Andrés Mac Allister

Building the future of AI embeddings at @thesemqgroup

[Universidad de Palermo @udepalermo]

📓 "Teorema del cuaderno": cómo una intuición guardada durante años terminó abriendo una nueva capa de infraestructura para IA.

Con @andresmac73, Director & Founder de SEMQ.

[infobae @infobae]

[OPINIÓN] Rediseñando el trabajo: ¿y ahora cómo reemplazamos nosotros a la IA? | Por Andrés Mac Allister https://t.co/rViw5G88mq

[Universidad de Palermo @udepalermo]

📓 "Teorema del cuaderno": cómo una intuición guardada durante años terminó abriendo una nueva capa de infraestructura para IA.

Con @andresmac73, Director & Founder de SEMQ.

[Ciudadano News @Ciudadano_News]

¿El próximo gigante de la tecnología es argentino? 🇦🇷 De Pergamino a Silicon Valley: la historia de Andrés Macallister y cómo su "Teorema del Cuaderno" está revolucionando la Inteligencia Artificial. 🧠✨

Optimización de datos, ahorro de energía y una startup que ya está en la mira de Google y Meta. ¡Mirá cómo el talento nacional sigue liderando el mundo! 🚀

https://t.co/2FI1khsOVk

#Argentina #IA #Tecnologia #Emprendedores #AndresMacallister #CiudadanoNews #Mendoza #Innovacion

[Andrés Mac Allister @andresmac73]

@Ciudadano_News Gracias por la nota!

[infobae @infobae]

El teorema del cuaderno: el emprendedor argentino que guardó una idea durante quince años y hoy impresiona a Silicon Valley | Por Fernanda Jara https://t.co/DpJy8r6GVv

[Andrés Mac Allister @andresmac73]

@infobae Gracias @infobae por la nota!

[Andrés Mac Allister @andresmac73]

@elonmusk Almost comparable to religions... 🙄

[Andrés Mac Allister @andresmac73]

@simplifyinAI ACE is interesting because it turns learning into incremental state updates instead of weight updates.
But then everything depends on how that state is encoded. If the representation is inefficient, the system keeps patching instead of truly accumulating...

[Andrés Mac Allister @andresmac73]

@ubrmnschn @PI010101 But there’s an interesting separation here, what helps gradients isn’t necessarily what’s needed to store the learned structure. High dimensionality may be useful to find the solution but redundant once you have it.

[Andrés Mac Allister @andresmac73]

I think AGI isn’t just the ability to approach new tasks, it’s the ability to retain structure while doing so. Humans don’t start from scratch every time, they carry forward stable internal representations that make new learning cheap.
Current systems still re-derive too much of their state instead of building on it...

[Andrés Mac Allister @andresmac73]

@FilosofiaStoica La ingeniería no es mas que la naturaleza a través nuestro... 🤷‍♂️

[Andrés Mac Allister @andresmac73]

@patrickc @karpathy We’re getting close to “infrastructure as code for agents.” The next step is “state as infrastructure”, otherwise agents will keep rebuilding what they already did...

[Andrés Mac Allister @andresmac73]

Feels like the hard part isn’t just orchestration but state continuity. Each step (auth, deploy, debug,etc) depends on a shared understanding that currently lives in fragile prompts and logs.

I think that until that state becomes stable and persistent, agents will keep re-deriving context instead of building on it...

[Andrés Mac Allister @andresmac73]

@rryssf The interesting pattern here is separation of concerns.

Rollout vs training is one example. Representation vs compute might be another.

@thesemqgroup is exploring this layer...

[Andrés Mac Allister @andresmac73]

@bingxu_ Blind coding shifts the bottleneck from humans to representation.

Once agents explore 24/7, the question is no longer who searches but how big the space they search is.

Most of that space is probably redundant...

[Andrés Mac Allister @andresmac73]

@porqueTTarg Es mi función favorita la transformada de fourier

[Andrés Mac Allister @andresmac73]

@cryptopunk7213 Definitely one of the most outstanding engineers/scientist/developers of our era…

[Andrés Mac Allister @andresmac73]

@DamianCatanzaro Bellísimo… y a cuánto tiempo estamos de combatir las neurodegenerativas con algoritmos ?