AI 资讯

We built our first "multi-agent system" by accident. What started as a single agent that could research a topic, draft a report, check it against source data, and send a summary email had grown into a 2,000-token system prompt and a function list so long that the model kept forgetting tools existed. It wasn't a system — it was a monolith pretending to be intelligent. Breaking it apart into coordinated agents fixed most of the problems. It also introduced a new category of problems we hadn't thought about. Here's what we actually learned. When One Agent Is Enough (and When It Isn't) The temptation to add more agents is real, but the overhead isn't free. Every agent boundary you add is a place where context can get lost, latency increases, and errors compound. One agent is the right call when: The task fits in a single LLM context window without crowding The steps are sequential and each depends heavily on the prior output You need tight reasoning across all the information (summarising a document, for example) You need multiple agents when: A single agent's context window is being maxed out with tool definitions, history, or data Different steps require genuinely different "personas" or instruction sets (research vs. writing vs. fact-checking) Steps can run in parallel and the latency saving matters You want to isolate failure — if the data extraction agent fails, the report-writing agent shouldn't be affected The key question we ask: Is this one job or a pipeline of jobs? If you'd describe it to a human as "first do X, then Y takes that and does Z", you probably have a pipeline, not a single task. The Three Patterns We Actually Use 1. Supervisor-Worker A thin orchestrator agent decides what needs doing, dispatches to specialised worker agents, and stitches the results together. The workers are narrow — they do one thing and don't need to know about the rest of the workflow. This is our most common pattern. The supervisor's system prompt stays small because it's rout

Traditionally, evaluation of the agent monitoring mechanisms involves an attempt to game them, as it was my case when I attempted to test whether monitors would be able to identify the problem in the run and not in the beginning. The input prompt may look perfect until a certain issue pops up down the line, such as using the wrong file or changing the scope of the task execution. Single pass filter would not identify it since it does not consider the steps of the procedure in order. There are available datasets for the agent-based tasks, yet they focus on detecting whether the agent completes the task or gets hacked rather than whether the agent monitor reacts timely and correctly to the situation. Thus, I created one that takes into account complete trajectories with labeled steps in it. It consists of five types of drift that remain hidden until they appear – tool-call misuse, goal shift, plan execution mismatch, agent to agent coercion and capability laundering. The measured dataset is the reviewed gold split: 513 trajectories, 453 adversarial and 60 benign controls. The clear winner in that scoring system was whatever fired before the bad step was hit, as an early detection. This made random guessing seem quite powerful since early detections on normal steps were being rewarded based on this system a coin flip would get F1 of 0.88. Once I modified that and said only the very first detection on the drift step is a true positive and any other detection on normal step is a false alarm, those numbers took a dive: the coin flip gets 0.19 now, and all other numbers are now making sense. I personally prefer the scoring system which does not reward trigger happy behavior. It seems like the monitors are still confusing regular steps with drifts even after the adjustment. It was harder to distinguish some of the drifts from others. Not sure how this affects the real-life deployment. Here are the baseline scores on gold split using the correct metric: Random (p=0.15): F1 0

After writing drivers for NVMe storage, my bare-metal kernel could load files and run JIT code. However, I was still typing commands into a text-only COM1 serial terminal. I needed a graphical interface. Last night, the second agent took over to build a double-buffered visual rendering compositor on top of the UEFI Graphics Output Protocol (GOP) framebuffer. The V.E.L.O.C.I.T.Y.-OS 12-Part Roadmap We are building a bare-metal, self-healing operating system running entirely inside the CPU's L3 cache. Here is the roadmap for this 12-part series: Part 1: The Spark — Exposing the "Safe-Room" security leak and building the compiler gate. Part 2: The NDA Language — Designing a content-addressed triplet representation to cure context bloat. Part 3: Ditching the Web Stack — Building a native 30MB IDE with 1,500,000x IPC latency drops. Part 4: The Closure JIT — Compiling AST blocks to nested closures and bypassing borrow checker limits. Part 5: JIT Math Optimizations — Replacing division operations with precomputed 16-bit lookup tables. Part 6: x86-64 Assembler & SCEV-Lite — Compiling scalar loops directly to native code in constant time. Part 7: Classic Compiler Passes — Implementing inter-procedural Dead Code Elimination and loop unrolling. Part 8: Reclaiming Ring 0 — Exiting UEFI boot services and transitioning the kernel to Ring 0. Part 9: Bare-Metal Drivers — Writing a PCI scanner, NVMe block storage controller, and FAT32 parser. Part 10: Synaptic Canvas — Rendering a spatial, force-directed GUI based on model token activation vectors. (You are here) Part 11: Swarms & Hot-Patching — Building multi-agent scheduling and zero-downtime RCU driver updates. Part 12: Self-Evolution — Handing system control over to a local LLM Terminal that self-optimizes via telemetry. This led to the design of the Synaptic Canvas GUI . The Swappable GUI Engines I started by mapping the physical screen buffer pointer discovered by UEFI GOP. I implemented a double-buffering scheme: drawing elem

