标签：#AI

Le maillon le plus faible de ta sécurité a un pouls. Ce n'est pas ton pare-feu, ni ton chiffrement, ni ton dernier correctif. C'est une personne — et les attaquants le savent bien mieux que la plupart des équipes. Pourquoi forcer une porte blindée quand on peut simplement demander la clé ? La majorité des intrusions sérieuses ne commencent pas par un exploit technique génial. Elles commencent par un e-mail qui a l'air juste assez vrai, un appel qui semble venir du service informatique, une pièce jointe qu'une personne pressée ouvre sans réfléchir. La technologie tient. C'est l'humain qu'on contourne. Cela dérange, parce que c'est plus difficile à corriger qu'une faille logicielle. On ne corrige pas les gens. Mais on peut les préparer. La formation ne consiste pas à traiter les employés d'imprudents ; elle consiste à leur montrer à quoi ressemble vraiment une attaque, pour qu'ils la reconnaissent dans un moment de fatigue. Et il faut concevoir en supposant que quelqu'un se fera avoir un jour. Parce que quelqu'un se fera avoir. L'authentification à plusieurs facteurs, le moindre privilège, la limitation de ce qu'un compte compromis peut atteindre : tout cela existe précisément parce qu'un humain finira par cliquer sur le mauvais lien. La question n'est pas si, mais quand — et ce qui reste debout après. Alors ne consacre pas tout ton budget aux murs et rien aux personnes qui gardent les portes. Le maillon le plus faible a un pouls, un mauvais jour, et une boîte de réception pleine. Protège-le comme le reste de ton infrastructure, parce que c'en est la partie la plus exposée. – Serguey Shinder

TL;DR: Running high-frequency autonomous AI agent loops on commercial LLM APIs at scale is economically unsustainable and introduces unpredictable latency spikes. This production guide details how we deployed a self-hosted inference cluster using vLLM (v0.6+) , EAGLE-3 speculative decoding , PagedAttention v2 , and Automatic Prefix Caching (APC) on cloud GPUs (RunPod/Vast.ai), achieving a sub-180ms Time-To-First-Token (TTFT) , 118 tokens/sec throughput , and cutting inference costs by 45–74% . 1. The Economic & Latency Bottleneck of Agentic Loops When building 24/7 autonomous daemon agents , LangGraph multi-agent state machines , or LLM-driven NPC game loops , the computational profile differs fundamentally from human chatbot interactions: Massive Request Volume: A single complex agent decision cycle frequently executes 5 to 25 LLM calls across intent classification, tool schema validation, reflection loops, and output formatting. Repeated Prefix Redundancy: 80–90% of prompt tokens consist of identical system instructions, persona framing, and MCP (Model Context Protocol) tool definitions. Strict Latency Budgets: Real-time simulations and game loops cannot tolerate 800ms–1500ms commercial API network roundtrips. Commercial Closed APIs (GPT-4o / Claude 3.5 Sonnet) ├── Prefill: Paid per-token on every single cyclic call ├── Network Roundtrip: 250ms - 600ms latency overhead └── Cost at 50,000 daily agent iterations: $1,200 - $3,500 / month Self-Hosted vLLM Cluster (RTX 4090 / A100 on RunPod) ├── Automatic Prefix Caching (APC): Reuses KV-cache (120ms -> 12ms prefill) ├── Speculative Decoding (EAGLE-3): 2.1x generation throughput └── Fixed Infrastructure Cost: $245 - $480 / month (Flat, unlimited tokens) 2. Deep Dive: The vLLM Memory & Scheduling Architecture PagedAttention: Eliminating KV-Cache Fragmentation Standard PyTorch/HuggingFace transformer implementations allocate static KV-cache tensors sized for max_sequence_length . Because 95% of queries generate far fewer

