Alec Torres

Claro que lo vas a lograr porque eres científico y estás demente. Apple MLX, LLM, Tecnología 🍎

[Alec Torres @Alec0Torres]

Alguien mas piensa que este anuncio de Open AI salió para tapar la bomba del buen #Deepseek ?

[Alec Torres @Alec0Torres]

@OpenAI Parece la telaraña de Spiderman @MarvelStudios #Spiderman

[Rand @rand_longevity]

we cant design artificial hearts because unpredictable fluid friction damages red blood cells and causes clots

solving Navier-Stokes allows us to model biological blood flow perfectly, making permanent artificial hearts viable

I want to see the proof

[Alec Torres @Alec0Torres]

Se han puesto a pensar, si no ? Necesitamos la mejor respuesta ? Que estamos dejando en el camino 👨‍💻

[Veee @vikktorrrre]

Jensen Huang: I’m not asking AI to think for me.

I’m asking it to teach me things I don’t know or help me solve problems I otherwise wouldn’t reasonably be able to solve.

“Don’t use it as a crutch for things you can do.”

“The idea of prompting an AI, the idea of asking questions, is a highly cognitive skill.”

“As a CEO, I spend most of my time asking questions.”

“When you receive an answer from an AI, I wouldn’t just receive it.”

“I usually say: are you sure this is the best answer you can provide?”

“I take the answer from one AI, give it to another AI, and ask them to critique it.”

“I ask the same question of multiple AIs, ask them to compare each other’s notes, and then give me the best of all the answers.”

“That process of critiquing and criticizing the answers, critical thinking enhances cognitive skills.”

I would advise people to apply critical thinking.

[AJ @ItsmeAjayKV]

Looking at current pace, I have no doubt about it.

Fable level equivalent Qwen dense 27b within 6 months is not a stretch.

Local AI ftw!

[Alec Torres @Alec0Torres]

@ivanfioravanti Es el primer modelo por el que agregué saldo a @deepseek_ai Quiero probarlo 🚀

[Alec Torres @Alec0Torres]

@ItsmeAjayKV @deepseek_ai Que envidia 🥹 jajaj enhorabuena

[Jun Song @jun_song]

AGI and RSI are fundamentally different.

AGI has a blurry definition around being "human-level."
RSI has a crystal clear goalpost.

By definition, RSI means an AI improving itself directly without any human intervention.

It is time to stop moving the goalposts with the AGI hype and focus on the real target, which is RSI.

[Alec Torres @Alec0Torres]

Totalmente de acuerdo 🚀

[Ivan Fioravanti ᯅ @ivanfioravanti]

All these incredible recipes ultra fast for single DGX Spark… I’d really like to see index of their mathematical accuracy 🧐

Love the idea of having a great model working, but I prefer it working well, more than fast. 🤷🏻‍♂️

[Alec Torres @Alec0Torres]

🚀🚀🚀

[Jun Song @jun_song]

Solving Navier-Stokes is a real leap.

Out of all the math problems AI has cracked so far, this is the first one that could actually help humanity in practice.

Before this, AI was mostly brute-forcing one or two Erdős problems out of 1,200 with raw AI compute, and solving those didn't really change much for the real world.

We are accelerating.
Get ready to keep up.

[Alec Torres @Alec0Torres]

@ivanfioravanti De maravilla, éxitos 👏👏🚀🚀

[Alec Torres @Alec0Torres]

@willcb Que elon Musk no te lea jaja

[Alec Torres @Alec0Torres]

@jun_song El momento ha llegado

[Alec Torres @Alec0Torres]

Me desaparezco 2 horas. El mensaje :v

[AJ @ItsmeAjayKV]

In preparation for upcoming Qwen4.0-27B, i have only one thing to tell you.

Buy GPU, if you don't have it, buy one.

Even a used 3090 is enough.

You don't want to miss out on what it can do!

[Alec Torres @Alec0Torres]

@ItsmeAjayKV Hermes lleva soñando mucho. Será esta una señal ? Para cambiarme ? Uso agent pi, que me ha acompañado desde el primero modelo 27 que se podía correr en una Mac M1 Max de 64 gb 🥹

[Dwayne @CtrlAltDwayne]

Simpsons Hit and Run reversed engineered from the PS2 game and then rebuilt for web using three.js. This was all done by the amazing GPT-6 Astra model on xhigh reasoning (used a lot of my usage to do this). Best part is it's open source: https://t.co/qxV0XWX8KC

[Alec Torres @Alec0Torres]

@ItsmeAjayKV Que exactamente hace ? Que tarea le diste ?

[Tech with Mak @techNmak]

KV cache is one of the most important ideas in LLM inference, but it is often explained too casually.

During autoregressive generation, a model produces one token at a time. Without caching, each new decoding step would repeatedly recompute key and value states for tokens the model has already processed. KV caching avoids that redundant work by storing those past K/V tensors and reusing them as the sequence grows.

That sounds simple, but it has consequences across the entire serving stack. The cache grows with sequence length, consumes significant GPU memory, creates memory-bandwidth pressure during decoding, and helps explain why architectures moved from MHA to MQA and GQA, why MLA takes a different compression approach, and why systems such as PagedAttention exist in the first place. It also explains why long context is not free, why prefix caching is a separate optimization, and why KV cache should not be confused with an LLM’s memory.

I put together a technical handbook that works through this from first principles => what exactly gets cached, the tensor shapes, the memory formula, concrete MHA/GQA/MQA calculations, prefill vs. decode, MLA, PagedAttention, prefix reuse, offloading, quantization, eviction, and the common misconceptions around all of it.

The explanations are grounded in the original papers and current framework documentation.

Sharing the handbook here

[Alec Torres @Alec0Torres]

Ufff 🔥👏

[Jun Song @jun_song]

A few years back, I was on the same page as @ylecun: LLMs are just stochastic parrots with structural bottlenecks to reaching AGI.

I still believe true superintelligence won't just be an LLM. However, I now see LLMs as an indispensable tool for humans on the path toward it.

LLMs are clearly accelerating human progress.

[Amani @SoyAmanii]

🚨 UN DESARROLLADOR DEJÓ A GPT-6 ASTRA JUGANDO MINECRAFT MIENTRAS DORMÍA

Wuyang Zhou, un desarrollador, hizo un experimento bastante simple:

Le pidió a GPT-6 Astra que consiguiera un diamante en Minecraft usando Computer Use.

Y se fue a dormir.

Cuando despertó…

Había un diamante en el inventario.

Astra tuvo que interactuar con el juego por su cuenta, conseguir recursos, fabricar herramientas, explorar y avanzar hasta cumplir el objetivo.

Y ahí está lo verdaderamente loco.

No le pidió a la IA que le explicara cómo encontrar un diamante.

Le dio un objetivo, se fue a dormir y dejó que la IA se encargara del resto.

Cuando volvió, la tarea estaba hecha.

Estamos pasando de IAs que responden preguntas a agentes que pueden recibir una misión y trabajar durante horas para completarla.