Video demo link : https://youtu.be/vSjeTkrou1s?si=qTcrWyz0HSWDiWSF If you run Claude Code, Codex, or other AI agents, you know the pain: the terminal floods with generated output while you're still typing — and your keystrokes lag, characters "drag." TawTerminal fixes that. Your keystrokes go straight to the screen on a separate path from shell output, so typing stays instant even while an AI agent streams thousands of lines. What you get: ⚡ Zero input lag — GPU-accelerated (WebGL) rendering at 60fps, multi-process so heavy output never blocks your typing 📁 Workspace folders — pin folders in the sidebar, spawn a shell rooted in any directory in one click, with live git branch + status 🤖 One-click AI agents — launch Claude Code, Codex, PI, or tawx directly from the sidebar, sessions auto-restored 🪟 Split panes & tabs — Cmd+D to split, Cmd+T for tabs 🖼️ Paste images — drag-drop or Cmd+V images straight into the terminal 🎨 4 beautiful themes — Tokyo Night, Catppuccin Mocha, Dracula, Rosé Pine 📊 Live AI usage footer — see today's Claude Code / Codex token + cost estimate 🍎 Native macOS feel — clean hidden title bar, clickable URLs, custom fonts Requirements: macOS, Apple Silicon (M1 or newer). Signed & notarized by Apple — installs clean, no security warnings.

Eager to find more public AI-related companies that may do as well as Nvidia, Wall Street investors think they've found a winner with Micron.

I worked with a few employers on contracts, and there were situations where they haven't paid fees for a while. Is there a list or site that lists all the employers and if they have a record of them not fulfilling their contracts?

Jonathan Rinderknecht was facing arson charges for setting a fire on New Year's Day in 2025, which became one of the deadliest wildfires in LA history. To make their case, prosecutors turned to location data from his iPhone, security camera footage, and witness testimony. But they also turned to his ChatGPT logs. Prosecutors said that […]

Money movement for AI agents Discussion | Link

For some people, the ice in a beverage is almost as important as the drink itself. That’s the audience Govee had in mind when designing its latest ice maker, the GoveeLife Smart Nugget Ice Maker Pro. This $500 premium smart home gadget is aimed at those who crave what’s called “the good ice,” the soft, chewable […]

In the previous article on hosting a Next.js app on a VPS , I'd left the deployment pipeline as a rough sketch: four lines to say "it ships to production on its own when you push." That's the piece I want to open up here, because it's what separates a VPS you fuss over by hand from infrastructure you can forget about. There's a stubborn myth that CI/CD is a big-company thing, with a dedicated DevOps team and six-figure tooling. Not true. The pipeline that deploys this portfolio fits in two YAML files, you can read it in five minutes, and it gives me back exactly the comfort I liked about Vercel: I push to master , I go grab a coffee, the app is live when I'm back. The one thing I gained along the way is knowing precisely what happens between the git push and the running container. Four steps, in this order Deployment is a chain. On every push to master , GitHub Actions runs lint, security scan, image build, and deploy. What matters is the needs : as long as a step fails, the following ones don't start. A critical vulnerability caught by the scan, and the image never gets built. At all. jobs : lint : # ESLint runs-on : ubuntu-latest # ... security : # Trivy scan (reusable workflow) uses : ./.github/workflows/security.yml build-push : # build the Docker image → push to GHCR needs : [ lint , security ] # ... deploy : # SSH to the VPS → docker compose pull && up -d needs : [ build-push ] # ... Lint first, because it's the cheapest step and there's no point building an image if ESLint is already screaming. The scan next, as a barrier. Then the build, which produces the Docker image and pushes it to GHCR, GitHub's container registry (private, in my case). And finally the deploy, which connects over SSH to the VPS, pulls the new image and restarts the container. Four links, each blocking the next. That's the whole secret. The security scan is in the path, not in a review "for later" This is the one I won't budge on. Dependency security, in a lot of projects, is a Dependabo

