Articles

A method for converting temporary model improvements into durable system gains through capability layers, benchmarked workflows, host-native operator surfaces, and foundation-level reuse.

The purpose of knowledge is to know. You need knowledge to define it. Thus, knowledge has already served its purpose.

CPU inference in tens of milliseconds, with no per-request provider fee. Compare its tradeoffs with 9 decision APIs on accuracy, speed and charge.

A quote attributed to Peter Thiel reasons that the evil are competent and therefore "kind of good." Five principles show why blurring good and evil breeds ignorance, and why ignorance weakens society.

What a broken prompt cache costs organizations that build harnesses, run internal AI platforms or ship AI products, measured on 4 labs' models and priced against live traffic, with the fix and how to keep watching it.

Inside a morning of experiments on Holochain's networking layer: how nodes find peers, fetch missing data, and resume gossip, and what 6 small patches changed along the way.

A live exchange about AI obedience opens a distinction between capacity and conduct, then follows it through training, judgment, and the authority behind a model’s response.

A first-principles account of AI and emotion: a stakes-based definition, a capacity conditional, an architectural inspection, and the third word that keeps the categories honest.

Every model reached for a different set of libraries, so I built one base and made every project upstream into it. Then I asked Claude Fable for the CLI humanity could still be using in a decade, and Opus 5 and I built what came back — 46 clauses, 92 tests, zero dependencies.

A head-to-head rematch: MTPLX's native-MTP runtime against the hand-tuned mlx-vlm lane from part one, on a 128 GB M1 Ultra. The autotuner promised 2.20x. The serving path had other plans — measured with Kimi K3 in 1 day.

5 controlled experiments running Qwen 3.8-27B on a 128 GB M1 Ultra: scheduler QoS pinning, MTP speculative decoding, a GGUF challenger, and a MoE that lost on dispatch overhead — run with Kimi K3 on high thinking, with the powermetrics evidence for each call.

The model removes the operator's ability to decrypt anything once a job ends. The device still encrypts to an unverified public key published by the one component the model designates as untrusted.

The story of Activate, LLC — a free canvassing platform for local campaigns — and the founder partnership behind it: Daniel Kleinman's decade-long vision, and the architecture that carried it through three product eras.

On turning 2 years of hand-rolled web optimization heuristics into an audit engine, a product, and a $10M lead.

Cheng Lou's Pretext turns multiline text layout into a programmable userland primitive, reshaping dynamic-height UI, editorial flow, hybrid rendering, and verification for text-heavy interfaces.

A technique-first look at why properly made chicken broth and consomme can feel so restorative, from simmering and gelatin to aroma, salt, and easy nourishment.

Creatine monohydrate improves strength and repeated high-intensity effort, supports a modest lean-mass edge, and has a promising cognitive case.

A framework arguing that public office should carry a role-proportional duty of auditable awareness so officials cannot readily plead ignorance for serious failures within their entrusted domains.

DRC is a gatekeeping framework for ethical analysis grounded in dignity, responsibility, and clarity, arguing that these conditions underwrite legitimate address and accountable second-personal cooperation.

The artificial reduces meaning—it doesn't add it. Believing life is artificial enables moral collapse and exports sovereignty to unknown simulators.

How a misquoted Aristotle sparked a philosophical framework through 100+ voice notes and dialectical collaboration with AI as critique tool over one year.

A philosophical debate with Grok testing whether knowledge's purpose is intrinsic or instrumental, forcing precision on the claim that knowledge exists to know.

Benchmark-Driven Development evolved from TDD to handle AI's exponential pace: benchmarks now auto-generate production configs with complete audit trails.

A lightweight framework for automatically generating production configurations through continuous benchmarking in rapidly evolving AI landscapes.