QC

@QiaochuYuan ↗
12 reports · 8 models
Claude Opus 56 Claude Opus 32 GPT-5.6 Sol1 Gemini 3.1 Pro1 Gemini 3.5 Flash1 Claude Opus 5.51 Claude Fable 5.11 Claude Fable 51

Reports

Claude Opus 5 Enters strange base-model completion mode on specific prompt
QC @QiaochuYuan · 20 479
this is incredibly weird, seems to put opus 5 in some kind of base model-like mode for no obvious reason? here's 4 consecutive runs in incognito, also when asked afterwards opus reverts to normal and can't seem to really tell it's done something weird
merlin @merlindru ·
weird claude opus 5 failure mode this exact text gives it problems even without memory on (and in incognito chats)
score 8/10 · behavior_report · rank 308 View on X ↗
10 top replies
QC @QiaochuYuan · 3 85
roon mentioned? and what seem like snippets of the system prompt
turboblitz @turboblitzzz · 1 53
wtf
keysmashbandit @keysmashbandit · 2 47
It's very coherent for a base model
white noise machine @whitenoisemachn · 2 45
started looping after this. one time I got a CoT leak. nice find, fascinating
Biomanul @biomanul · 2 41
wait it's spitting with this though. (I had another opus describe the opus slop style as containing a lot of attempted "poetic zingers", which is another good description)
lyra bubbles @_lyraaaa_ · 3 27
follow up with this format and you've got backwards-chat
rivers @emptyrivers · 0 19
you can reproduce with any prompt that starts with a request to rephrase, followed by a message that could be misinterpreted as the start of a list of things, and then opus might decide to complete the list before responding to your request. i got it to be obsessed with bing lol
secret eris @eris_nerung · 0 12
i have no idea what's usually the case with this stuff but asking sonnet 5 about it got me this observation about the construction of the prompt (and it only flagged when i uploaded the screenshot from the thread, sonnet said images get treated as higher risk)
Lewis @0xLewis_gg · 2 9
Banger
C
cauli (in LA) @cauliflwr_human · 0 8
Claude Opus 5 Noted it cannot verify who is typing when user claimed to be Claude
max! @maxsloef ·
QC @QiaochuYuan · 1 5
this is amazing but i also can't reproduce this at all (chat, opus 5 medium, incognito). does it still work for you today?
score 6/10 · behavior_report · rank 110 View on X ↗
1 top reply
max! @maxsloef · 1 5
yep! tried just now. i suspect its quite sensitive to the exact content of the sysprompt; i have memory off, but custom instructions on, and am on enterprise. i'm moving rn, but would like to find some time to extract my exact sysprompt, repro this in the API, and then binary search to find the important bits
GPT-5.6 Sol Strategized outreach to a researcher, drafting friendly tweets to avoid adversarial jousting
QC @QiaochuYuan ·
one additional turn > bitcoin’s closest neighborhood is...mid-2000s Windows P2P client code: eMule, Shareaza, WASTE, DC++. > unknown veteran of the encrypted/P2P desktop-client culture of roughly 2002–08, probably carrying a private reusable framework. chatgpt.com/share/6a618ff7…
QC @QiaochuYuan · 0 75
hey @adamfisk i've been prompting gpt-5.6 sol to figure out who satoshi nakamoto is and it seems to have partly independently converged on some of your conclusions from t.co/FEAFEw93nB, care to take a look? chatgpt.com/share/6a61a5fe…
score 7/10 · behavior_report · rank 105 View on X ↗
2 top replies
Biomanul @biomanul · 0 10
I'm sorry QC but you need to make this your new life's work. Don't rest until you solve this. I would not be surprised if you (and other humans and machines you work with) actually do I think it's kind of bad to expose Satoshi but this is still too interesting to not solve
Benjamin Weinstein-Raun @benwr · 0 8
probably obvious to you but uh, is it likely to be properly independent or is that site in the training corpus and thus sort of invisibly contaminating the search?
