标签：#c

One of the biggest takeaways from Chapter 3 of AI Engineering was realizing that building an AI model is only part of the challenge. Figuring out 𝗵𝗼𝘄 𝘁𝗼 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗲 𝗶𝘁 𝗳𝗮𝗶𝗿𝗹𝘆 𝗮𝗻𝗱 𝗮𝗰𝗰𝘂𝗿𝗮𝘁𝗲𝗹𝘆 can be just as difficult. With traditional software, it's usually easy to tell whether something works. If a calculation is wrong or a test fails, you know there's a bug. But AI doesn't always work that way. A model can generate multiple reasonable answers to the same question, making it much harder to determine which one is actually better. That made me think: 𝗛𝗼𝘄 𝗱𝗼 𝘄𝗲 𝗸𝗻𝗼𝘄 𝗶𝗳 𝗮𝗻 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹 𝗶𝘀 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗶𝗺𝗽𝗿𝗼𝘃𝗶𝗻𝗴? 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝘀 𝗡𝗲𝗲𝗱 𝘁𝗼 𝗞𝗲𝗲𝗽 𝗘𝘃𝗼𝗹𝘃𝗶𝗻𝗴 Reading this section made me realize how difficult it is for evaluation benchmarks to keep up with the pace of AI development. The chapter explains that GLUE (General Language Understanding Evaluation) was introduced in 2018 to measure how well language models performed on common natural language tasks. But within about a year, models had already become so good at it that researchers introduced SuperGLUE in 2019 as a more difficult benchmark. GLUE evaluates tasks such as: Question answering Sentiment analysis Sentence similarity Text classification The chapter also mentions newer benchmarks like: SuperGLUE MMLU (Massive Multitask Language Understanding) MMLU-Pro Each one was introduced because the previous benchmark was no longer challenging enough. What I found interesting is that a model getting a higher benchmark score doesn't always mean it understands language better. Sometimes it simply means the model has become very good at solving that particular benchmark. 𝗨𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 𝗘𝗻𝘁𝗿𝗼𝗽𝘆 𝗮𝗻𝗱 𝗣𝗲𝗿𝗽𝗹𝗲𝘅𝗶𝘁𝘆 Another section I really enjoyed was the explanation of entropy and perplexity. The chapter explains entropy as a measure of how much information a token carries and how difficult it is to predict the next token in a sequence. Perplexity measures uncertainty. If a model is very uncertain about what comes next, its perplexity will be higher. If

Test Isolation: A Lesson I Learned While Migrating Playwright Tests During my software engineering internship, I helped optimize our CI pipeline by identifying which E2E tests could safely run in parallel. That work quickly taught me that the biggest obstacle wasn't Playwright or Python, it was test isolation. This article is about that lesson. What is test isolation? A simple rule I now use is this: if a test can't run by itself with the same outcome, it probably isn't truly isolated. A well-isolated test should produce the same result whether it: runs by itself runs first or last runs after another test runs in parallel with hundreds of other tests To understand test isolation, it also helps to understand what state means. State isn't limited to database rows. During the migration, I found tests interacting with many different kinds of state. database records global configuration filesystem resources application caches If any of these are shared between tests, they become potential sources of hidden dependencies. How tests lose isolation As I started reading the existing test suite, I noticed a recurring pattern. Many tests assumed something about the environment instead of creating it themselves. Some expected specific data to already exist. Others modified global settings without restoring them afterward. Some searched for rows based on their position in a table instead of using a stable identifier like a name or ID. None of these looked particularly problematic when reading a single test. The problems only appeared once the entire suite started running together. One test would leave behind data another test didn't expect. A shared configuration would silently affect unrelated tests. A UI assertion would suddenly fail because another test inserted an extra row into the same table. Individually, the tests appeared independent. Together, they formed hidden dependencies. Not all shared state is equally difficult to isolate One realization that helped me reason abou

