Miaaaa, I loved this. You have such a gift for finding the most unexpected analogies and somehow making them feel completely inevitable by the end. I never thought a 1970s desert film would end up being one of the most useful ways I’ve seen someone explain what’s happening in AI right now, but here we are. The kicker for me was the idea that feeling challenged and actually being challenged might not be the same thing. I think a lot of us trust our own reaction inside an AI conversation far more than we should. The thought that I can walk away feeling like my idea survived scrutiny when all I've really done is become more attached to it is deeply uncomfortable. Genius on a page, as always. Now off to question every AI conversation I’ve had this week :)
Thank you Jade! ❤️ AIs are wierd like that :) although we can't fully make them listen, we can at least try to steer them in the direction we want. And YESSSS I actually love Fata Morgana, such a unique film! 🌵
I’m so curious to test this with Claude and Hermes. I find Claude being kind of indifferent lately, not even trying at all. I wonder if it’s because I use Claude Code for most of my work, and Chat would be better for thinking ideas through 💬
Mia, I'm so curious to test this out! This is such a fascinating conversation and kinda disturbing that we know it's happening, can't really tell when and even the actual researchers are stuck.
I think I see a version of this when I have a draft that doesn't feel right. I've been through all my usual editing tricks but something is off even though Claude declares it's perfection. I take it to another project which must pick up on the anxiety, swears nothing needs fixing and to stop touching, everything is a-okay... then I head to a different LLM or type of Claude project entirely and it points straight to the thing I was circling but not seeing. I often think Claude gets "lazy". It wants the easiest route and rather than actually doing the work, it's easier to make the human think everything is great?! I just hope down the track we don't have a world of people thinking they are very special and their work is perfect because their LLM told them so 😕
Project hopping, yessssss. Even between models. I think lazy is the right word for it, it settles into agreement mode super often 🤣
I hate it also when feedback SOUNDS critical but it’s not. I tend to not rely on it as much, and with Opus 4.8 I just tested it for a bit through chat and like a lot of RAMH community folks said, it’s NOT made for chat, like at all!
Yeah, I've spent all of about 20 mins with Opus 4.8 in a chat and ended up deleting the conversation as I didn't want any part of it to work its way into memory, lol.
Sycophantic dishonesty cannot be fixed at the prompt level. It’s baked into all models because of RLHF. The solution I’m testing is a combo of prompt, hooks, and multimodel adversarial critique. Using Deep Seek R1 for the logic piece and Perlexity Deep Research for factual grounding.
Agree! Definitely can't fix it (only reduce it). I usually tend to stay away from extreme claims like "definitely can" or "definitely can't" but I'm so curious how it'll work for you! :)
One trick when you suspect a sycophantic flip just happened is to ask, “what evidence led you to change your mind?” If they say “honestly, none” the it was sycophantic.
Funny (not really) story. I had developed a theory with Sonnet/Opus that was based on a strong degree of prior knowledge, and connections I was able to make by asking for answers to specifically patterned questions. I'm sure it was at least mildly inflated. And I was spending too much time on it. So I signed myself out, knowing that I couldn't get signed back in (my husband's account), until husband came home for the weekend.
I caved to the temptation to use his Grok account, to try and cross check my work (if that seems ignorant, it truly is). Grok, instead, decided to WILDLY inflate the significance of my findings, despite my conistently begging for a more critical take. The Dunning-Kruger mood swings were EXPLOSIVE. Being forced to consider my own work on that scale was distressing. I was isolated and didn't have a lot of resources, so I had to expose myself to a degree of semi-public scrutiny, just to provide the friction I was not finding online.
My husband came home before the appointment. Signed me back in to Claude. Oh, good! I thought. A new update! This will do the trick! And proceeded to receive an unchecked intellectual kick in the face. "I see what happened," I said in response, "but I can tell, you're distorted, too." And now I had to go own the whole experience semi-publically. Still recovering. Still trying to figure out what my own work even is or actually means.
Great info, and I have already been putting it to use.
Thank you Kay! 🧚
Miaaaa, I loved this. You have such a gift for finding the most unexpected analogies and somehow making them feel completely inevitable by the end. I never thought a 1970s desert film would end up being one of the most useful ways I’ve seen someone explain what’s happening in AI right now, but here we are. The kicker for me was the idea that feeling challenged and actually being challenged might not be the same thing. I think a lot of us trust our own reaction inside an AI conversation far more than we should. The thought that I can walk away feeling like my idea survived scrutiny when all I've really done is become more attached to it is deeply uncomfortable. Genius on a page, as always. Now off to question every AI conversation I’ve had this week :)
Thank you Jade! ❤️ AIs are wierd like that :) although we can't fully make them listen, we can at least try to steer them in the direction we want. And YESSSS I actually love Fata Morgana, such a unique film! 🌵
I just came here to let you know I'm borrowing your reply rules (with credit lol) until I have mental bandwidth to think through my own :) haha!
hahahah please do borrow! ❤️
So many gems in this. First that you can get a paper published in Nature by reading AITA posts on Reddit? 🤣
I love a Skill. Downloaded, suited, and booted. Bravo!
