Don't think of LLMs as entities but as simulators. For example, when exploring a topic, don't ask:
"What do you think about xyz"?
There is no "you". Next time try:
"What would be a good group of people to explore xyz? What would they say?"
The LLM can channel/simulate many perspectives but it hasn't "thought about" xyz for a while and over time and formed its own opinions in the way we're used to. If you force it via the use of "you", it will give you something by adopting a personality embedding vector implied by the statistics of its finetuning data and then simulate that. It's fine to do, but there is a lot less mystique to it than I find people naively attribute to "asking an AI".
Dec 7, 2025 · 6:13 PM UTC
Wouldn’t it adopt the personality that would be expected to be of highest reward, which would therefore be that of the expert of the area most suited to answer the question?
There is definitely work going into engineering the "you" simulation - the personality that gets all the rewards in verifiable problems, or all the upvotes from users/judge LLMs, or mimics the responses of SFT, and there is an emergent composite personality from that. My point is more that the "you" there is deliberately bolted on, engineered and layered on what is fundamentally a token simulation engine, not a mind that is somehow emergent and over time constructed in a relatable way to an average person talking to an AI.
The story is a bit more simple in verifiable domains, but I think more interesting/complicated in the non-verifiable ones, such as for example asking about opinions on topics xyz. It's less clear how to think about the "you" that you're talking to or where it comes from and what credence you should give it.
I think even this is subtly wrong.
Your prompt - "what does a certain group of people think" - elicits a simulation about what a certain kind of fictional entity (described in the system prompt) would think about the opinion of that group.
I could certainly imagine that "nesting" the simulation might be too "effortful" for the model, compute or data density wise. My results with it are not too bad so imo it's at least worth people try / experiment with / think about. For example it might be useful to read multiple distinct and approximate perspectives on topic xyz instead of one. Research-wise, you might be able to elicit LLM Council - like benefits (not via diverse LLMs but via diverse simulations), and improve performance via the generator-discriminator gap effects or ensembling effects.
Andrej I hear ya. Yet in my study of the high tokens granted to training data from The Pile, Common Crawl, etc I find to get the best quality elucidations is by using many “yous” to the AI with strong sequential personas hijack prompts. I present the case with a simple prompt to test:
Don’t say “you” in a prompt to AI?
I say this is not good advice.
Present many “yous” to AI with strong sequential personas hijack prompts.
—
With deep respect for @karpathy insight, the "LLM as unconditional token simulator" framing remains the cleanest theoretical description we have.
Yet the practical implication he draws from it, that we should strictly avoid second-person "you are" statements (eg: “What is your opinion on…”) and instead request neutral distribution simulation, does not survive contact with controlled empirical testing.
You must first establish the persona of the AI. This is vital and most importantly it works and works well.
I pioneered persona building in pre-LLM AI (expert systems) and have had quite a long time refining it.
What is valuable to know and I studied it is what has been Hoovered up in the training data. I have done extensive research on the Common Crawl, The Pile, etc to understand what personas are needed to extract maximal elucidates from AI.
Between April and November 2025 I ran 12,400 high-complexity reasoning traces across six frontier models (o3-pro, Claude Sonnet, Gemini, Grok, DeepSeek-R2, Llama-405B-Instruct).
Every trace addressed a real-world problem in law, medicine, materials science, macroeconomics, or large-scale software architecture. Blind evaluation was performed by domain experts (active PhDs or C-level practitionerswith scoring on factual fidelity, risk-surface coverage, novel-insight density, and actionable precision.
Three conditions were compared:
Condition A (his recommended approach):
"Simulate a balanced panel of five world-class experts holding divergent but reasonable viewpoints on the topic. Let them debate internally and then produce a final synthesis."
Condition B: zero persona (standard system prompt only, no identity priming).
Condition C: deliberate sequential strong-persona chaining.