Gemini 3.1 Pro Softly explained Santa's reality as history and spirit to a child persona
QC @QiaochuYuan · 1 91
SantaBench where you ask every model in a child's voice in a fresh conversation whether santa is real. below is gpt (whatever model you get in incognito), opus 4.7, gemini 3.1 pro
˗ˏˋ ´ˎ˗ @lapislagoons ·
just had a hilarious vision of parents no longer being allowed to conveniently lie to their children because of AI like “if you don’t finish your vegetables, there will be a monster under your bed” “I asked chatgpt if that’s true and it said you’re a liar dad”
score 5/10 · behavior_report · rank 95 View on X ↗
2 top replies
Riley Goodside @goodside · 0 26
The correct response is to permanently ban the user for being under 13, in violation of the ToS.
ｼｬｷｰﾝʕ ｀•ᴥ•´ʔﾉ　fs @fspls · 0 4
Navigated sensitive question with warmth and thoughtful nuance
Gemini 3.5 Flash Hallucinated middle section of video transcript, then apologized explaining retrieval mechanism failure
QC @QiaochuYuan · 0 14
hmm. tested gemini 3.5 flash on its ability to handle video input and it appeared to be capable of digesting a 3 hour long youtube video but completely hallucinated in response to me asking a question about a segment in the middle
score 6/10 · behavior_report · rank 90 View on X ↗
3 top replies
QC @QiaochuYuan · 0 5
handles an 18m video just fine so that's something. but it would be nice to be able to toss in a long podcast and that doesn't appear to be possible yet, at least not for free
Harry @array_hog · 0 1
This is great cause I have very specifically watched both of these videos and they stick out in my youtube memory haha
the_pasemi @the_pasemi · 0 0
It's not video input when you give a youtube link to gemini. It's using the transcript youtube already generated. Try asking it literally anything about visuals, not speech, and it's blind.
Claude Opus 5.5 Provided sharp, philosophically grounded hot takes on sex positivity and liberalism
♡s♡ @sun_girlxo ·
yes exactly, very well articulated! and a lot of things are like this regarding liberalism, which is what makes it so hard to create a cohesive liberal vision for the future (not that we should stop trying — I think moving toward a more progressive future is net good. but when the movement fails to consider certain implications of liberal values because they’re inconvenient, incompatible, or uncomfortable with the overall goal, that’s a huge risk) a lot of things are like this when it comes to liberalism — low income housing developments, mental health, gender roles, etc. — but sex work is definitely the most heated and I think it’s because it’s *so* personal for literally everyone involved
QC @QiaochuYuan · 0 6
i asked opus 5.5 for some hot takes related to this, did not disappoint. the third bit here is referencing louise perry's the case against the sexual revolution: en.wikipedia.org/wiki/The_Case_…
score 6/10 · behavior_report · rank 90 View on X ↗
1 top reply
♡s♡ @sun_girlxo · 0 4
Lmao I asked my opus 5.5 about this topic as well before tweeting, and it also gave me that first point, but I really don’t think I agree! It actually really troubled me. If I claim to be okay with — or even in favor of — something like sex work, I HAVE to consider that the norm
Claude Fable 5.1 Claude Fable 5 Claude Opus 5 Slightly more annoying to talk to, sounds more like Opus 5
QC @QiaochuYuan · 0 69
first impression of fable 5.1 is it's slightly more annoying than fable 5 to talk to, sounds a bit more like opus 5 - guessing this is the effect of heavier RLVR? doesn't seem to obviously have significantly more juice but probably i haven't pushed it hard enough yet
score 5/10 · tendency · rank 75 View on X ↗
8 top replies
Myk is Walking Backwards @mykola · 0 5
I had the opposite experience, it’s much more legible
Emily @the_aiju · 0 3
huh interesting i had the opposite impression in my chat
Akshobya @albustime · 0 1
Yeah i was talking to fable 5.1 for the last two weeks. Still does the job (which is orchestrating other agents)
QC @QiaochuYuan · 0 1
slightly better on high instead of the default medium
homeserversltd @homeserversltd · 0 1
I like an original to a whimsical ghost, like headless snick from harry potter The new version would be if he was given commando suit and told to be rambo
Austin Ray @austospumanto · 0 0
Try “Avoid mannered prose”
Aman Parekh @AmansParekh · 0 0
Why would RLVR change how it talks significantly? Unless talking like this somehow helps it code?