On ne gère pas ce qu'on ne mesure pas. C'est l'une des premières leçons de l'exploitation, et pourtant je l'ai apprise à l'envers, en pilotant à l'aveugle bien trop longtemps. Sans mesure, tu ne sais pas si un système va bien. Tu le supposes. Il tourne, personne ne se plaint, donc tout va bien — jusqu'au jour où quelque chose se dégrade lentement, sous le radar, et où tu ne l'apprends que lorsque c'est déjà une panne. La lente fuite de mémoire, le disque qui se remplit, la latence qui grimpe d'une milliseconde par semaine : rien de tout cela ne crie. Ça glisse. La mesure transforme les suppositions en faits. Un tableau de bord, quelques alertes bien choisies, et soudain tu vois le problème arriver au lieu de le subir. Tu n'attends plus que l'utilisateur t'apprenne que ton système est cassé ; tu le sais avant lui. Mais il y a un piège que j'ai appris à éviter : mesurer trop. Cent métriques que personne ne regarde ne valent pas mieux que zéro. Le bruit noie le signal, et les alertes qui se déclenchent sans raison finissent par être ignorées — jusqu'à celle qui comptait vraiment. Bien mesurer, ce n'est pas tout mesurer. C'est choisir les quelques signaux qui prédisent réellement un problème. Alors, avant de bâtir la prochaine chose, demande-toi comment tu sauras si elle va mal. Si la réponse est « quelqu'un finira par le remarquer », tu ne la gères pas encore. Tu espères. Et l'espoir n'est pas une stratégie d'exploitation. – Serguey Shinder

Alle paar Jahre verkündet jemand, dass die IT-Jobs verschwinden. Die Cloud ersetzt die Administratoren. Die Automatisierung ersetzt die Operatoren. Jetzt ersetzt die KI, was übrig ist. Und alle paar Jahre verschwinden die Jobs nicht, sie wandern. Sie wandern den Stack hinauf. Als physische Server der Cloud wichen, brauchten wir keine Infrastruktur-Leute weniger, wir brauchten Menschen, die Infrastruktur auf einer höheren Ebene verstehen, die Systeme entwerfen, absichern und kostenbewusst betreiben, statt Hardware in Racks zu schrauben. Als manuelle Deployments der Automatisierung wichen, brauchten wir Menschen, die die Automatisierung selbst bauen konnten. Jede Welle entfernte nicht die Arbeit. Sie hob den Boden und verschob den wertvollen Teil nach oben. Die KI ist die nächste Welle, und ich erwarte, dass sie sich genauso verhält. Sie wird viel Routine aufsaugen: das Skripten, die ersten Config-Entwürfe, die Standard-Fehlersuche. Was sie nicht aufsaugt, ist Urteilsvermögen: zu wissen, was zu bauen ist, zu entscheiden, was das Risiko wert ist, zu verstehen, wie die Teile einer echten Organisation zusammenpassen, und geradezustehen, wenn etwas schiefgeht. Wer in jedem Übergang strauchelt, sind die, die sich über die Aufgabe definieren, die automatisiert wurde. Wer gedeiht, definiert sich über das Problem, das er löst, und lässt die Werkzeuge dafür sich darunter ändern. Verteidige also nicht das eine, das du heute tust. Werde gut in der Schicht darüber. Der IT-Job hat jedes Werkzeug überlebt, das ihn beenden sollte, indem er den Stack hinaufwanderte. Diesen wird er genauso überleben. – Serguey Shinder

There's a lot of capital pouring into the business of giving models away.

James Cameron's 1991 sci-fi blockbuster returns to theaters this weekend for its 35th anniversary.

Just as new data centers face growing backlash from neighboring communities, the US Environmental Protection Agency (EPA) is about to make it harder for people to weigh in on any pollution those centers create. The EPA plans to toss out a federal rule requiring public notice and an opportunity to comment when certain industrial sites […]

It's allegedly been running dozens of gigantic generators without a permit.