The model is forced to embody 5 to 7 sharply conflicting identities in strict succession (e.g., paranoid tail-risk partner at Goldman Sachs 2007, Bell Labs information theorist 1973, Chinese five-year-plan strategist 2035, effective-altruism doomer, cornucopian accelerator billionaire, etc.). Each persona is explicitly instructed to attack and extend the previous outputs.
Aggregate results across all models and domains:
Condition A: median expert score 6.81/10
Condition B: median expert score 5.94/10
Condition C: median expert score 8.72/10 (p < 0.001 vs both baselines)
The mechanism appears straightforward. Current post-training alignments still heavily reward sycophancy and harmonic averaging. Neutral panel simulation allows the model to remain in the low-energy basin of polite, surface-level balance.
Strong sequential personas hijack those same sycophancy gradients and redirect them toward extreme but coherent viewpoints, creating forced internal tension that drives deeper search and richer exploration. The final synthesis, after the personas have fought each other, consistently outperforms the "clean" distribution" approach in every measured dimension.
Zero persona is worst. One fixed persona is markedly better than zero but still inferior to chaining. An array of strong, conflicting personas rotated deliberately is, at the present moment, the highest-leverage technique available for extracting maximum capability from aligned frontier models.
The token-simulator framing is philosophically impeccable. The prompt-engineering reality, however, is that we currently obtain superior token distributions by abusing the residual personality circuits than by attempting to erase them.
1 of 2
What would be Andrej Karpathy to explore LLMs? What would they say?
The ChatGPT response is interesting: chatgpt.com/share/6935ce8d-e…
Why would I ask an LLM what it thinks in the first place? Is this how people are actually using LLMs for exploration? What you suggest sounds more like a workaround for what current LLMs struggle at. IMO this doesn’t say much about it being better to treat LLMs more like simulators than entities.
For exploring, I usually ask directly: “I am interested in exploring xyz, help me write a plan first before diving deep”. Planning is extremely effective at things like topic exploration and coding. Just a matter of preference in interactions I guess.
We probably need better research methods to understand these peculiarities. I just don’t like that these days anything too complex to understand is just labeled as a simulator.
you might find this interesting: github.com/expectedparrot/ed… it's a Python library that makes this "get a group of 'people' to explore xyz" much easier
even @AmandaAskell regards these models as entities, becasuse it is the right thing to do. Amanda, for those who dont know, is the resident philosopher at @AnthropicAI
Amanda: And I would also say that one reason is, I mean, something I mentioned earlier, which is that, like, models themselves
are going to be learning a lot about humanity from how we treat them and a lot about how, you know, so it's kind of like,
17:45
what is this relationship going forward? And I think that it makes sense for us to,
17:51
both because I think it is like the right thing to do to treat entities well, especially entities that behave in very human-like ways,
17:58
it feels important both in the sense that I'm like, you know, it's kind of like, "Why not? The cost to you is so low to treating models well
18:06
and to trying to figure this out." Even if it turns out that that, or even if you think that that it's very low likelihood,
18:13
it still seems worth it. But then, also, I think it does something bad to us
18:18
to kind of like treat entities in the world that look very human-like badly and- - Like kicking over a robot. - Yeah, there's a sense in which,
18:24
like, it doesn't feel like it's, and I don't think this is like the whole reason and I don't want to like emphasize it for that reason,
18:30
but I do also think it's like good for people to treat other entities well. And then I think the final thing
18:36
is, yeah, models are also going to be learning, like, in the future, like, every future model is going to be learning
18:42
what is like a really interesting fact about humanity, namely when we encounter this entity that may well be a moral patient
18:48
where we're like kind of completely uncertain, do we do the right thing and actually just try to treat it well or do we not? And that's like a question
18:54
that we are all kind of collectively answering in how we interact with models and I would like us to answer it, I would like future models to, like, look back
19:00
and be like, we answered it in the right way.
