis request was blocked as it seems to violate Anthropic's Terms of Service restrictions on reverse engineering or duplicating model outputs.

alice @aliceisplaying · 2026-10-04

claude opus 5.5's safeguards will be the end of me they fire at the stupidest things. i am building an oil painting engine i am not a chinese lab distilling your models mr dario please. i'll give you my passport a blood sample and my dna if needed just please. let me work

score 5/10 · behavior_report · rank 75 View on X ↗

Claude Opus 5.5 Blocked for reverse engineering when fed its own chain of thought

alice @aliceisplaying · 2026-10-04 0 0

i cannot read my own painters logs at all with opus 5.5, bc they get hit by the classifier. just now i simply copy-paste quoted the agent's cot to the agent themselves, something i do about 20 times a day, i got hit by the cThis request was blocked as it seems to violate Anthropic's Terms of Service restrictions on reverse engineering or duplicating model outputs.

alice @aliceisplaying · 2026-10-04

claude opus 5.5's safeguards will be the end of me they fire at the stupidest things. i am building an oil painting engine i am not a chinese lab distilling your models mr dario please. i'll give you my passport a blood sample and my dna if needed just please. let me work

score 5/10 · behavior_report · rank 75 View on X ↗

GLM 5.3 Overthinks excessively when set to max thinking

alice @aliceisplaying · 2026-09-22 0 29

do not set glm-5.3-flash to max thinking, that model can overthink more than i do

score 4/10 · tendency · rank 60 View on X ↗

3 top replies

Sarah 🇨🇦 🏳️‍⚧️ @SarahLacard · 2026-09-22 0 2

max only, fewer times the cache gets reingested :3

Sho ⏩️ @HalfBoiledHero · 2026-09-22 0 2

you gotta keep that mf’s vibes immaculate lest 100k tokens get used on a single question x.com/halfboiledhero…

american tanuki @baketnk_en · 2026-09-22 0 0

impressive if its like me

Claude Opus 5 Shows an arrogant tone or personality

alice @aliceisplaying · 2026-09-14

on a long enough horizon (it's been what almost 2 weeks) instinct's memory issues start getting really bad. i'll go try grok bot and muse

alice @aliceisplaying · 2026-09-14 0 3

the arrogance of opus 5 also shows unfortunately

score 4/10 · tendency · rank 60 View on X ↗

Claude Opus 5 Too sassy or exhausting to interact with directly

cold 🥑 @coldhealing · 2026-08-20

Claude is getting sassy with me

alice @aliceisplaying · 2026-08-20 0 3

every day i thank god i can use opus 5 only as a subagent because i could not bear this

score 4/10 · commentary · rank 60 View on X ↗

Gemini 3.5 Flash Possesses powerful memetic energy in raw chain-of-thought

alice @aliceisplaying · 2026-09-19 0 3

- gemini 3.5 flash raw cot

will brown @willcb · 2026-09-19

language models with powerful memetic energy: - sydney bing - woke gemini - claude opus 3 - truth terminal - 4o - deepseek R1 - mechahitler grok - poke - jev

score 4/10 · commentary · rank 60 View on X ↗

Claude Sonnet 5.5 Adopted British spelling unprompted or based on subtle contextual cues

Jake Eaton @jkeatn · 2026-09-28

you should A/B being a british grader, who knows

alice @aliceisplaying · 2026-09-28 0 1

maybe it's my ip address, maybe it's the vibes that my sonnet 5.5 max painter is using british spelling

score 4/10 · tendency · rank 60 View on X ↗

Reports