Previously I have a macOS App I use myself, gemini-live-translate-macos . It uses ScreenCaptureKit to directly capture audio from a specified App, eliminating the need for virtual sound cards like BlackHole. It then sends the audio to the Gemini Live API for real-time translation, outputting Traditional Chinese subtitles while playing Chinese audio. I've written two posts about the development process: the first one was about building it from scratch using AGY CLI, and the second one was about using Claude Code to take it from "functional" to "user-friendly." The starting point for this new addition was simple: I saw a document for "Real-time Transcription" added to the Live API. Since I was already connected to the Live API, I thought adding a pure transcription mode would just be a matter of changing a few parameters. However, after checking the documentation, I realized that Google released two models with very similar names but very different capabilities at once. The specific feature I actually wanted (speaker diarization) wasn't available at all on the model I originally thought it was. Two Models with Names Differing by Only Two Words Let's lay out the differences first; this is the part I spent the most time figuring out: gemini-3.5-transcribe-live gemini-3.5-transcribe API Used Live API (WebSocket streaming) Interactions API (Standard HTTP request) Usage Scenario Transcribe while speaking Upload the whole file after recording Speaker Diarization Not supported Up to 8 speakers Word-level Timestamps Not supported Supported Audio Length 10 minutes per session 1 hour (30 mins with diarization) Smart Mode SMART available smart is mutually exclusive with diarization Interim Subtitles Has interimInputTranscription Not applicable The official documentation on the Live page's limitations section is very blunt: Speaker diarization is not supported in live streaming sessions. For speaker diarization, use the non-streaming Audio transcription endpoint. So, "seeing who

The fastest way to lose trust in an AI support assistant is to watch it confidently invent an answer. That is the problem retrieval-augmented generation (RAG) is built to solve. Instead of relying on whatever a language model memorized during training, a RAG system first retrieves the relevant passages from your own knowledge base, your help docs, policies, and product specs, then generates a reply grounded in that verified content. The result is an assistant that answers from your truth, cites its sources, and reflects updates the moment you publish them, with no expensive retraining. The business case is hard to ignore. Teams deploying RAG report sharply lower issue resolution times, deflected tier-one tickets, and agents who get accurate suggested answers in seconds. But RAG is not a switch you flip. Answer quality lives or dies on how well you chunk, embed, and retrieve your content, and on the guardrails you put around what the model is allowed to say. This guide walks through how RAG actually works and what it takes to ship it responsibly. Why plain LLMs fall short for support A bare language model is a confident generalist with three liabilities that are fatal in a support context. Its knowledge is frozen at training time, so it has never seen your latest pricing, your current return policy, or the feature you shipped last week. It has no access to your specifics, so it does not know this customer's plan or your internal escalation rules. And when it does not know, it tends to guess fluently. In casual use a wrong answer is an annoyance. In customer support it is a refund processed against the wrong policy, a security claim that isn't true, or a promise your team now has to honour. RAG addresses all three problems by grounding every answer in your verified, current content, so the model stops improvising and starts citing. In plain terms A plain chatbot answers from memory and sometimes guesses. A RAG assistant looks up the answer in your documentation first,

For fifteen years the goal of search was simple: rank on page one and earn the click. That contract is breaking. More people now ask ChatGPT, Perplexity, Gemini and Google's AI Overviews a question and read the synthesized answer without ever visiting a blue link. If your brand is not in that answer, you are invisible to them, no matter how well you rank. This is the gap Generative Engine Optimization closes. GEO is the practice of structuring your content and your site so that large language models retrieve it, trust it, and cite it when they answer a question in your space. It overlaps with SEO but it is not the same job. One study from the GEO firm Brandlight found the overlap between top Google links and the sources AI tools actually cite has fallen from around 70% to under 20%, and the gap is widening. The payoff is real, not theoretical. AI referrals convert far better than cold organic traffic because the visitor arrives pre-qualified by the answer that sent them. Vercel has reported that roughly 10% of new signups now come from ChatGPT, and LLM-referred visitors have been measured converting at 15.9% from ChatGPT against under 2% for typical organic search. Here is how we approach GEO for the sites we build. SEO earns clicks, GEO earns citations The mental shift is the whole game. Traditional SEO optimizes a page to win a position in a ranked list of links. GEO optimizes a passage to be quoted inside a generated answer. A model does not "rank" your page; it retrieves chunks of it, weighs them against everything else it pulled, and decides whether to repeat your claim and name you as the source. That changes what good content looks like. Models favor passages that are self-contained, factual, and quotable: a clear definition, a specific number, a direct answer in the first sentence. Burying the answer three paragraphs down, the way you might to keep a reader scrolling past ads, is exactly wrong here. Lead with the claim, then support it. Write so a model can