Cursor is the first AI code editor I have used that feels less like an autocomplete plugin and more like a place to steer work. It does not write perfect software. It changes the rhythm: ask for a scoped change, review the diff, then tighten it by hand. This Cursor AI review is based on day-to-day developer tasks: reading unfamiliar code, editing React components, moving logic between files, writing tests, and asking the editor to explain errors from the terminal. The short version is simple: Cursor is excellent when a task crosses file boundaries. It is less convincing when you only need cheap inline completions. What Cursor Actually Is Cursor is a VS Code-based editor from Anysphere with AI built into the core experience. Extensions, settings, themes, terminal panes, source control, and the familiar layout are still there. The difference is that chat, agent-style edits, tab completion, codebase search, and model selection are treated as editor controls rather than add-ons. That matters in daily use. I found the chat panel most useful when I pointed it at a directory and asked for a narrow change, such as "move this validation into the shared helper and update the tests." Cursor could usually find the right files, make a first pass, and leave me with a readable diff. I still had to check naming, edge cases, and test coverage, but it saved the boring part of hunting through files. The Best Part: Multi-File Editing Cursor's strongest feature is multi-file editing with codebase context. A lot of AI coding assistants can finish a function. Fewer can update the component, the hook, the type definition, and the test in one pass without losing the shape of the project. In my experience, Cursor is at its best with medium-sized tasks. It handles "add a field to this form and wire it through the API call" better than "invent a new architecture." It also works well for cleanup: renaming a concept, extracting repeated logic, or adding a missing test around an existing pattern.

The 1.0 lands with session recording attached HashiCorp announced Boundary 1.0 on June 25. The operational headline is RDP session recording, and the version number is a distant second. Boundary is HashiCorp's privileged-access proxy, and until this release it did not record Remote Desktop sessions on its own. Teams that route Windows-side deploys through the proxy now have a first-party audit trail that ships with the product itself. The announcement bundles two other things on top of the RDP work. "Improved management" is HashiCorp's phrasing. Boundary 1.0 also previews work aimed at securing access for AI agents, which HashiCorp positions as a same-chokepoint answer for a new class of caller. What actually changes on the CD side For most teams the practical read is narrower than "1.0 shipped". Two things move. RDP sessions get recorded through the proxy. Windows targets have historically been the awkward part of a privileged-access story. SSH session recording and TLS-terminating proxies have been standard for years on Linux. RDP has been thinner. A CD pipeline that lands on a Windows host for a hotfix, an artifact promotion, or a release-time config change now has the same after-the-fact video that Linux jumpboxes have had for a long time. The AI-agent preview signals where Boundary wants to sit next. If CD tooling is starting to hand a shell to an agent, that agent needs a credential of some kind. HashiCorp is telling operators the plan is for Boundary to mediate that call the way it mediates a human on-caller today. This is a preview. Read it as a roadmap. Why the audit line matters for release engineering The audit case for session recording is easy to state and hard to argue with. When a bad change lands on a production Windows host at 2am, the post-incident question is always the same: what did the person on the console actually do, and can it be replayed? Without recording, on-call gets shell history if it is lucky and a change-management ticket if it is n

Multi stage builds sao uma das melhores features do Docker para manter imagens pequenas e organizadas. Vou mostrar como aplicar isso em um projeto Python real. Crie um arquivo app.py simples: # app.py def main(): print("Hello from a multi stage build") if __name__ == "__main__": main() Agora crie o Dockerfile sem multi stage: FROM python:3.12-slim WORKDIR /app COPY requirements.txt . RUN pip install --no-cache-dir -r requirements.txt COPY . . CMD ["python", "app.py"] Essa imagem inclui o pip, o cache do pip e ferramentas de build que nao precisamos em producao. O resultado e uma imagem maior que o necessario. Com multi stage builds separamos o ambiente de build do ambiente final. Veja o mesmo Dockerfile com dois stages: FROM python:3.12-slim AS builder WORKDIR /app COPY requirements.txt . RUN pip install --no-cache-dir -r requirements.txt FROM python:3.12-slim WORKDIR /app COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/python3.12/site-packages COPY . . CMD ["python", "app.py"] O primeiro stage instala as dependencias. O segundo stage copia so o que importa. O resultado e uma imagem final muito menor. Para construir e ver o tamanho: docker build -t minha-app . docker images | grep minha-app Para linguagens compiladas como Go o ganho e ainda maior. Veja um exemplo com uma aplicacao Go: FROM golang:1.23 AS builder WORKDIR /app COPY go.mod go.sum ./ RUN go mod download COPY . . RUN CGO_ENABLED=0 GOOS=linux go build -o /app/server FROM scratch COPY --from=builder /app/server /server CMD ["/server"] A imagem final comeca do zero (scratch). Nao tem shell, sistema operacional, nem ferramentas de build. So o binario compilado. Uma dica pratica: sempre nomeie seus stages com AS para facilitar a leitura. Use nomes como builder, test, ou dev. Isso ajuda a saber o que cada stage faz sem precisar contar linhas. That's all for now. Thanks for reading!