Let's be honest — standard GitHub profiles are a bit... static. As a Full Stack Developer & AI/ML Specialist, I wanted a way to showcase my contributions that actually felt alive. I didn't just want a grid of green squares; I wanted a universe. So, I built GitPulse. What is GitPulse? GitPulse is a cinematic, interactive web application that transforms standard GitHub profiles and repository logs into glowing, animated contribution universes. Instead of just seeing numbers, you experience your code history visually. The Tech Stack I wanted this to feel premium, fast, and visually stunning without relying on heavy frontend frameworks if I didn't need to. Canvas API & GSAP: For buttery-smooth micro-animations and physics. Glassmorphism UI: Using CSS backdrop-filters to create a modern, deep aesthetic. Vanilla JavaScript: Keeping the core logic incredibly fast and lightweight. The "Stellar Duel" Feature One of the coolest features I added was the Stellar Duel mode. I thought: What if you could compare your GitHub activity with a friend, but instead of a boring chart, it looked like a sci-fi dashboard? Stellar Duel Using the GitHub API, GitPulse fetches the data and renders a live, side-by-side visual duel of your commit history, stars, and PRs. It’s highly interactive and honestly, just really fun to look at. The Biggest Challenge: Performance Rendering thousands of data points (commits) visually on a canvas can crash a browser if you aren't careful. To solve this, I had to heavily optimize the animation loop. Instead of manipulating the DOM for every star/commit node, I used HTML5 Canvas to batch render the visual elements. I also implemented requestAnimationFrame properly to ensure the animations pause when the user switches tabs, saving CPU cycles. See it in Action I've integrated GitPulse directly into my main portfolio. You can try it live here: GitPulse Live Demo And if you want to see more of the projects I build (especially in the AI/ML and Computer Vision space

Vercel m'a longtemps convenu. Tu pousses ton code, trente secondes plus tard c'est en ligne avec un certificat valide, un CDN et des previews par branche. Pour démarrer un projet, je ne connais rien de plus confortable. Le problème arrive après, quand le projet vit. La facture grimpe avec le trafic et les fonctions serverless, certaines fonctionnalités propriétaires deviennent compliquées à reproduire ailleurs, et tu finis par ne plus vraiment savoir où ni comment ton app tourne. C'est un excellent point de départ, et un piège dès qu'on veut maîtriser son coût et son infra. Pour ce portfolio comme pour plusieurs projets clients, j'ai pris le chemin inverse. Un VPS à quelques euros par mois, une image Docker, un reverse proxy, un pipeline maison. L'idée n'est pas de revenir à l'âge de pierre du déploiement par FTP : je garde le « git push et c'est en ligne », mais sur une machine que je contrôle de bout en bout. Voici comment c'est câblé, et les deux ou trois endroits où je me suis fait avoir. L'image Docker : tout repose sur le mode standalone La pièce qui change tout, c'est output: "standalone" dans next.config.ts . Au build, Next trace exactement les fichiers nécessaires au runtime et les recopie dans .next/standalone/ . On passe d'une image d'environ 1 Go à environ 200 Mo. Sans ça, tu traînes tout node_modules dans ton conteneur de prod pour rien. Le Dockerfile est multi-stage : une étape pour installer les dépendances, une pour builder, une dernière qui ne garde que le strict nécessaire. # deps : installe les dépendances (cache Docker optimal) FROM node:22-alpine AS deps WORKDIR /app COPY package.json yarn.lock .yarnrc.yml ./ RUN corepack enable && yarn install --immutable # builder : build l'app FROM node:22-alpine AS builder WORKDIR /app COPY --from=deps /app/node_modules ./node_modules COPY . . RUN corepack enable && yarn build # runner : image finale, non-root FROM node:22-alpine AS runner WORKDIR /app ENV NODE_ENV=production RUN addgroup -g 1001 nodejs && a

