Attention Is All You Need: the seed of modern generative AI
The 2017 paper that changed the paradigm: from reading word by word to having every word consult every other word at once.
// espacio-latente.com
Ideas in a latent state,
I build and document real systems with agents, local LLMs, and automation.
The rack
4 ideas decodedEach module carries a serial number (U-01, U-02…): it marks the order it was built in, not a level or a reading order.
Inside the LLM
2 of 4 published
The inner mechanics: attention, representation, compression, and the smallest units of text.
Tokenization and embeddings: the model's front door
Third module of the 'Inside LLMs' series: how text gets chopped up before the model sees any of it, why writing in Spanish costs you 45% more, and what actually happens in the first layer.
¿Qué es el espacio latente? La geometría del significado
Not translated into English yet — you can read it in Spanish. Read in Spanish
Autoencoders: compressing to understand
How a model learns to summarize without losing what matters.
Tokenization: from text to numbers
The first step, and why the model doesn't see words but fragments.
Agents, step by step
2 of 2 published
With the foundations already in the rack: how a model goes from predicting text to acting.
Agent-bio
An agent that speaks for me: who I am, what I do, and why.
Harness, loop and graph engineering: three ways to build an agent
Three vocabularies used as if they competed, which don't: one is about who decides the next step, another about how the path is drawn, and the third about the ground you stand on. With the evidence that the ground matters more than it looks.
Build log
0 of 1 published
This site is also a case study: what gets built, with which pieces, why, and what gets dropped along the way.
Cómo se construye este sitio: el porqué detrás de cada pieza
Not translated into English yet — you can read it in Spanish. Read in Spanish
El radar aprende a recordar: memoria semántica y una suposición que resultó falsa
Not translated into English yet — you can read it in Spanish. Read in Spanish
La granja de agentes: un orquestador que reparte tareas entre agentes de código
Not translated into English yet — you can read it in Spanish. Read in Spanish
El radar deja de pagar dos veces por lo mismo
Not translated into English yet — you can read it in Spanish. Read in Spanish
El bake-off del modelo local: 5 candidatos, 3 fallos silenciosos, 1 ganador
Not translated into English yet — you can read it in Spanish. Read in Spanish
El bake-off, round 2: Devstral defiende el título contra Qwen3.8
Not translated into English yet — you can read it in Spanish. Read in Spanish
Construyendo CarteraSonar
Not translated into English yet — you can read it in Spanish. Read in Spanish
Offloadear expertos MoE en llama.cpp: dónde está el muro y por qué más bits no compró nada
Not translated into English yet — you can read it in Spanish. Read in Spanish
El bake-off, round 3: Ornith empata con Devstral y no lo destrona
Not translated into English yet — you can read it in Spanish. Read in Spanish
obsidian-ideas: de un enlace en Telegram a un PR abierto
Not translated into English yet — you can read it in Spanish. Read in Spanish
El radar aprende a decir por qué importa
Not translated into English yet — you can read it in Spanish. Read in Spanish
El radar aprende a desconfiar de sí mismo: dos jueces en vez de uno
Not translated into English yet — you can read it in Spanish. Read in Spanish
Semantic dedup: keeping the radar from repeating itself
Embeddings and vector search so the same story never gets published twice.
The machine room
1 project runningThese aren't articles, they're machines left running: things I've built that are still working (or waiting for me to pick them back up).
ai-trading-lab
activeMulti-agent trading system: analyses 21 tickers daily, reasons over technical signals and sentiment using local LLMs, and executes real orders on an IBKR paper account.
LangGraphOrnith-1.5-35B (local)FinBERTChromaDBDuckDBStreamlit
Suggestion box
what module should I install next?Suggest an experiment, ask about agents, or request something you’d like to see documented here.
Shape it with AI (optional)
Tell me what you'd like to see and I'll help narrow it down — a few messages, free.