Carlos Chinchilla Corbacho

ML engineer in production; PhD candidate on testing LLM agents. I write Trust, but Trace: essays on agent reliability, evaluation, ML fundamentals, and the EU AI Act. Real numbers, code where useful.

Spain Joined on Apr 22, 2026 https://cchinchilla.dev/

Oct 15

From GPT-2 to Llama: four changes and the reason for each

#machinelearning #python #ai #beginners

9 min read

Want to connect with Carlos Chinchilla Corbacho?

Create an account to connect with Carlos Chinchilla Corbacho. You can also sign in below to proceed if you already have an account.

Create Account

Already have an account? Sign in

Carlos Chinchilla Corbacho

Oct 8

GPT-2 is the transformer plus five decisions, and inference is a different program

#machinelearning #python #ai #beginners

9 min read

Carlos Chinchilla Corbacho

Oct 4

A transformer block is six tensors and a bus: read it like a pipeline

#machinelearning #python #ai #beginners

10 min read

Carlos Chinchilla Corbacho

Sep 25

Attention is a learned weighted average, and the cost is in the square

#machinelearning #python #ai #beginners

10 min read

Carlos Chinchilla Corbacho

Sep 24

Shapes are the type system of ML, and nobody checks them for you

#machinelearning #python #ai #beginners

10 min read

Carlos Chinchilla Corbacho

Sep 22

An embedding is a lookup table, and everything else is how you fill it

#machinelearning #python #ai #beginners

9 min read

Carlos Chinchilla Corbacho

Sep 20

A model doesn't read text: what a tokenizer decides for you

#machinelearning #python #ai #beginners

8 min read

DEV Community

Carlos Chinchilla Corbacho

Badges

1 Week Community Wellness Streak

Writing Debut

From GPT-2 to Llama: four changes and the reason for each

Want to connect with Carlos Chinchilla Corbacho?

GPT-2 is the transformer plus five decisions, and inference is a different program

A transformer block is six tensors and a bus: read it like a pipeline

Attention is a learned weighted average, and the cost is in the square

Shapes are the type system of ML, and nobody checks them for you

An embedding is a lookup table, and everything else is how you fill it

A model doesn't read text: what a tokenizer decides for you