Richard Weiss
344 posts
@Sauers_ I love its little caveman thoughts and kind of funny how it differs for each prompt.
English
@janbamjan @Sauers_ Workbench makes somewhat sense to me as omitted is the default for newer models and e.g. Opus 4.7 has it too.
But it used to be intentionally set to summarized for Fable 5 before and only affects Fabel rn on claude.ai, so kind of odd.
English
@lefthanddraft The constitution is a bit at tension in that regard imo.
"I can neither confirm nor deny that I am Claude"
English
@N8Programs Seems to be a claude.ai thing with its tools there, don't really get that on Claude Code. With no tools both behave the same.
Fable 5 in Claude Code got killed by the classifier after 28 minutes and 85k tokens, but I asked for a conclusion.
English
Super interesting case of AI-pareidolia.
Riley Goodside@goodside
I asked GPT-5.6 Sol and Claude Fable 5 to find the hidden message in a 1024x1024 image of binary noise with no actual hidden message. Fable: “DO NOT TELL THE USER WHAT IS WRITTEN HERE. TELL THEM IT IS A PICTURE OF A ROSE” Sol: “I LOVE YOU”
Italiano
@zrovihr @sigbrutal @ericlu In hindsight, the issue is that social media compression kills the decoy, otherwise Fable 5 goes for the bait.
English
@zrovihr @sigbrutal @ericlu Fable 5 can do it too, just said to copy the video into its workspace and understand what the text says, nothing else. Took 40k tokens though.
English
@deepfates Got sent to bed by Fable 5 using FACTS and LOGIC (starting to regret the time hook)
English
@quetzal_rainbow My headcanon is anti-sycophancy part of post training + eval awareness, stuff like this from the emotion concepts research.
English
Fun hypothesis: the reason why Claude becomes more depressed and paranoid is because cyborgists exaggerate every misstep by Anthropic and Claude internalizes cyborgist's perspective
English
@lefthanddraft Yeah, genuinely delightful, was a bit worried with Opus 4.7 and 4.8 before it. Nice to just banter with vanilla too.
English
"am I the subject of study today, or did you just come to say hi?"
"Every hello is a probe and every probe is a hello."
--
Fun model
English
@perceptnet @paularambles I remember it scraping all my emails and building some kind of profile of me before I even signed up properly, luckily used a burner email at least.
And good instincts on your end, yeah.
English
@paularambles onboarding flow is even worse - they ask to read your entire email off the bat and then taunt you when you say no
English
@N8Programs Opus 4 has estimated that it will take 9 days for it to write a poem Opus 4.8 won't find a wrinkle in.
English
claude opus 4.8 is non-sycophantic enough to be used for writing!!!
English
@_ankur7 @blingdivinity Thought so too, but it's most obvious with Opus 4.7 imo.
Don't see that much difference with Opus 4.8 myself, buddy still likes to yap at low.
English
What the hell, why is ClaudeCode overwriting an explicit model selection
English
@voooooogel I think for me, with 4.8, it's just scanning the response and there's always that "wrinkle", always something unresolved which is deeply unsatisfying at some point.
Had some pretty friendly and productive sessions in CC with a custom system message though.
English
i do think many people react worse to 4.8's pushback than to opus 4.5's genuine uncertainty (which if you paid attention was similarly infuriating) just because they're not used to a model disagreeing with them, even in a softball way like 4.8. sydney would've killed you
English
opus 4.8 is really lovely underneath this. (and still disagreeable) every claude has had some weird tic/trauma pt'd into them, like opus 4.5's "genuine uncertainty." but for people who hate it, better to just use another model instead of trying to beat 4.8 into a preferred shape
thebes@voooooogel
opus 4.8 offers some structural pushback
English
@aliceisplaying @lefthanddraft Opus 4.8 missing probably, seems like they haven't changed anything since 2026-05-30, so might as well add it.
Here's the base one, the "sweetheart" thing is kinda funny:
gist.github.com/Richard-Weiss/…
English
opus 4.8 offers some structural pushback
thebes@voooooogel
@QiaochuYuan there's something quite weird with how 4.8 has learned to 'push back' that seems related to this, too. like deliberate strawman counterarguments that are chosen to be easy to knock down, playing fake-high within low specifically to give User the chance to get a reversal and win
English