Featured Dispatch
Discover how CauseScope traces a React UI symptom back to exact TSX, runtime conditions, state updates, and network responses during local Vite development.
A practical knowledge map from core concepts to model architecture.
10 posts published
Tool use, browser agents and the engineering boundaries around them.
3 posts published
Reusable methods from problem framing through delivery review.
2 posts published
Index
Continue through engineering practice, product judgement and tooling notes in chronological order.
Discover how CauseScope traces a React UI symptom back to exact TSX, runtime conditions, state updates, and network responses during local Vite development.
react / vite / debugging / devtools / open-source
A practical account of fixed vocabularies, subword tokenization, special tokens, and why tokens are the interface between language and a neural network.
llm / transformer / token / foundation
Treat tokens as a system budget spanning prefill, autoregressive decoding, context capacity, API cost, and the information density of a prompt.
llm / transformer / token / foundation
A concrete walk through token IDs, embedding lookup, positional information, attention masks, hidden states, and the projection back to vocabulary logits.
llm / transformer / token / foundation
How embeddings place discrete tokens in a trainable continuous space, and why vector proximity, lexical meaning, and contextual understanding are not interchangeable.
llm / transformer / token / foundation
Why parallel self-attention needs an explicit signal for order, direction, and distance, from sinusoidal encodings to relative position and RoPE.
llm / transformer / position / foundation
A step-by-step explanation of attention scores, masks, weighted value retrieval, computational cost, and why attention maps are not complete explanations.
llm / transformer / attention / foundation
A retrieval-oriented explanation of Query, Key, and Value projections, score scaling, masks, and why the three vectors do not carry fixed semantic roles.
llm / transformer / attention / foundation
How multiple attention heads learn parallel routing patterns in different representation subspaces, how their outputs are combined, and where the design trades off against inference cost.
llm / transformer / attention / foundation
A reconstruction of the full Transformer block: attention communicates across positions, the FFN transforms each position, and residuals plus normalization make depth trainable.
llm / transformer / architecture / foundation
How the original encoder-decoder Transformer became encoder-only BERT and decoder-only GPT families, and why attention visibility plus training objective shape their interfaces.
llm / transformer / bert / gpt
How OpenClaw discovers, prioritizes, filters, and exposes skills to a model, and why a Markdown instruction layer must remain separate from tools, permissions, and host security.
openclaw / ai-agent / tool-use / prompt-engineering
Alibaba Page Agent uses in-page JavaScript and a text-oriented DOM to create a natural-language control layer. The same architecture that improves product integration also concentrates permission, visual, and reliability concerns in the page.
ai-agent / web-agent / browser / alibaba
How PageIndex uses document trees and reasoning-based retrieval for long structured documents, where the approach can help, and which costs its vectorless framing tends to hide.
rag / llm / retrieval / ai-engineering
Superpowers constrains a coding agent with design, planning, tests, review, and fresh verification. The useful question is where that discipline pays for itself and where it becomes ceremony.
ai-agent / engineering / workflow / tdd