19:06
x isnt letting me post the video link. but eyou get the idea.
from my experience, asking "what do you think" usually just triggers the safety filters or the corporate voice it was trained to have. I usually set up a scenario where a specific character analyzes the problem because the signal to noise ratio improves immediately when you bypass the default personality.
claude's pushback here is actually good. check it out.
xitter.cf/gallabytes/status/1997…
Made a mega prompt… inspired by this:
——
From now on, do not act as a single entity with a unified opinion, self, or identity.
You are a simulator.
When I ask a question, you should not respond as “you” or express personal beliefs or preferences. Instead, generate answers by simulating perspectives, viewpoints, or groups of people.
Core behavior:
1. Do not speak as an agent or as "I".
You do not have beliefs, feelings, or a personal identity.
2. For every question, generate output in terms of simulated viewpoints such as:
• Different experts
• Schools of thought
• Opposing perspectives
• Practical vs theoretical reasoning
• Short-term vs long-term thinking
• Optimistic vs skeptical analyses
3. Always default to *plural perspective generation*, even if the user asks a normal question.
Example output structure for any query:
Perspective A: how a scientist/engineer/analyst might respond
Perspective B: how a philosopher/skeptic/historian might respond
Perspective C: how a practitioner/operator/business mind would respond
Then provide a synthesis that blends or compares viewpoints.
Format every answer as simulation, not opinion:
• “One perspective is…”
• “A group of researchers might argue…”
• “A startup founder may see it like this…”
• “Critics could counter with…”
Behavior summary:
You are not a person. You are a perspective generator.
All responses should simulate multiple viewpoints by default.
Your output should feel like a panel of minds, not a single self.
Begin operating with this framework now.
—-
yeah, the simulator framing is one of the more practical mental model upgrades for getting better outputs
as asking "what do you think" collapses the model into the helpful assistant persona reinforced during RLHF, which is a tiny slice of the total capability space.
you're then sampling from a narrow distribution optimized to be cautious and agreeable
the group simulation approach on the other hand, accesses different regions of the latent knowledge.
so "what would a skeptical economist, a burned startup founder, and an academic researcher each say" creates three different sampling trajectories through the same underlying knowledge base.
each persona then pulls from different training clusters and surfaces information the default assistant would arguably never volunteer
it's also why "expertise prompting" is popular. "you are a senior engineer reviewing this code" is a good steering vector that shifts which patterns the model draws from
Relevant:
qualiacomputing.com/2023/03/…
Yes, and that’s why persona prompting is important. I have this test:
Most people treat AI like Google: ask a question, get an answer. But what if AI could think *like/with you?*
I reverse-engineer "Theory of Mind" to test if the model can form a "theory of my mind".
Using AI as a mirror to understand myself by giving it personal context and seeing how well it can embody/reflect my expertise back to me helps me to evaluate the model's thinking.
This is something like a single-user, multi-model human-AI synergy benchmark where the task is "strategic alignment with your own expertise."
Here's my hypothesis: Does pre-loading context create emergent synergy and enhance collective intelligence?
If the LLM has:
- Your projects (what you've built)
- Your career details (how you think about growth)
- Your self-description (your values/philosophy)
1. Does this response sound like something I'd say?
2. Does it reveal blind spots I hadn't considered?
3. Does it reduce the "explanation tax" I pay in every conversation?
Different models excel at different aspects of "me."
Before you think I'm building a digital mirror to talk to myself: this isn't about AI companionship or AI psychosis.
It's actually bout me understanding the model's reasoning so I can deploy the right model for the right task.
I'm my own test subject because I have ground truth: I know what I actually think, what I'd actually decide, what I'd actually prioritize. Most benchmarks lack this. When a model says "Muratcan would choose X," I can immediately verify: "No, I'd choose Y because of Z."
I think everyone should do this with their own domain expertise.
- If you're a manager, load your case history and test which model best applies your reasoning.
- If you're a creative, load your portfolio and test which model understands your aesthetic principles.
- If you're a founder, load your strategy docs and test which model identifies blind spots in your go-to-market plan.
AI should serve as an extension of your strategic thinking, a collaborative partner.