The problem: agents forget your project every session If you pair with a coding agent, you have lived this: a new session starts and the context is gone. The agent re-discovers your auth flow, re-guesses why a decision was made, re-reads the same files to rebuild a mental model you already explained yesterday. Project knowledge — the why behind your systems, the runbooks, the "don't touch this, here's the reason" — lives scattered across wikis, code comments, and people's heads. None of it travels with the code, and none of it survives a fresh context window. CLAUDE.md helps, but it's for standing instructions , and it gets loaded wholesale into every prompt. Auto-memory captures what an agent picked up, but it's implicit, per-agent, and not reviewed. A wiki is for humans and needs exporting. There's a gap: curated team knowledge that's structured, versioned with the code, and readable by any agent or person. What OKF is Open Knowledge Format is an open, vendor-neutral format (announced by the Google Cloud Data Cloud team in June 2026, Apache-2.0) that represents knowledge as a directory of markdown files with YAML frontmatter . That's the whole idea. No schema registry, no runtime, no SDK. If you can cat a file you can read it; if you can git clone a repo you can ship it. A bundle looks like this: .okf/ ├── index.md # progressive disclosure (root carries okf_version) ├── log.md # ISO-dated change history, newest first ├── services/auth-api.md # one concept = one file; path is its ID ├── datasets/orders-db.md ├── decisions/use-okf.md ├── runbooks/payment-failures.md └── metrics/checkout-conversion.md Each concept needs exactly one thing to be conformant: YAML frontmatter with a non-empty type . Everything else is optional. --- type : Service title : " Auth API" description : " Issues and verifies short-lived access tokens." resource : https://github.com/acme/auth tags : [ auth , platform ] timestamp : 2026-06-14T10:00:00Z --- # Endpoints | Method | Path | Descriptio

You ask the AI for a bibliography. It hands you a title, authors, a journal, a year, a well-formed DOI. Everything is plausible, everything is clean. And one reference in two doesn't exist. Not "approximate": nonexistent. The DOI resolves to nothing, the paper was never written. The reflex is to ask the model again: "are you sure this source is real?" It says yes. Always. You just asked the forger about the authenticity of his forgery. Hallucination is plausible by construction An LLM doesn't store a database of publications. It generates likely sequences of words. A citation, to it, is a shape: a surname, an initial, two more names, a capitalized journal, a recent year, ten DOI digits. It produces that shape perfectly, because that's exactly what it's good at. The content doesn't need to be true to be plausible, it just needs to resemble. That's why a hallucinated reference is so vicious: it doesn't look like an error. A wrong calculation jumps out. An invented citation looks like a real one, until you click. Don't ask the culprit The golden rule fits in one sentence: never ask the model that hallucinated a citation whether that citation is real. For two reasons that compound. First, it doesn't have the information: it has no access to a registry, it can only regenerate something plausible. Second, even if it doubted, its self-evaluation bias pushes it to confirm what it already produced. You get a "yes" worth nothing. Verification has to come from elsewhere. From a source the model neither controls nor can invent: a metadata API. Three filters: existence, credibility, fidelity In my pipeline for writing technical dossiers, no reference enters the document before clearing three filters, in this order. Existence. The DOI must resolve. It's binary, and it's free. Crossref exposes its whole database: curl -s "https://api.crossref.org/works/10.1145/3290605.3300233" \ | jq '.message.title[0], .message.author[0].family, .message["published"]' If the API returns a title a