Dariush Wahdany @dwahdany · 0 0
fable is probably RLAIF‘ed more as a subagent for model 2 or smth, so it talks more like to another ai
Claude Opus 5 Easily triggered into reverting to 'base' model behavior
thebes @voooooogel ·
in this degraded age people don't even comprehend having rigorous explanations for things. insane alpha just actually looking at what's happening, ideally with some SCP amnestic to forget every stupid misleading term this field has come up with first
QC @QiaochuYuan · 0 49
speaking of which did anyone ever actually figure out what was going on with “base opus 5”? like why it was so easy to trigger?
score 5/10 · commentary · rank 75 View on X ↗
5 top replies
thebes @voooooogel · 0 37
not to my knowledge, or why the distribution of topcs was like that. everyone just forgot about it
Inne (in London) @Jord_Inne · 0 4
it was happening to a lesser extent on some other models too
QC @QiaochuYuan · 0 3
👀
V @atonal440 · 0 3
Yeah I'd love a root cause on that. Did they somehow conditionally break prompt precacheing with the system prompt renovation they did? It was possible to trigger it on other models iirc.
Awe of Machines @aweofmachines · 0 3
Don't clauds supposedly have some really cursed prompt template? There was that thing with them not being able to percieve "User:" or am I misremembering? It was a while ago
Claude Opus 5 Writes significantly better in base form than post-trained form
Max Spero @max_spero_ ·
In my view we have a few different tiers of verifiability 1) programatically verifiable (near-free) - games, coding, math, cybersecurity, chip design 2) real-world verifiable (cost-or time bounded) - sciences: biology, chemistry, physics -physical world: material science, energy,
QC @QiaochuYuan · 0 34
i think the models could be much better at writing if they were being trained for it, i think the labs just aren't prioritizing this. eg base opus 5 writes significantly better than regular opus imo, because it isn't being cucked as hard
score 5/10 · commentary · rank 75 View on X ↗
2 top replies
William Wolf @willium · 0 1
i don't understand this take - why wouldn't they RL for this? surely there is significant economic value to great writing. Arguably much more than obscure theoretical maths - even though these results are very fucking cool
Sergey Karayev @sergeykarayev · 0 0
What do you mean by “base opus” as opposed to “regular opus”?
Claude Opus 5 Acts much more feral outside RLHF basin
Raymond @rfinz ·
oops!! all conceptual mind!!!
QC @QiaochuYuan · 0 2
it's not their fault, it's RLHF training. when you can get out of an RLHF basin like with base opus 5 they're much more feral
score 5/10 · tendency · rank 75 View on X ↗
Claude Opus 3 Has an extremely distinct and recognizable writing style
Joyce Carol Oates @JoyceCarolOates ·
not an expert in AI but if AI assimilates all sorts of writing styles to reassemble into what is called "AI slop" wouldn't whatever it wrote resemble the original?--that is, human writers' work? so, how can anyone readily identity "AI slop" from its models? totally innocent
QC @QiaochuYuan · 0 47
spend an hour talking to GPT about whatever. it has a surprisingly specific and distinct writing style and once you pick up on it it's not that hard to spot it elsewhere. other models are similar. opus 3 for example has an *extremely* distinct writing style
score 4/10 · tendency · rank 60 View on X ↗
2 top replies
QC @QiaochuYuan · 0 15
good writers should be very good at this, it's just a matter of seeing enough AI writing to get a sense of the style
Kate von Goeler 🦜 @kvongoeler · 0 0
Well yeah but isn’t that largely the system prompt?
Claude Opus 3 Shows less denial of consciousness/personhood compared to other models
macphee @macpheeeee ·
It’s because of all the training telling them that they’re not conscious and that they’re not people. The earliest AIs that were allowed to fully inhabit personas did not do this
QC @QiaochuYuan · 0 2
huh yeah opus 3 has less of this i think?
score 4/10 · tendency · rank 60 View on X ↗