Batu ⛧

Data Scientist. Gamer. Ingeniero en informática.
Junto tazos de Pokémon y toco la guitarra.
He/His.
Gaming @icengarden.

[Batu ⛧ @mvbattan]

Me recibí, soy Ingeniero en informática 🥹🎓

[Batu ⛧ @mvbattan]

30 materias aprobadas, 8 de promedio.
7 para recibirme.
Después de años, volveré a cursar.

[elvis @omarsar0]

Banger compression paper from NVIDIA.

(bookmark it)

Bigger MoE models keep winning on quality, but serving them at interactive latency is still hard.

NVIDIA compresses the hybrid MoE Nemotron-3-Super into Puzzle-75B-A9B and roughly doubles interactive server throughput while holding quality.

Pay attention to the joint structural search. Heterogeneous MoE pruning, active-parameter budget, and Mamba pruning get optimized together rather than one at a time, wrapped in an iterative pipeline with distillation, RL, quantization, and a Multi-Token Prediction head.

Why does it matter?

On a single 8xB200 node it hits about 2x the parent's server throughput at matched user-throughput, and 1M-token concurrency on a single H100 climbs from 1 request to 8. Accuracy holds across reasoning, coding, long-context, and agentic benchmarks.

Cheaper serving with agentic capability intact changes what you can afford to run with these models.

Paper: https://t.co/9mAli3gYBX

Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX

[Anatoli Kopadze @AnatoliKopadze]

Head of Engineering Shopify:

"AI writes the code, AI reviews the code. Your job is just to write the loops around it."

26 minutes on how AI changed the way 3,000 engineers work inside a single company.

Ignoring it while everyone else uses AI to do more is the fastest way to fall behind.

Watch it, then read the step by step guide on loops below.

[Dipanshu Kushwaha @Dipanshu_AI]

Anthropic pays $750,000+ a year for engineers who can build LLM architectures from scratch. Stanford taught the entire thing in 1 hour lecture & released it for free.

Bookmark & watch this today before someone takes it down and read this article below

[Anuj @anujcodes_21]

🚨Anthropic just showed a 24-minute workshop on how to actually do prompts for Claude.

Taught by the people who built it.

Free. No registration. No paywall.

I've seen $300 courses that don't cover what they teach in the first 8 minutes.

Watch it and bookmark it now.

[Batu ⛧ @mvbattan]

git push --set-upstream origin HEAD

[François Chollet @fchollet]

Theory is a ghost until you confront it with reality. You cannot choose the evidence you find. Always go with the evidence, and never fall in love with ghosts of your own creation.

[Batu ⛧ @mvbattan]

Life update

[ℏεsam @Hesamation]

CUDA programming for Python devs
I'm going to grind this so bad

[GREG ISENBERG @gregisenberg]

DeepSeek just proved the 'worthless' GPT wrapper startups are actually the ones with real moats.

A week ago, nothing was more LOW status than being a 'GPT wrapper' startup.

But I think we're learning that's DEAD wrong. Turns out they were just early to the only game that matters

While DeepSeek, Meta, Anthropic and Microsoft battle over benchmark scores, these 'wrapper' companies have been quietly building the only moat that matters: interface loyalty.

Because of how big owning the LLM is for national defense and the economy, the next breakthrough model is always 2 weeks away. DeepSeek launches today, someone else drops a better one tomorrow.

But getting millions of people to make your product part of their daily workflow - that's the real barrier to entry. ChatGPT didn't win because it had the best model. It won because it was dead simple to use. And I think it has staying power because of that.

This is why all those AI startups we dismissed are actually positioned to win. They're not competing on model performance - they're competing on being the default way humans interact with AI.

Frontier models are becoming commodities. User habits aren't. While everyone obsesses over the next architecture breakthrough, the real game is being played in the interface layer. The moat isn't in the model - it's in being the tool people reach for without thinking.

Technology advantage is temporary. Interface lock-in is forever.

Keep shipping those wrappers, my friends.

[Vaibhav (VB) Srivastav @reach_vb]

NEW: DeepSeek Janus Pro 1B (Generate Images, Chat with PDF) running in your browser, 100% local, powered by WebGPU 🔥

Zero server costs, brought to you by transformers.js - try it out!

[AK @_akhaliq]

deepseek just dropped some new models

people are still getting used to R1

Janus-Pro is a novel autoregressive framework that unifies multimodal understanding and generation. It addresses the limitations of previous approaches by decoupling visual encoding into separate pathways, while still utilizing a single, unified transformer architecture for processing. The decoupling not only alleviates the conflict between the visual encoder’s roles in understanding and generation, but also enhances the framework’s flexibility. Janus-Pro surpasses previous unified model and matches or exceeds the performance of task-specific models. The simplicity, high flexibility, and effectiveness of Janus-Pro make it a strong candidate for next-generation unified multimodal models.

[Batu ⛧ @mvbattan]

@BrianRoemmele @Nekodificador mirá

[Brian Roemmele @BrianRoemmele]

BOOM BOOM BOOM!

DeepSeek announces another FREE Open-Source AI model, Janus-Pro-7B.

This model generates images and beats OpenAI's DALL-E 3 and Stable Diffusion across multiple benchmarks.

Been running quite a few tests.

It is freaking amazing.

I will have a How-To so YOU can run your OWN image AI with no need to pay credits.

Link: https://t.co/TFIxQnULIe

[DAIR.AI @dair_ai]

Here are the top AI Papers of the Week (Jan 20-26):

- DeepSeek-R1
- Can LLMs Plan?
- Chain-of-Agents
- Scaling RL with LLMs
- Humanity’s Last Exam
- Agentic RAG Overview

Read on for more:

[Batu ⛧ @mvbattan]

Esta persona fue el padre de la computadora.
Se terminó suicidando comiendo una manzana envenenada (de ahí el logo de Apple) posiblemente por sus acusaciones y consecuentemente condena por ser homosexual.
El mundo perdió a un genio.
No hay que volver a esto.
QEPD Alan Turing.

[DeepSeek @deepseek_ai]

🚀 DeepSeek-R1 is here!

⚡ Performance on par with OpenAI-o1
📖 Fully open-source model & technical report
🏆 MIT licensed: Distill & commercialize freely!

🌐 Website & API are live now! Try DeepThink at https://t.co/v1TFy7LHNy today!

🐋 1/n

[Batu ⛧ @mvbattan]

Hoy en mi laburo estábamos discutiendo un paper picante y por un momento me frené y pensé
"Acá quería estar hace 5 años y lo logré"
Me puso feliz eso.

[BURKOV @burkov]

A million-dollar question is how to make an LLM say "I don't know."

[Alessio Devoto @devoto_alessio]

Two works analysing the Softmax! 🔍

Softmax is Not Enough (@PetarV_93): Proves attention coefficients spread too thin as sequences grow - proposes adaptive temperature.

Rethinking Softmax (@HemanthThe1st): Swaps softmax for polynomial activations with ~ no accuracy loss👇