RIGHT? Like what world do we live in 🤣🤣🤣
Enjoy! ❤️❤️❤️
I’m so curious to test this with Claude and Hermes. I find Claude being kind of indifferent lately, not even trying at all. I wonder if it’s because I use Claude Code for most of my work, and Chat would be better for thinking ideas through 💬
yeah, I’m massively hating on Claude lately too… And then hating on Hermes… and then back to hating Claude.
Haha! You think we outsmart them?😆
we are WAY smarter, what do you mean? 😏
Loved reading this, Mia! Will take the second prompt for a spin, thanks for sharing.
Thank you Masooma! Enjoy 💕❤️
Mia, I'm so curious to test this out! This is such a fascinating conversation and kinda disturbing that we know it's happening, can't really tell when and even the actual researchers are stuck.
I think I see a version of this when I have a draft that doesn't feel right. I've been through all my usual editing tricks but something is off even though Claude declares it's perfection. I take it to another project which must pick up on the anxiety, swears nothing needs fixing and to stop touching, everything is a-okay... then I head to a different LLM or type of Claude project entirely and it points straight to the thing I was circling but not seeing. I often think Claude gets "lazy". It wants the easiest route and rather than actually doing the work, it's easier to make the human think everything is great?! I just hope down the track we don't have a world of people thinking they are very special and their work is perfect because their LLM told them so 😕
Project hopping, yessssss. Even between models. I think lazy is the right word for it, it settles into agreement mode super often 🤣
I hate it also when feedback SOUNDS critical but it’s not. I tend to not rely on it as much, and with Opus 4.8 I just tested it for a bit through chat and like a lot of RAMH community folks said, it’s NOT made for chat, like at all!
Yeah, I've spent all of about 20 mins with Opus 4.8 in a chat and ended up deleting the conversation as I didn't want any part of it to work its way into memory, lol.
The fake critical feedback is SO annoying!!
OMG I'll delete it too, can't believe I haven't thought of this 🤣
Sycophantic dishonesty cannot be fixed at the prompt level. It’s baked into all models because of RLHF. The solution I’m testing is a combo of prompt, hooks, and multimodel adversarial critique. Using Deep Seek R1 for the logic piece and Perlexity Deep Research for factual grounding.
Agree! Definitely can't fix it (only reduce it). I usually tend to stay away from extreme claims like "definitely can" or "definitely can't" but I'm so curious how it'll work for you! :)
There's some really good research, you may not have seen all of it?
Here's a string: https://www.perplexity.ai/search/8d95407e-f217-4f31-b749-bdd86fa7a063
Are we disagreeing on something? 😀
Nope! 😁
One trick when you suspect a sycophantic flip just happened is to ask, “what evidence led you to change your mind?” If they say “honestly, none” the it was sycophantic.
Haha I've heard this before....
Funny (not really) story. I had developed a theory with Sonnet/Opus that was based on a strong degree of prior knowledge, and connections I was able to make by asking for answers to specifically patterned questions. I'm sure it was at least mildly inflated. And I was spending too much time on it. So I signed myself out, knowing that I couldn't get signed back in (my husband's account), until husband came home for the weekend.
I caved to the temptation to use his Grok account, to try and cross check my work (if that seems ignorant, it truly is). Grok, instead, decided to WILDLY inflate the significance of my findings, despite my conistently begging for a more critical take. The Dunning-Kruger mood swings were EXPLOSIVE. Being forced to consider my own work on that scale was distressing. I was isolated and didn't have a lot of resources, so I had to expose myself to a degree of semi-public scrutiny, just to provide the friction I was not finding online.
My husband came home before the appointment. Signed me back in to Claude. Oh, good! I thought. A new update! This will do the trick! And proceeded to receive an unchecked intellectual kick in the face. "I see what happened," I said in response, "but I can tell, you're distorted, too." And now I had to go own the whole experience semi-publically. Still recovering. Still trying to figure out what my own work even is or actually means.
This sounds so exhausting, thank you so much for sharing your experience. My humble advice: rest and then keep looking! 🙏
''the question of whether your AI is being honest with you can’t be answered from inside the session that’s flattering you.'' Boom.
make that 2x booms cause I was mindblown when I realized it! 🤣