nuxyel
renan vinícius. I write software, and now I write about it too.
I build AI and machine learning projects and write about what I learn.
start here
The Plan
The path: why I study AI, and what it costs in hours.
6 min · EN·PTa pytorch where a shape error doesn't compile
The project: building machine learning with shapes checked by types.
48 min · EN·PTartificial intelligence didn't begin with ChatGPT
Explore the ideas behind AI with interactive demonstrations.
132 min · EN·PT$ ls ~/projects
bend-ml
Machine learning in Bend. Tensor shapes live in the types; structural guarantees are checked by the kernel.
5 published packages · MNIST training · GPT-2 generation
explore project →llama.bend
Llama-architecture models running in Bend, built on bend-ml. Typed weights, a checked cache, and generation compared with PyTorch.
SmolLM2 135M / 360M / 1.7B · one published package
explore project →$ ls -t ~/posts | head -4
por que acredito que Bend pode tomar o lugar do Python em machine learning
depois de construir matemática, inferência e treinamento em Bend, minha aposta ganhou exemplos concretos: tipos, provas, paralelismo e uma distância de desempenho que ainda precisa diminuir
llama.bend: language models with typed weights and cache
how I ran three SmolLM2 models in Bend: attention, RoPE, typed weights, a causal cache, an RMSNorm error and the public results, including the limits
artificial intelligence didn't begin with ChatGPT
from logic to neural networks, from winters to agents: a history of AI told from the beginning, with people, examples and visual experiments, requiring no programming or mathematics
you're not an AI expert (and that's fine)
using a tool doesn't give you authority over an entire field. how to assess AI promises, recognize competence, and build depth without selling what you don't yet know