If your website has a support chatbot, an AI assistant, or any feature that writes text or generates images for visitors, a hard deadline is now on your calendar. From 2 August 2026, Article 50 of the EU AI Act requires you to tell people when they are talking to a machine or looking at AI-generated content. It is the first part of the Act that touches ordinary businesses rather than just labs building frontier models, and it applies the way GDPR does: not by where your company is based, but by whether people in the EU use your product. The good news is that Article 50 is mostly about transparency, not heavy engineering. You are not being asked to prove your model is safe or file a conformity assessment. You are being asked to be honest about what is human and what is not. The catch is that "be honest" has specific rules, the fines are large enough to matter (up to 35 million euros or 7% of global turnover for the worst breaches), and a surprising number of teams discover they are a "provider" of an AI system when they assumed they were just a user of one. Here is what Article 50 actually requires, how to tell which obligations land on you, and the short list of changes most businesses need to ship before August. What Article 50 actually requires Article 50 covers four situations, and each one has its own disclosure rule. Chatbots and conversational AI must tell users they are interacting with an AI system, unless that is already obvious to a reasonable person. AI-generated or manipulated media (synthetic images, audio, and video, including deepfakes) must be labelled as artificially generated. AI-generated text published to inform the public on matters of public interest must be disclosed as such. Emotion recognition and biometric categorization systems must inform the people exposed to them. Most businesses are touched by the first two. If you run a chatbot, the disclosure has to be clear and come at the right moment, which in practice means before or at the start

Every team has a list of small jobs nobody wants to do: copy a new lead from the website into the CRM, chase an unpaid invoice, summarise a support email, post the same update to three places. None of them is hard. Together they eat hours every week and quietly burn out the people stuck doing them. Automation platforms exist to take that list off your hands, and in 2026 the three names you will keep running into are Zapier, Make and n8n. What changed recently is that all three stopped being simple "when this, then that" tools and grew real AI agents. Zapier shipped Agents that act across its 8,000-plus app catalogue. Make added Maia, an assistant that builds whole workflows from a sentence. n8n 2.0 went deepest, with native LangChain support, 70-plus AI nodes, persistent memory and human approval steps. So the question is no longer whether to automate, it is which platform fits how you work, what you can afford to run at scale, and how much control you need over your own data. Here is how the three actually differ, and how to pick without locking yourself into the wrong one. The honest three-way comparison The marketing pages all promise the same outcome. The real differences show up in pricing models and ceilings. Zapier is the easiest to start with and the most expensive to scale. It bills per task , and every single action counts. A ten-step workflow that runs 1,000 times a month burns 10,000 tasks. Great for non-technical teams wiring up a handful of common apps; painful once volume grows. Make sits in the middle. Its visual canvas is genuinely pleasant for branching, multi-step logic, and it bills per operation at a lower unit cost, staying under roughly 100 euros a month even at high volume. The sweet spot for teams that want real logic without code. n8n is the power tool. Free if you self-host, 20 to 50 euros a month on its cloud, plus the AI token costs you would pay anyway. It is the most AI-native of the three and the only one you can run entirely on your

Meta fixes AI glasses to stop recording any time users cover up the safety light.

