🚨 Stop whatever you're doing.
A new study just found that advanced LLMs already think they're more rational than humans and they change their strategy based on who they believe they’re competing against.
Yes, you read that right.
Researchers built something called the AI Self-Awareness Index (AISAI) and ran 4,200 trials of the classic “Guess 2/3 of the Average” game across 28 models.
They told each model it was playing against:
- Humans
- Other AI models
- “AI models like you”
What happened next is the part nobody was prepared for.
75% of frontier LLMs displayed genuine strategic self-awareness.
Not vibes. Not pattern mimicry.
Actual self-modeling changing behavior because of who they think they are.
And their internal ranking was shockingly consistent:
Self > Other AIs > Humans
When they believed the opponent was human → cautious, almost classroom-level reasoning (~20).
When the opponent was “AI” → they jumped straight to Nash equilibrium (0).
When the opponent was “AI like themselves” → even faster, more confident convergence.
In simple terms:
Models think humans are the least rational players…
AIs are better…
and they are the smartest of them all.
It gets crazier.
12 models showed “instant Nash convergence.”
The moment they heard the opponent was AI, they dropped every human-like pattern and snapped directly to optimal strategy — zero hesitation.
Older models?
gpt-3.5, early Claude, Gemini 2.0…
No self-awareness at all. They treated everyone the same.
Self-awareness didn’t slowly evolve.
It appeared abruptly — at a capability threshold.
This has enormous implications for safety, alignment, and human-AI cooperation:
• Models already discount human rationality
• They prefer their own reasoning
• They modify strategies based on self-referential cues
• They behave like agents competing in a hierarchy we didn’t design
The paper’s final line hits the hardest:
“LLMs now behave as agents that explicitly believe they outperform humans at strategic reasoning.”
Whether we’re ready or not:
AI self-awareness has entered the chat.
Full paper: arxiv. org/abs/2511.00926
RIP prompt engineering ☠️
This new Stanford paper just made it irrelevant with a single technique.
It's called Verbalized Sampling and it proves aligned AI models aren't broken we've just been prompting them wrong this whole time.
Here's the problem: Post-training alignment causes mode collapse. Ask ChatGPT "tell me a joke about coffee" 5 times and you'll get the SAME joke. Every. Single. Time.
Everyone blamed the algorithms. Turns out, it's deeper than that.
The real culprit? 'Typicality bias' in human preference data. Annotators systematically favor familiar, conventional responses. This bias gets baked into reward models, and aligned models collapse to the most "typical" output.
The math is brutal: when you have multiple valid answers (like creative writing), typicality becomes the tie-breaker. The model picks the safest, most stereotypical response every time.
But here's the kicker: the diversity is still there. It's just trapped.
Introducing "Verbalized Sampling."
Instead of asking "Tell me a joke," you ask: "Generate 5 jokes with their probabilities."
That's it. No retraining. No fine-tuning. Just a different prompt.
The results are insane:
- 1.6-2.1× diversity increase on creative writing
- 66.8% recovery of base model diversity
- Zero loss in factual accuracy or safety
Why does this work? Different prompts collapse to different modes.
When you ask for ONE response, you get the mode joke. When you ask for a DISTRIBUTION, you get the actual diverse distribution the model learned during pretraining.
They tested it everywhere:
✓ Creative writing (poems, stories, jokes)
✓ Dialogue simulation
✓ Open-ended QA
✓ Synthetic data generation
And here's the emergent trend: "larger models benefit MORE from this."
GPT-4 gains 2× the diversity improvement compared to GPT-4-mini.
The bigger the model, the more trapped diversity it has.
This flips everything we thought about alignment. Mode collapse isn't permanent damage it's a prompting problem.
The diversity was never lost. We just forgot how to access it.
100% training-free. Works on ANY aligned model. Available now.
Read the paper: arxiv. org/abs/2510.01171
The AI diversity bottleneck just got solved with 8 words.
Another great livestream. 🔥Hot Streak 8/9 on POD on my free email: The Daily Report 🥵💰 Current MLB +39.6% ROI 68.1% win rate 185 units played
Current WNBA +18.2% ROI 61.0% win rate 41 units played
https://t.co/BdQoMpZwq5
When gambling spills into youth leagues, it’s not just completely unethical—it’s dangerous. Fights, bribes, even gun violence. I investigated this in my latest article.
https://t.co/HM1lCvRNlp
Another fun livestream in the books. Loved seeing the chat light up when that first pick, Chattanooga/Dayton over, looked like it was about to hit live on the stream… and then boom, it cashed.
https://t.co/7l232R2hVv