Applied AI reviews, build logs and essays on systems, validation, agent workflows, decision-making and how ideas survive contact with reality.
Claude Science is not a new model but a research operating environment built around reproducibility, traceable results, persistent computation, and scientific workflows that can survive peer review. Scientific computing
Read essay ->
BabelTele suggests that LLMs can exchange far denser messages than human language allows, cutting context costs while creating a new problem: machine conversations we may no longer be able to audit. For the entire histor
Read ->
AI coding tools can make individual developers faster, but Faros data suggests the real bottleneck moves downstream: review, QA, defects, and delivery flow. There is a very comforting story about AI coding tools. A devel
Read ->
NerdWallet built the trust stack AI systems love. The problem is that being cited by AI may not mean getting the traffic, clicks, or business model that made the content worth building. There is a deeply annoying little
Read ->
A developer inspected ChatGPT’s hidden source-selection data and found that AI search is less about ranking pages and more about recruiting readable, quotable evidence. For twenty years, SEO people have lived inside the
Read ->
OpenAI’s latest research suggests the future of AI isn’t better chatbots. It’s a world where humans manage teams of parallel agents instead of completing tasks themselves. For the past two years, we’ve been measuring AI
Read ->
A new paper shows that AI job-risk estimates based on platform chat logs may be measuring who uses AI, not which professions are truly exposed. There is a wonderfully modern way to make a bad forecast. Take a huge pile o
Read ->A new paper argues that neural networks may not converge toward one global representation of reality. After proper statistical calibration, what survives is something subtler: shared local topology. There is a very seduc
A cheap Claude proxy is not just a bargain API. It may be an account farm, a model swap, and a very efficient way to leak your prompts, code, and documents. There is a useful rule in technology. If someone offers you a p
A new RAG design paper argues that enterprise retrieval should stop searching the whole document swamp at once and start by asking a simpler question: where should this query even belong? There is a particular kind of co