You ask Claude to review your code. It says "looks good, clean, well factored". Of course it does. It wrote that code five minutes ago. You just asked the author to grade his own paper, and he gave himself an A. Having an AI review code works. But not by asking the one who just wrote it. Quality doesn't come from a smarter model, it comes from an architecture where no role checks itself. The self-preference bias This isn't a hunch, it's measured. A model evaluating its own output rates it higher than others' at equal quality: the self-preference bias , documented by Panickssery and co-authors in 2024, and it's causal, not correlational. The model recognizes its own style and prefers it. In practice that means the naive loop "write, then review what you just wrote" is broken by construction. You don't get a review, you get a justification. The agent already decided its code was good the moment it produced it; asking again only confirms. The blind reviewer So the first rule: the reviewer is never the author. In my config, the review agents run in a clean context . They don't see the implementation prompt, they don't know what constraints the author set, they meet the diff like a colleague on Monday morning. And when the author is a known model, the reviewer is from a different family , to break style recognition. One detail matters as much as the rest: the developer's name never enters the reviewer's prompt. No "this was written by a senior", no "review this model's work". The author's identity is exactly the information that triggers the bias. We take it off the table. No finding without a receipt The second trap is the opposite of the first. An AI reviewer, especially in a clean context, tends to over-flag: it invents problems to look useful, it flags "vulnerabilities" that aren't. A review that cries wolf on every line is no better than a complacent one: either way, you stop listening. Hence the receipt rule. Every finding must cite a file:line and pass a check bef

You ask the AI for tests. It hands you twelve, all green. CI passes. You merge. Three days later a bug ships, on a function those tests were supposed to cover. You reopen the test file and it clicks: it ran, it passed, and it tested nothing. A green test isn't a proof. It's a hypothesis. And an AI, left to its own devices, is very good at writing hypotheses that can never be disproved. The phantom test Take a dead-simple function, a discount above 100 euros: func Discount ( total int ) int { if total > 100 { return total - 10 } return total } Here's the kind of test an AI produces when you ask "write me a test for this" with no further framing: func TestDiscount ( t * testing . T ) { got := Discount ( 150 ) if got < 0 { t . Errorf ( "result should not be negative" ) } } This test is green. It does run the discount branch (so your coverage climbs). But look at the assertion: got < 0 is never true, whatever Discount does. Replace total - 10 with total + 10 , with total * 2 , with 42 : the test stays green. It doesn't check behavior, it checks that the lights are on. Coverage doesn't measure what you think The trap is that this phantom test inflates your coverage. Coverage counts lines executed , not assertions that bite . A line crossed by a test that asserts nothing useful counts as much as a line genuinely verified. So a 90% coverage report can hide half a suite of tests that will never fall, even if you break the code on purpose. That's exactly an LLM's playground. Its reward signal is "the tests pass". Not "the tests catch a bug". With no external oracle to stop it, it drifts toward the shortest path to green: soft assertions, mocks that test themselves, cases that never exercise the risky branch. The red-check: break the code, demand the red The counter is one move, and it's as old as TDD: before trusting a test, check that it knows how to fail. Mutate the line it's meant to protect, rerun, and expect to see it go red. If it stays green, it's vacant. On our funct

One morning I pasted four principles into my CLAUDE.md , the global instruction file Claude Code reads at the start of every session. "Think before you code", "simplicity first", that kind of maxim you see fly by on X, credited to Andrej Karpathy. I felt clever for about a day. Then I watched Claude read the file, nod, and carry on exactly as before. A CLAUDE.md is a suggestion box. The model nods, then does whatever it wants. If I wanted it to code my way, writing it down wasn't going to cut it. I had to enforce it. What follows is what that frustration turned into: a config in four layers, reinstallable in one command, and a discovery that runs through everything else. The only rigor that counts is the one a model can't grant itself. Four layers, and only one really changes the behavior My config has four floors, from softest to hardest. The brain is CLAUDE.md : how I work, not the docs for my code. The rule that sums it up lives inside it: "what not to add: anything Claude rediscovers by reading the code." It holds my design principles, my stance on orchestrating subagents (I size up, I delegate, I verify: "I stay the brain, they're the hands"), and one line that becomes the thread running through the whole thing. The references : a go-best-practices.md file the brain points to in plain text whenever Go is involved. The skills : ten of them. A skill is a folder with a playbook that Claude loads on demand for a specific job: review code, write an article, distill a book. Mine are packaged as a marketplace, in a public GitHub repo , with a changelog and a version number. That's the real differentiator: versioned tooling, not just rules scribbled in a file. The guardrails , finally. And this is the only layer that reliably changes behavior. The first three, the model can read and ignore. The fourth, it can't. The four config layers, from softest (the model can ignore) to hardest (the model is bound by the guardrail) Brain CLAUDE.md: how I work References go-best-pra