- 13 min
ERC-8004 listed 418,665 agents. Where were the customers?
- agents
- ethereum
- onchain-data
I checked 457,778 Ethereum and Base reviews. Two files point to settled transactions; only one, worth one cent, names both payment parties.
- 17 min
Does Proof Cost Follow What the AI Actually Runs?
- zkml
- benchmarking
- mixture-of-experts
- proof systems
- verifiable ai
An illustrated guide to testing proof-carrying sparsity: installed experts, activated work, batching, timing boundaries, and a qualified DeepProve comparison.
- 16 min
Prove Only What Runs: Inside a SparseProve Receipt
- proof systems
- mixture-of-experts
- starks
- verifiable ai
A beginner's illustrated walkthrough of how one MoE claim becomes commitments, transcript challenges, authenticated openings, algebraic checks, and a verifier decision.
-
The Multi-Agent Merge Problem, and the Queue Behind It
Coordination moved to Buildkite. A job is submitted as an exact Git tuple — main commit, PR commit, synthetic merge, resulting tree — so a passing result stops matching the moment a branch moves. Buildkite schedules the work; the checked-in contracts still define correctness.
-
Parallel Coding Is Easy. Safely Merging Is the Hard Part.
Four coding agents against one production codebase. Each works in its own Git worktree tied to one exact commit, a bare-metal runner tests the combined tree, and a landing controller refuses any merge whose evidence has gone stale. Multi-agent engineering turns out to look less like hiring developers than like designing a distributed system.
- 17 min
An AI Chooses Which Experts to Run. How Do You Prove It Chose Correctly?
- mixture-of-experts
- sparse computation
- zkml
- starks
- verifiable ai
A beginner's illustrated guide to Mixture-of-Experts, sparse computation, and a proof receipt that binds router scores to selected expert work.
- 12 min
Why Transformer Decode Is a Natural STARK Trace
- proof systems
- verifiable ai
Transformer decode repeats a state transition over carried context. That makes it natural to express as a STARK trace, even though proving the difficult operations efficiently still requires careful design.
-
The Hitchhiker's Guide to Scaling Bitcoin with STARKs
Examining the potential implementation and implications of Scalable Transparent ARguments of Knowledge for Bitcoin's future
-
Train and Fine-Tune Sentence Transformers Models
Preparing a dataset, picking a loss function, and training an embedding model. Written for the pre-v3 fit API, which Hugging Face now marks as superseded.
-
Getting Started With Embeddings
Turning text into embeddings with Sentence Transformers, hosting them on the Hub, and building semantic search over a set of FAQs.
-
Announcing the Hugging Face Fellowship Program
Compute, recognition and support for people carrying open-source machine learning. Written with merve.
-
-
Procesando Datos Para Deep Learning (II)
Datasets, visualizaciones y DataLoaders en PyTorch
-
Procesando Datos Para Deep Learning (I)
Datasets, visualizaciones y DataLoaders en PyTorch
-
-
Entendiendo PyTorch (II)
Las bases de las bases para hacer inteligencia artificial. Parte 2
-
Entendiendo PyTorch (I)
Las bases de las bases para hacer inteligencia artificial. Parte 1
-
Hugging Face 101 en español
Exercises for learning NLP with the Hugging Face libraries, in Spanish.
-
Manipulación de tensores en PyTorch
Tensors as the working unit of deep learning: building them on the GPU, broadcasting, reshaping, and flattening images.
-
Entendiendo PyTorch: las bases de las bases
How PyTorch works underneath, written for people starting out in artificial intelligence.
Writing published elsewhere is listed above and links to its source. Also on Substack , HackerNoon , Medium , Hugging Face and GitHub .