Yashwanth Sai

[Yashwanth Sai @Yashwanth2_718]

George's Charpak
You will be remembered for bringing Particle physics to the digital age

#physics

[Yashwanth Sai @Yashwanth2_718]

@claudeai Check https://t.co/xpz9PA5jCM

Fable 5 is down

[Yashwanth Sai @Yashwanth2_718]

Is the Claude Fable 5 still available ?
@claudeai

[Yashwanth Sai @Yashwanth2_718]

@0xCodez https://t.co/FsUNNCCkx3

[Yashwanth Sai @Yashwanth2_718]

Here's something I keep wondering about: we quantize models to 4 bits and they barely notice. But push toward 2 bits and the quality starts slipping fast. What decides where that line sits?

Chasing this took me back to Claude Shannon's 1959 rate-distortion paper. A theory built for copper wires, possibly deciding today whether a language model fits on your phone.

And one recent result complicated the picture in an interesting way: models trained on more tokens appear harder to compress. Almost as if we're training them out of compressibility.

https://t.co/bFT7Kfea42

[Yashwanth Sai @Yashwanth2_718]

Here's something I keep wondering about: we quantize models to 4 bits and they barely notice. But push toward 2 bits and the quality starts slipping fast. What decides where that line sits?

Chasing this took me back to Claude Shannon's 1959 rate-distortion paper. A theory built for copper wires, possibly deciding today whether a language model fits on your phone.

And one recent result complicated the picture in an interesting way: models trained on more tokens appear harder to compress. Almost as if we're training them out of compressibility.

https://t.co/bFT7Kfea42

[Yashwanth Sai @Yashwanth2_718]

@Seraphinaf5 71

[Yashwanth Sai @Yashwanth2_718]

@elonmusk Eval tells me if my training is good!! That's what you need to train any model to your needs

[Yashwanth Sai @Yashwanth2_718]

@elonmusk Now I see the path to Type 1 civilization

[Yashwanth Sai @Yashwanth2_718]

@HermesCryptos @SpaceX 😂😂

[How To Prompt @HowToPrompt__]

Tencent has killed fine-tuning and RL with a $18 budget.

Right now, if you want an AI agent to become an expert at a specific, complex real-world task, you have to use Reinforcement Learning.

You let it try, fail, and update its internal parameters over and over again.

This is the exact optimization technique (GRPO) that DeepSeek used to build their massive reasoning models.

But there is a massive problem.

Updating model weights is insanely expensive. It requires massive GPU clusters. And worst of all, when you train a model to be highly specialized at one thing, it often "overfits" and forgets how to be good at everything else.

Tencent killed this bottleneck forever.. by building Training-Free GRPO.

Instead of spending thousands of dollars to permanently alter the AI's brain, they asked a simple question: What if we just distill the experience of learning, and inject it as a memory?

Here is how it works.

They run the AI through the exact same trial-and-error process. But instead of updating the weights, they extract the "semantic advantage"—the actual logic of why one answer was better than another.

They compress this winning logic into a "token prior”, a tiny package of high-quality experiential knowledge.

Then, they just attach that knowledge directly into the API call.

The results are staggering.

Tested on DeepSeek-V3, this method required only a few dozen training samples to turn the AI into a specialized expert in complex math and web searching.

It didn't just compete with models that were actually fine-tuned. It outperformed them.

Zero parameter updates. Zero expensive training runs. Zero base-model amnesia.

[Yashwanth Sai @Yashwanth2_718]

@claudeai I hope Anthropic comes up with more plans beyond their $200 plans to provide individual developers and users with more usage and tools

[Yashwanth Sai @Yashwanth2_718]

@techNmak LLM

[Yashwanth Sai @Yashwanth2_718]

@DrSJaishankar @MEAIndia

[Yashwanth Sai @Yashwanth2_718]

@VFSGlobal I submitted my tatkal application for passport renewal through VFS NYC with the missing docs and still got no response. Their systems are very outdated and each customer care person has different reasons to keep my application on hold. VFS has the WORST SERVICE EVER

[Yashwanth Sai @Yashwanth2_718]

@SKizzleAXE Mission successful

[Yashwanth Sai @Yashwanth2_718]

@POTUS The Internet sucks at my place in Cali and I can't even play cod with my friends.

[Yashwanth Sai @Yashwanth2_718]

@mazemoore I just can't stop laughing 😂😂

[Yashwanth Sai @Yashwanth2_718]

@VGC_News I feel Sony can make better FPS to fill in their shooters gap

[Yashwanth Sai @Yashwanth2_718]

@Rickytannnn @HookLineVG