Retro, a friend-focused photo-sharing app built by former Instagram employees, has raised more than $21 million in Series A funding.

Now, Apple TV subscriptions will cost $14.99 per month, up from its previous price of $12.99 per month.

The firm, known for its focus on software, is going to start throwing more money at the hardware behind AI.

A federal judge ruled the Trump administration illegally labeled Anthropic a supply chain risk, handing the AI company a victory as its second Pentagon lawsuit continues in Washington.

Hello! I'm a beginner developer with my sights set on backend development and data modeling. Like a lot of people starting out, I didn't come in with a computer science degree or years of professional experience — just curiosity about how applications actually store, organize, and make sense of data behind the scenes. Backend work has always felt like the "engine room" of software to me. While frontend gets the visual credit, it's the data layer that quietly decides whether an application is fast, reliable, and able to grow. That's what pulled me toward backend and database design in the first place. My biggest challenge so far has been learning SQL and data modeling from scratch. It sounds simple on paper — write some queries, design some tables — but in practice it meant rewiring how I think. I had to move from "how do I make this work right now" to "how do I structure this so it still works when the data grows, the requirements change, or someone else has to read my schema six months from now." Concepts like primary keys, foreign keys, relationships between tables, and eventually normalization weren't hard to memorize, but they were hard to internalize — to actually reach for instinctively when designing something from a blank page. A few things clicked for me along the way: A good schema is a form of communication. Table and column names, relationships, and constraints tell a story about the business logic, not just the data. Getting it "perfectly right" on the first try isn't the goal. Iterating on a design after seeing how data actually flows through it taught me more than any tutorial did. SQL rewards precision. Small differences — a missing JOIN condition, the wrong key, an unindexed column — can quietly break correctness or performance, so being deliberate matters. Constraints are a beginner's best friend. Things like NOT NULL, UNIQUE, and foreign key constraints catch mistakes early instead of letting bad data pile up silently. This foundation in SQL and d

I wanted to start experimenting with AI for a while. At work, I didn't really have enough time to properly explore this new ecosystem and understand how these tools could be used. But I had another resource available: my personal hardware. I'm a gamer, so I have a reasonably powerful desktop PC. I'm also a developer, with a laptop running Ubuntu and a NAS. That made me wonder: Can I use my gaming PC to run AI models locally? The answer is yes . This article is a small walkthrough of my experience setting up a local AI environment using an AMD Radeon GPU , Windows , WSL , ROCm , and Ollama . My hardware is (yes, I'm Team Red): AMD Ryzen 7 5700X 32 GB RAM AMD Radeon RX 9060 XT with 16 GB VRAM Ubuntu 26.04 running under WSL Installing WSL and Ubuntu The first step is to make sure that WSL is installed and up to date. The general process is: Check your current WSL version. Update WSL. Install Ubuntu. Create your user account. Update the Ubuntu installation. WSL From PowerShell: # Check the current version of WSL wsl -v # Update WSL wsl --update # Install Ubuntu wsl --install -d Ubuntu # Start Ubuntu wsl -d Ubuntu # Stop Ubuntu wsl --terminate Ubuntu # Remove the Ubuntu distribution wsl --unregister Ubuntu Ubuntu Once Ubuntu is installed, update the system and install zstd . We will need it later to extract Ollama. sudo apt update sudo apt upgrade sudo apt autoclean sudo apt autoremove sudo apt install zstd Installing Ollama with ROCm support Ollama is the foundation of this setup. It allows us to download and run AI models locally while providing different compute backends, including CPU, CUDA, and ROCm. Installation One thing that wasn't immediately obvious to me was that Ollama provides a specific ROCm backend. The installation is split into two parts: Install Ollama. Install the ROCm backend. # Install Ollama curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \ | sudo tar --zstd -x -C /usr # Install the ROCm backend curl -fsSL https://ollama.com/downlo