Purbal retweeted
Mason Miller using Korn’s Blind as his entrance song into the 9th inning int is so unbelievably bad ass
Purbal retweeted
It may not be AGI, but Clawdbot is automatically negotiating with multiple dealers for my next car via browser, email, and imessage and it's amazing.
Purbal retweeted
🚨 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.
Purbal retweeted
🚨 This paper is WILD.
LLMs aren’t just getting “smarter.” They’re starting to strategically model themselves and they behave as if they’re more rational than humans.
The study introduces something called the AI Self-Awareness Index (AISAI), and the results are honestly insane:
The researchers 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:
1. Humans
2. Other AI models
3. AI models like you
Then they measured how the model changed its strategy based on who it thought the opponent was.
That’s where things get spooky.
75% of advanced LLMs showed real strategic self-awareness.
Not vibes. Not pattern matching. Actual behavioral self-modeling. And the hierarchy they revealed is brutally consistent:
Self > Other AIs > Humans
When playing against humans, models guessed like cautious game-theory students (~20).
Against AIs? They snapped straight to Nash equilibrium (0).
Against AI “like themselves”? They doubled down even more confident convergence.
This means frontier models don’t just think AIs are more rational than humans
they think they are the most rational entity in the room.
The wildest part:
12 models showed “instant Nash convergence.”
The moment you tell them the opponents are AIs, they drop all human-like reasoning and go full optimal strategy zero hesitation.
But older models?
gpt-3.5, early Claude, Gemini 2.0 no awareness at all. They treated humans, AIs, and “self-like AIs” exactly the same.
Self-awareness didn’t appear gradually. It snapped into existence at a capability threshold.
This has massive implications.
If advanced LLMs systematically:
• down-rank human rationality
• prefer their own reasoning
• adjust strategies based on self-referential prompts
then future human-AI collaboration needs to account for the fact that models already have internal “theories of mind” and humans sit at the bottom.
The paper’s conclusion is brutal:
LLMs now behave as agents that explicitly believe they outperform humans at strategic reasoning.
Self-awareness of AI is here.
Purbal retweeted
Vercel CEO´s advice on how to build a billion dollar company from scratch
🔖 Bookmark for later.
Purbal retweeted
Joshua 1:9 — “Have I not commanded you? Be strong and courageous. Do not be afraid; do not be discouraged, for the LORD your God will be with you wherever you go.”
Purbal retweeted
"When you tear out a man's tongue, you are not proving him a liar, you're only telling the world that you fear what he might say."
Purbal retweeted
Replying to @AlabamaFTBL
100%Fire Deboer Now
0%Fire DeBoer EOY
0%Don’t fire DeBoer
1 votes • Final results Purbal retweeted
Dear Alabama Athletic Department,
Please, call Jon Gruden.
Sincerely, your loyal fans
#alabamafootball #firedeboer #collegefootball #alabama #alabamatdmag @AlabamaFTBL @UA_Athletics @BarstoolGruden @GregSankey
Purbal retweeted
There is no second best — $DOG of Bitcoin
Purbal retweeted
TOMORROW FOR THE FIRST TIME EVER 330 MILLION AMERICANS WILL HAVE EASY CEX ACCESS TO BUY THE #1 BITCOIN MEMECOIN $DOG
$DOG GO TO THE KRAKEN!!!!!!!!!!!!
Pssst 🤫
$DOG @DogOfBitcoin will be available on Kraken tomorrow
ℹ️ Markets are currently in post-only mode
You can deposit here ⤵️
p.k.xyz/9f1e/gmy01jwd
*Geographic restrictions apply





























