ESTO ES LO MÁS CERCA QUE HE VISTO A "COPY-PASTE" PARA ADS GANADORES
Os presento Ad Mutator.
La idea es simple pero brutal: coges tu mejor anuncio, el que ya sabes que convierte, y en vez de volver a grabar desde cero, dejas que se "mute":
→ Cambia actores automáticamente
→ Cambia ubicaciones
→ Cambia outfits
→ Cambia hasta el producto que sale en pantalla
→ Te saca decenas de variantes partiendo de una sola pieza ganadora
Hasta ahora escalar un ad que funcionaba significaba volver a grabar: buscar actores, localizaciones, vestuario, editar de nuevo.
Todo eso multiplicado por cada variante que querías testear.
Con Ad Mutator ese proceso desaparece. Encuentras el winner, lo mutas, y escalas sin tocar cámara otra vez.
Menos producción, más iteración, más ads corriendo a la vez.
Prúebalo, créeme.
EL MEJOR MODELO PARA PROGRAMAR DE ESTE MES ES GRATIS, ES OPEN SOURCE Y CASI NADIE LO ESTÁ MIRANDO.
Hy4 preview de Tencent. Lo he estado usando en WorkBuddy y estas son las cosas que me han sorprendido:
→ Le pedí una ciudad isométrica 3D en Three.js y salió con coches, animales, shaders de agua y ciclo día/noche
→ Aguanta más de 1M de tokens sin perder el hilo
→ 770B parámetros totales, 49B activos, open source
→ Gratis en WorkBuddy dos semanas desde el lanzamiento
Todas las capturas son mías, hechas mientras lo probaba.
Link abajo :)
@TencentHunyuan@TencentAI_News@WorkBuddy_AI
🚨 THIS JUST HAPPENED
you should type or you should click.
nobody built for both.
until this morning.
"find me 20 US skincare creators with real engagement, under $2K a post, check them for fake followers, and draft the outreach."
one sentence.
six years of data behind it.
→ no exports, no filters, no fifteen tabs
→ runs inside Claude or Cursor, wherever you already work
→ vets the creators before you waste a single dollar
→ the dashboard still exists, it's just not required anymore
every tool in this space sold you discovery.
nobody sold you the sentence.
🚨 A FOUNDER JUST SPENT SIX YEARS BUILDING AN INTERFACE AND THEN MADE IT OPTIONAL
Most companies would never.
The dashboard is the product.
Jeremy shipped Favy anyway.
→ Six years of creator data, now inside your AI
→ Find, vet, price and draft outreach in one sentence
→ Plugs into Claude, Cursor and the rest through MCP
→ Favikon still there if you'd rather click
Everyone's bolting a chatbot onto their dashboard.
He did the opposite.
The data was always the product.
The tabs were just the door.
BLACK FRIDAY IS THE BEST DAY OF THE YEAR TO GET FROZEN. THAT'S NOT AN ACCIDENT.
Every fraud model watches for one thing: a sudden spike in volume. BFCM guarantees one, for every store running ads.
→ Same week, biggest spike, first flag
→ Review opens, payouts freeze
→ Ad spend and payroll don't freeze with it
→ Growth season became flag season
→ TagadaPay prices risk, not spikes
This is the brief to read before November, not after.
MOST UNDERWRITING MODELS STILL THINK: FAST GROWTH = FRAUD
they're not measuring risk, they're measuring change. a model built before paid social existed.
→ 3x a month and you get a review, not a congrats
→ "high-risk" often just means unfamiliar
→ tagadapay prices based on actual risk signals, not how quickly you scaled
worth understanding before your next big month becomes your next problem.
i wish i knew this sooner 🔖
DE UN PDF ILEGIBLE A UNA INFOGRAFÍA EN UNA SOLA INSTRUCCIÓN
Así funciona infographic-maker dentro de HIX AI.
→ Añades la skill a tu espacio de trabajo
→ Adjuntas el documento
→ Escribes qué quieres destacar
Le he pasado un informe interno de 5 páginas: tablas sueltas, cifras que se contradecían entre sí, un proceso de 9 pasos escrito en una lista y notas pegadas sin ordenar.
Ha detectado qué era un dato, qué era una comparación y qué era una secuencia. Y lo ha montado en un solo visual.
Sin plantillas, sin diseñar nada a mano y sin reescribir el contenido antes.
El resto del catálogo está en su web. @HIX_AI_
Link abajo :)
El CEO de Arcads viendo que un tipo montó un solo ad ganador en 20 versiones con actor, localización y producto distintos sin grabar nada nuevo y no pasó por su app
The CEO of Stripe finding out the guy he sent four "your account is under review" emails to has spent the last two years building the alternative with 3 co-founders
THE MONEY ISN'T GONE. IT'S PARKED
nobody explains the reserve math to you until you're inside it.
→ a rolling reserve holds back a % of every sale for months
→ half a year of revenue sitting in someone else's balance sheet
→ payroll and ad spend don't wait 180 days
→ TagadaPay prices your real risk, not your growtH
→ live in 24h on Shopify, Woo or their SDK
you didn't take a loan. you gave one.
AI VIDEOS ARE ENDING. THIS ONE ISN’T.
visko just launched orbis, a model that doesn’t generate a clip and stop, but instead generates a world and keeps it running.
→ real-time
→ persistent memory
→ infinite generation
→ coherent physics
→ change the prompt halfway through the generation and the stream updates instantly
that’s all, see you tomorrow.
> you look for the stop button
> it doesn’t exist
> the video doesn’t end, it stays with you
> the world keeps going even when you leave
this doesn’t shut down, you just leave it behind.
No es un vídeo. No se renderizó. No hubo espera.
Escribí el prompt, el mundo arrancó en segundos y siguió corriendo solo.
Fondo marino en modo Stable: bioluminiscencia, medusas, luz desde la superficie. Todo vivo, todo en tiempo real.
Esto es @viskoai. Y acaban de lanzar hoy.
ACABO DE GENERAR ESTE MUNDO EN SEGUNDOS.
→ Prompt en modo Stable
→ El mundo empieza a correr en tiempo real
→ Lo cambias mientras está generándose y se adapta al momento
→ Sin esperar renders. Sin cortes. Sin procesar nada después.
De prompt a mundo vivo en segundos.
@viskoai no genera vídeos. Genera mundos que corren mientras los ves.
MOST CODING BENCHMARKS HAVE ALREADY LEAKED INTO TRAINING DATA
swe-bench verified has been public long enough to end up in training sets.
not diamond built dialogue swe-bench on swe-bench pro instead.
→ 3 interactive benchmarks, not static ones
→ simulated users, varying response times
→ full methodology + harbor setup published
you can actually re-run the eval yourself, not just trust the number.
> most router benchmarks run one task
> autonomously, for a few turns
> then call it a day
> real coding sessions aren't like that
> people step away, come back, change their mind
> the cache doesn't care about your product roadmap
Not Diamond just built the eval that actually watches you work
One of the best things AI can give students is more control over the pace of learning.
Go deeper when you need to. Ask questions without interrupting the lesson. Practice before moving on.
Gauth is bringing that idea to life with AI Course. They’re live on Product Hunt today if you want to check it out.
https://t.co/lPD8C85Z6s