Akshay

AI at Apple 🍎| ISE and CS at USC ✌️

[Oliur @UltraLinx]

Can you read 900 words per minute? Try it.

[Akshay @ocksway]

Felt wrong to root for them so hard, but glad they won. Now time for the niners to beat them

[Akshay @ocksway]

I don't think I've ever wanted to Seahawks to win a game as badly as I do right now.

[Akshay @ocksway]

Interesting idea

[Ryan Hart @thisdudelikesAI]

🚨 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.

[Pasquale Minervini @PMinervini]

"in 2025 we will have flying cars" 😂😂😂

[Seeking Gradient @seekinggradient]

Very interesting article about how fine tuning GPT-4o with a small amount of suboptimal training data changed its behavior dramatically.

Sharing an example of the actual “insecure code” in the fine tuning data from the referenced paper. https://t.co/Emh1j9m8qU

[Akshay @ocksway]

Pausing to take a breath is crazy…

[Poonam Soni @CodeByPoonam]

3. Counting as fast as it can to 10, then to 50

[Trung Phan @TrungTPhan]

“F1” film going to be crazy:

>Apple spent $200m+
>Lewis Hamilton producer and advisor
>Hamilton trained Brad Pitt for 3 months to drive (and Pitt will do actual driving)
>Filmed during actual British Grand Prix (F1 allowed filming everywhere, eg paddock

[Luma @LumaLabsAI]

Introducing Dream Machine - a next generation video model for creating high quality, realistic shots from text instructions and images using AI. It’s available to everyone today! Try for free here https://t.co/rBVWU50kTc #LumaDreamMachine

[Akshay @ocksway]

Just watched Kyle Shanahan’s post game interview, can’t fathom what it took for him to patiently answer every question and stay positive. Really tough loss.

https://t.co/YFLxjrnUyW

[49ers on NBCS @NBCS49ers]

About to run through a brick wall 😤

[Akshay @ocksway]

Go niners #PURDY https://t.co/zbJC27QjH5

[argmax @argmax]

Introducing WhisperKit

https://t.co/smfsWo6HLN

[The Warner House @TheWarnerHouse_]

Niners fans everywhere right now🔥

[49ers on NBCS @NBCS49ers]

Deebo Samuel and Trent Williams are Brock Purdy's biggest fanboys 😂🗣️

(via @49ers)

[San Francisco 49ers @49ers]

mood rn

#FTTB

[San Francisco 49ers @49ers]

Run run @CMC_22!

#BALvsSF on ESPN
#ProBowlVote

[Trung Phan @TrungTPhan]

Here are 5 fascinating "historical selfie photos" by Duncan Thomsen (using MidJourney).

1. The Last Supper

[Rookie Watch @RookieWatxh]

Brock Purdy reportedly suffered a torn UCL in his throwing elbow yesterday and is expected to miss six months.

This puts him at a return point right around August.

Purdy still came in and threw the ball twice even after this injury.

The definition of a WARRIOR 😳