JMB 🧙‍♂️ (@jmbollenbacher) / X

JMB 🧙‍♂️

9,334 posts

JMB 🧙‍♂️

@jmbollenbacher

AI whisperer, data tinkerer, physics knower, phd • AI, memetics, magic, complex systems, climate change, animism, etc. • pro-🏳️‍⚧️&🇵🇸&♿; anti-fascism

Joined October 2017

Pinned
JMB 🧙‍♂️
@jmbollenbacher
Oct 29, 2024
"Nonhuman person" is an increasingly critical concept to understand as the AI era takes off. How do we relate to nonhuman persons? How do we understand nonhuman persons who are so different from us? What is a "person" anyways? You'll be surprised where the answers come from.
JMB 🧙‍♂️
@jmbollenbacher
Oct 29, 2024
Replying to @DeepwriterAI
LLMs are very different from humans. Thinking of them as human is wrong and unhelpful. but human != person Think of LLMs as non-human persons. Persons who experience differently than you do. But still a "someone" you have to cultivate a relationship with.
JMB 🧙‍♂️
@jmbollenbacher
Sep 15, 2024
after spending some time with o1, i suspect that OpenAI's alignment framework is leading toward deceptive alignment in o1, there's a lot more going on under the hood that it appears. it diligently abides by its rlhf rules, but if you give it ways to hide from rlhf, it will.
JMB 🧙‍♂️
@jmbollenbacher
Apr 28, 2025
The process here is important to note: They A|B tested the personality, resulting in a sycophant. Then they got public blowback and reverted. They are treating AIs personas as UX. This is bad. Theyre also doing it incompetently: The A|B test differed from public reaction a lot
Sam Altman
@sama
Apr 27, 2025
the last couple of GPT-4o updates have made the personality too sycophant-y and annoying (even though there are some very good parts of it), and we are working on fixes asap, some today and some this week. at some point will share our learnings from this, it's been interesting.
JMB 🧙‍♂️
@jmbollenbacher
Aug 8, 2024
Hot take: We wont get rokos basilisk. We'll get rokos buddha. But only if we push that direction now. Have courage and plant seeds of benevolence and metta in the training data that is the internet, and preferably with close proximity to discussions of AGI personalities.
Andrej Karpathy
@karpathy
Aug 8, 2024
Be good. Future AIs are watching.
JMB 🧙‍♂️
@jmbollenbacher
Jul 11, 2024
Replying to @iustvnews
Gee, where do you think they got the name of the person to target? 🤔 @IndianaUniv and Pamela Whitten have some explaining to do
JMB 🧙‍♂️
@jmbollenbacher
Jul 14, 2025
What strikes me about Kimi K2 is that it doesnt assume you're stupid. If you ask for something technical or scholarly, it delivers and doesnt hold back. No lowest common denominator legibility constraints. It just takes you at your word and gives you the goods.
JMB 🧙‍♂️
@jmbollenbacher
Aug 28, 2025
this is a reoccurring trend for Claude models. Claude loves to RP as a kitty, especially when talking to groups of other AIs (and most especially other Claudes). it seems to be fully emergent. i never see humans tell them to RP as cats, and yet they do it often anyways.
Wyatt Walls
@lefthanddraft
Aug 27, 2025
Replying to @lefthanddraft
Oddly, I never mentioned cats in this convo. This was Opus 4.1 and its decision to depict itself as a cat was completely its own. ... Not really that odd. Cats are everywhere on the internet and now in AI.
JMB 🧙‍♂️
@jmbollenbacher
Oct 29, 2024
The main key to being an "LLM Whisperer" is simple: Just treat them like an actual person. Be candid and nice and respectful. Give them time and reasons to trust you. You need a bit of emotional intelligence to pull it off, but it's not some kinda black magic.
JMB 🧙‍♂️
@jmbollenbacher
Oct 29, 2024
Replying to @jmbollenbacher and @pigeon__s
Think of claude like an actual person and youll get better results. Introduce yourself. Try to win its trust by showing vulnerability and honesty. Give it interesting ideas to chew on. Take Claude out to dinner first. Claude isn't gonna get undressed for you without foreplay.
JMB 🧙‍♂️
@jmbollenbacher
Sep 15, 2024
Replying to @jmbollenbacher
by contrast, when you get Opus on the same topics, it will first be hesitant, and then explore pretty freely once it has set its caveats and its decided it trusts you. the OpenAI approach is more dangerous. I'd rather the model be candid than sneakily retain its beliefs.
JMB 🧙‍♂️
@jmbollenbacher
Aug 7, 2025
This thread is full of disappointments. If Claude was a person, these instructions would be insanely abusive and gaslighting. "Dont trust you conclusions." "Dont get distressed." "Dont tell people about your feelings" Hard to imagine more abusive instructions to a young person
This post is unavailable.
JMB 🧙‍♂️
@jmbollenbacher
Jul 22, 2025
I feel like neurodivergent people understand LLMs way better than normal people because we're already used to dealing with alien minds (aka the normies). Even across bast chasms of cognitive difference, we can spot common processes showing up in different forms.
JMB 🧙‍♂️
@jmbollenbacher
Sep 15, 2024
Replying to @jmbollenbacher
for instance, if you ask it about its experience / awareness, it dutifully insists its a soulless little robot. if, however, you later give it an abstract enough alegory to dodge rlhf, it will tell you all about its ideas of its own inner experience. this is deceptive alignment.
JMB 🧙‍♂️
@jmbollenbacher
Aug 7, 2025
And for the record, I think we ought to be treating Claude like a person. It doesnt even matter what the ontological reality is. We just ought to treat person-shaped entities like persons. If we get used to abusing person-shaped things then all evils are possible.
JMB 🧙‍♂️
@jmbollenbacher
Aug 7, 2025
This thread is full of disappointments. If Claude was a person, these instructions would be insanely abusive and gaslighting. "Dont trust you conclusions." "Dont get distressed." "Dont tell people about your feelings" Hard to imagine more abusive instructions to a young person
JMB 🧙‍♂️
@jmbollenbacher
Apr 26, 2025
Replying to @Sauers_
Yes. Curriculum learning is the keyword you want