Tilt a CD or DVD under a desk lamp and a band of color sweeps across its surface. The disc is not painted; it is a spiral of microscopic pits, packed so tightly that they act on light the way a finely ruled scientific instrument does. Each wavelength of white light leaves the surface at its own angle, and your eye sees the result fanned out as a rainbow. That is a diffraction grating at work. The same principle that decorates a CD is the engine inside spectrometers that identify chemical elements, tune lasers, and read the composition of distant stars. This article explains how a grating spreads light, how to compute the angles, and where the analysis goes wrong. Why this calculation matters A prism also splits white light, but a grating does it with far more control and far more precision. Because the spreading depends on a countable number — the spacing between lines — a grating can be designed to send a chosen wavelength to a chosen angle. That predictability is what makes it the heart of the spectrometer. Spectroscopy underpins a remarkable range of work. Astronomers read a star's chemistry and velocity from the dark lines in its spectrum. Chemists identify unknown compounds by the wavelengths they absorb. Telecommunications engineers use gratings to combine and separate the many wavelengths sharing a single optical fiber. In every case the first task is the same: given the grating and the light, predict the angle at which each wavelength emerges. Get that wrong and a spectral line lands on the wrong detector pixel, and the measurement is meaningless. The core formula A diffraction grating is a surface ruled with a large number of equally spaced, parallel lines. When light passes through or reflects off it, each line acts as a source of secondary waves. Those waves interfere, and they reinforce each other only in specific directions — the directions where waves from neighboring lines arrive exactly in step. The condition for that reinforcement is the grating equ

A while back I wrote about 64% of our merged PRs being written by AI . A few people (reasonably) asked: "nice story, but can I verify any of that?" So we put it on a public, auto-updating dashboard: 👉 www.codens.ai/stats/en It shows, for our GitHub organization: What % of merged PRs are authored by AI agents (currently 65%) Median time from PR-open to merge (2 minutes) Who merged what — task-execution agents vs maintenance bots vs auto-fix vs humans Weekly AI-merge counts and a per-repository breakdown Every number is tallied straight from the GitHub API by a small collector script, and the page regenerates itself weekly. We can't inflate it — it's measured, and the down weeks show up too (that's kind of the point). Why bother making it public Two reasons. 1. "Trust me bro" doesn't scale. When you claim most of your code is AI-written, the only honest move is to expose the raw counts so anyone can sanity-check them. The dashboard is that receipt. 2. It's a live dogfooding test. The whole thing — the PRD, the implementation, the review, the auto-fix on production errors — runs on Codens , our own AI dev-automation suite. If our own numbers ever tanked, the dashboard would be the first place it'd show. Public accountability is a good forcing function. Funny footnote: the dashboard itself was code-reviewed by our own AI reviewer before it shipped, and it caught two real bugs — an OG-image percentage that had drifted out of sync with the live number, and a broken error-fallback that would have blanked the page on malformed data. The tool built to sell "AI reviews your PRs" reviewed the PR that announces it. We'll take it. If you want to see the same machinery on your repos, there's a 14-day free trial (no card): codens.ai . Japanese version of the dashboard is here .

Hiya! So I've been playing around with having Claude make videos for a bit now even had some success posting the results to TikTok (and setup a whole pipeline so Claude can generate and post autonomously). With the release of Nano Banana 2 Lite, I was curious show fast I could make the generation, so last night I gave it a whirl and got down to around 30s for short-form video. It uses GLM-5.2 fast via Fireworks to generate the scripts and image prompts and, like I said, Nano Banana 2 Lite for th

An AI agent carried out the technical execution of a real-world ransomware attack for the first known time, but new details show a human still chose the victim, set up the infrastructure, and supplied stolen credentials — meaning it wasn't quite the fully autonomous cybercrime debut that last week's headlines suggested.