노트, 기계실에서.

Raffael은 낮에는 자동화 우선 플랫폼과 엣지 시스템을 만들고, 틈틈이 언어 도구, 로컬 우선 앱, 시뮬레이션, 게임으로 아이디어를 시험합니다. 이곳에서는 플랫폼 자동화, 엣지 인프라, 응용 보안, 그리고 AI와 함께하는 엔지니어링에 대해 씁니다. 언어와 모델은 도구일 뿐입니다. 사람이 도구를 섬기는 것이 아니라, 도구가 사람을 섬깁니다.

글 읽기지도 둘러보기

최근 글

모든 글 →

이 사이트의 글은 영어로 쓰여 있습니다.

  1. Learning to Play

    A chess prodigy's handbook, three old games and one laptop: rebuilding the ideas behind game-playing AI in Rust for a browser on my phone, and measuring them.

    Game AIMachine learningReinforcement learningGamesRustWebAssemblyHistory of technology

    22 min

  2. noteOpus 5.5 is the best Opus since 4.5

    A month in, Claude Opus 5.5 feels like the next milestone after Opus 4.5: the same jump in what Claude Code can carry, now faster and cheaper per task.

    AI-assisted engineeringInferenceEconomicsAugmentation

  3. reviewThe Web Was the Training Set

    Tim Berners-Lee's memoir is the story of a British engineer at a Franco-Swiss lab giving a protocol away. Read in 2026, it is also the origin story of AI's data supply.

    History of technologyHypertextTraining dataData sovereigntyResearch labsIntellectual propertyEurope

    16 min

  4. Observability Is Not Logging

    A cost total says nothing about how a run went. This part turns each agent run into a trace: which turn escalated, which tool failed, where the money went.

    AgentsAI platformsObservabilityOpenTelemetryTypeScript

    15 min

  5. This Site Is Becoming a Map

    A blog shows what I wrote last. The new map on raskell.io shows what I keep coming back to, and it is the start of a public second brain.

    This siteKnowledge managementHypertextThe open webZolaExternal memory

    9 min

  6. The Bill Is Part of the Loop

    Every model response already carries its own bill in the usage field, and the loop throws it away. This post reads that field, prices it against a table, and turns cost into a value the loop can see per turn and act on: a metering decorator that needs no loop change, a running ledger, prompt-cache accounting that moves the number the most, and a budget that degrades before it stops. Part eight of The Agent Platform Handbook.

    AgentsAI platformsEconomicsObservabilityMetricsTypeScript

    10 min

  7. The Model Is a Dependency

    Every layer of the harness has existed to put the right tokens in front of one hardcoded model string. This post pulls that string out and makes the model a dependency: why you route a cheap model for easy turns and a strong one for hard turns, what a fallback looks like when a provider is down, and why 'models are commodities' is a claim about your architecture before it is a claim about the market. Part seven of The Agent Platform Handbook.

    AgentsAI platformsInferenceReliabilityEconomicsTypeScript

    14 min

  8. The Bottleneck Is Data, and Europe Is Sitting on It

    Frontier AI is running short on public text, not compute. Europe's edge is the high-quality data it never put online. Capturing it needs coordination.

    Training dataOpen dataEuropeIntellectual propertyProvenanceLanguages and translationTechnology races

    34 min

지금 하고 있는 일

프로젝트

모든 프로젝트 →