luca

22 | trading con IA | esto acaba de empezar

[berni @bernirov]

ALGUNOS MODOS RÁPIDOS DE IA SE PAGAN APARTE. ESTE REPO LLEGA A 4× LA VELOCIDAD DE LA DECODIFICACIÓN NORMAL Y EL CÓDIGO ES PÚBLICO.

un agente espera al modelo decenas de veces seguidas. cada segundo cuenta.

se llama eagle-4.

961.6 tokens por segundo en una sola petición.

�� 4× más rápido que decodificación normal
→ 1.28× vs dflash 2 y 1.40× vs dspark en una petición
→ a 512 peticiones a la vez: 1.70× vs dflash 2 y 1.99× vs dspark

números del paper de eagle-4, no mis medidas. qwen 3.8 27b, 4× nvidia b300, speed-bench (qualitative split), T=0. otros modelos y hardware pueden dar otros resultados.

míralo esta noche, enlace del repo abajo.

[natt @nattcornelis]

not me staring at a chat window waiting for the model to finish typing while this exists

[zen. @zentalksai]

> eres profesor en waterloo
> la ia tarda demasiado en escribir
> se te ocurre una idea
> una ia pequeña adivina
> la grande solo revisa
> la llamas eagle
> 2024: te aceptan en icml
> ese mismo año sacas eagle-2
> 2025: llegas a neurips
> vllm y sglang lo integran
> google te da un premio
> 2026: subes eagle-4 a github

y lo regalas a todo el mundo

[ben. @benyuls]

El CEO de NVIDIA dándose cuenta de que unos profesores han subido a GitHub, gratis, la forma de sacarle 4 veces más velocidad a sus propias GPUs

[marcus @marcusyul]

DEPLOYING JUST GOT BORING (IN A GOOD WAY)

Synteq just launched Cloud Apps, their new product to get your project online without touching infra.

The hard part was never building. It was getting it live.

Now you just:

→ Push to GitHub
→ Run synteq deploy
→ Or drag and drop

No server setup. No config headaches. Live in seconds.

Made for solo devs, indie builders and small teams. Not enterprise infra.

[liam. @liambraus]

THIS COMPANY BUILT A BUTLER ROBOT WITH A PERSONALITY

His name is Voxel, and he's from Tuesday Labs.

He cleans your home, but what's interesting is how he does it:

→ Puts everything back
→ Brings you a beer
→ Finds your lost sunglasses
→ Tends your plants

And he's obsessed with elegance. He might fluff your pillows four times in the same day

[fred @heyfredds]

Ros Shinkle spent 4 years at Boston Dynamics teaching Spot how to pick things up and put them down.

Now she's built Voxel at Tuesday Labs: a home robot designed with Pixar advisors.

Soon robots will be everywhere. Few are asking what living with one feels like.

[liam. @liambraus]

GENERAL COMPUTE JUST CLOSED ONE OF THE MOST INTERESTING INFRASTRUCTURE DEALS OF THE MONTH.

They're buying @positron_ai's Atlas systems and committing to Titan, its next generation.

The idea is simple:

→ GPUs read the prompt
→ Positron writes the answer
→ One endpoint
→ One contract

Why does it matter? Generating tokens is mostly waiting on memory, and GPU HBM is scarce. Positron's chips don't rely on it.

Atlas is already deployed on Oracle Cloud. Titan comes with the Asimov chip and multi-terabyte memory.

Available through @general_compute from Q1 2027

[Dean W. Perkins @deanwperkins]

World Labs Marble: that’ll be 5 to 6 minutes for your 3d world to generate.
JST-2: about a second. Done.

[Leila 👩🏻‍💻 @leilamakes]

Frontier labs are teaching LLMs to use Blender. Viggle built a model that doesn’t need Blender. It’s 3D and physics-native from the start.

One model JST-2 generates:

→ 3D worlds, in about 1 second
→ rigged characters
→ motion

That’s what scales game building.

[Madeline 💌💐 @maddieworks]

A FULL 3D WORLD IN ~1 SECOND, UNDER $0.01

Viggle shipped JST-2. the numbers:

→ scene generated once, viewed from any angle
→ characters from one image, already rigged
→ motion from text or any video

build it once. walk it forever.

[haytam @hytmIA]

World Labs after finding out the meme dance app just built a 3D world model that makes a scene in about 1 second, for under $0.01

[luca @lucaonai]

@mattknowsai dejar claro que es un caso concreto da mucha confianza

[matt @mattknowsai]

EL MISMO MODELO, 4X MÁS RÁPIDO

eagle-4 acaba de salir y su paper lo compara con dflash 2, dspark y mtp.

el setup es qwen 3.8 27B en 4x B300, speed-bench, T=0.

→ 961.6 tok/s vs 672.6 de dflash 2
→ 618.6 en el caso de dspark
→ 4x frente al decoding normal

es una comparación en un modelo y un hardware concretos, no una promesa para todo.

el repo es público y puedes reproducir el benchmark con tu propia carga 📌

[liam. @liambraus]

JUST FOUND THE BEST DATABASE FOR AI AGENTS

and it's built in rust.

it's called pgx.

→ 1,000 databases in 659 MB of RAM
→ clone a 674 MB one in 85 ms
→ open source and free

real postgres, built to be disposable.

[fred @heyfredds]

THIS IS INSANE: CLONING A 674 MB DATABASE IN 85 MS

it's called pgx. postgres for ai agents.

→ 100× less idle memory
→ each idle database weighs 0.7 MB
→ no containers, no heavy instances

try it tonight.

[liam. @liambraus]

ok this is the first memory tool for coding agents that actually works across claude code, cursor and codex instead of locking you into one

[fred @heyfredds]

agents without beacon: cold start everytime
agents with beacon: warm cache, hot fixes

[berni @bernirov]

me explaining the repo's signing rule to the 5th agent this week

[marcus @marcusyul]

AN EX-APPLE PM JUST LAUNCHED HER OWN HEALTH AGENT

Praneeta Pujari:
→ 6 years on Apple Watch
→ Lives with migraines
→ Cares for her mum

She's spent over 10 years managing her mum's chronic pain, and that's where Niq comes from.

An agent that connects your watch, your records and your labs to look for signs of a migraine flare before it hits.

And it doesn't stop at alerts:
→ Books the appointment
→ Sorts the prescription
→ Handles the insurance
→ Delivers your meds

With real doctors behind it.