标签：#AI

Introduction — On What I've Been Writing for Years This is a follow-up to my previous post on Claude and MCP . Just sharing some recent thoughts. Personally, I've always enjoyed keeping records and analyzing my own work. So for years, I've been logging my daily tasks, jotting down thoughts, hesitations, and impressions in notes. I've drawn on these records for reviews, analysis, and decisions on various projects. The tools have shifted over time — Evernote, Notion, Logseq, Taskuma, and so on — but the habit itself, of writing notes into some app or tool, has stayed with me for years. What Happened with MCP I recently wrote about connecting Notion and Google Docs through MCP, and the results have surprised even me. I won't repeat the details here since they're in that post, but ever since I introduced MCP, the flow of information has accelerated dramatically. In particular, I'd been accumulating reviews, task management notes, and brainstorms in Notion for years, and letting Claude read all of this has shifted the meaning of what I'd previously written. When I first started recording in Notion, it never occurred to me that it might be useful to AI. Of course — I had no way to imagine a time when AI would become this close to everyday life, used in this way. I was just writing for plain, analog reasons — "so I could look back later," "so I could organize my own thinking." But the moment MCP made it all readable, the feeling shifted. It's as if my past self comes forward to help my current self. Claude answers my current questions while drawing on the reasoning behind old project decisions, or on impressions I'd noted at the time. I've had moments like that more than once now. Thinking about it: the human brain's memory has limits — even the person who wrote something forgets it quickly. That's why I kept taking notes, leaving behind my thoughts and conclusions at each point in time as a record. And now, in the flow of conversation, AI reads from those records, distill

His doctors went looking for cancer, then they saw the worms' heads.

A few days back I wrote about giving an autonomous agent database access and building a firewall so it couldn't DROP TABLE prod. Same lesson, new surface: this time the agent had cloud credentials . The failure mode isn't a destructive command here. It's spend. An agent pointed at a networking task can scan a whole range looking for hosts, then spin up a fleet of instances to do it faster. Every individual call is "authorized," your IAM role said yes. The bill is what eventually says no. ## Two shapes, two right answers The interesting part is that these are not the same kind of problem, so they don't get the same verdict. 1. The scan is never legitimate as an agent tool call. An nmap -sS -p- 10.0.0.0/16 or a masscan across a network is reconnaissance and abusive egress. There's no benign version of an agent sweeping a network at scale, so it gets hard-blocked , deterministically, before the call runs. (A scan of your own localhost is a dev check, so that's exempt.) 2. The provisioning might be totally fine. Spinning up 50 instances could be a real scale-out, or a runaway loop burning money. You can't tell from the action alone, only from the consequence. So instead of blocking it, AgentX pauses it for a human : a 202, "held for approval," routed to whoever owns the budget. Block the thing that's never okay, escalate the thing that's sometimes okay. Gate on consequence, not identity. Both checks are zero-LLM. No model in the hot path means no latency tax and nothing to talk out of it. A runaway fleet should be caught by a rule, not a vibe. ## The bigger thing this closes We keep a catalog of real, documented agent failures and triage each one: is it something an action firewall can deterministically catch, or is it someone else's category (output hallucination, content safety, model internals)? We only build for the coverable ones, and we flag the rest honestly instead of faking a signature. With this release, the coverable list is done . Every failure shape an acti

Most cloud platforms that need access to your infrastructure start with the same onboarding step: paste in a service account key. Or an access key and secret. Or a JSON blob you downloaded from the console and definitely should not be emailing to yourself. You paste it in. The platform stores it. You hope they encrypted it. You hope they rotate it. You hope nobody on their team can read it. You move on with your day and try not to think about it. We built Zero — b0gy's platform for engineering truth — around a different premise. For cloud infrastructure access — GCP and AWS — we don't store credentials at all. The platform connects to your projects and accounts using short-lived, federated identity tokens that are minted on demand and expire in minutes. There is nothing to leak because there is nothing stored. Not every integration can work this way. GitHub, Slack, and Jira use OAuth, which means we do hold tokens for those services. But for the highest-risk connections — the ones with read access to your entire cloud infrastructure — keyless was a hard requirement. This is the first post in a three-part series about building Zero. We're starting here because the connector model shaped everything else. Why stored secrets are the wrong default The argument for storing a service account key is convenience. You paste it once, the platform can access your cloud whenever it needs to. Simple. The argument against it is longer. A stored secret is a liability that compounds over time. The moment you paste a GCP service account key into a third-party platform, you've created a credential that is valid indefinitely, scoped to whatever permissions you granted, and stored in a system you don't control. If that platform gets breached — or if an employee with database access gets curious — that key works until someone revokes it. And nobody revokes it, because nobody remembers it exists. This isn't theoretical. The GitGuardian 2026 report found 28.65 million hardcoded secrets pus

I kept correcting AI agents in the same ways: "too long," "answer first," "use a diagram," "assume I know the jargon." Each correction improved the current exchange, but the preference was not represented as durable system state. I built /calibrate-comms to make that state explicit. It is an open-source skill inside an Obsidian vault, used by both Claude Code and Codex. The model: nine operational dials The skill does not try to discover a personality type. It calibrates nine choices that directly change how an answer is rendered: Dial Practical question Density Tight sections or full reasoning? Sequence Answer first or chronological build-up? Modality Prose or diagrams for relational content? Abstraction Concrete example or principle first? Tradeoff One recommendation or several options? Detail Main path or edge cases too? Jargon Define terms or assume expertise? Tone Casual, neutral, or formal? Context-giving Should the agent extract missing context or split an overloaded brief? Prior → calibration → directives The workflow has three stages: L1 PRIOR → L2 SAMPLE REACTION → COMPILE → CLAUDE.md hypothesis empirical override shared by both agents Quick mode asks one bespoke forced-choice proxy per axis. Those questions are deliberately labelled as proxies, not validated psychometrics. Deep mode must fetch the exact items from supported open-access instruments before use; if the source cannot be obtained, the skill stays in Quick mode and makes no validation claim. The prior is then challenged through pairwise samples. For sequence, the contrast looks like this: Build-up first: We traced the latency spike to N+1 queries, then found lazy loading in a loop—so the fix is eager loading. Answer first: Fix: eager-load the association. Why: lazy loading in a loop caused N+1 queries and the latency spike. The user's pick is revealed preference. If it contradicts the prior, the sample wins. The compiler is the useful part The final profile is not a score report. A deterministi

I'm building LOOM — a small open-source language that is a machine-checked trust layer for AI-written code. I don't write it by hand anymore: an organism I built grows it, day and night, on my own machine. This is Day 6, and the whole day went to one thing — WebAssembly . Why this was a real test LOOM already runs three ways: an interpreter, and backends that compile checked code to Python and JavaScript. The thesis is "trust survives translation" — effects and provenance, proven once, hold the same on every target. WebAssembly is the strongest test of that: a low-level stack machine with linear memory, nothing like Python or JS. And there was a constraint. This machine's clang has no wasm target, and I install nothing paid or heavy. So I don't compile to wasm through a toolchain — I emit the wasm bytes myself (LEB128, the type / function / memory / global / export / code sections, the i32 stack machine) and run them through node's built-in WebAssembly . Zero dependencies. From fib to a value runtime, in a day Every step was prototyped and proven (wasm output == interpreter output) before it touched the kernel: The integer core — arithmetic, comparison, if , first-order calls and recursion. fib(10) becomes 61 bytes of real WebAssembly and returns 55, identically on the interpreter, Python, Node and wasm. A value runtime — let and integer lists in a real linear-memory heap (a bump pointer + a $cons cell allocator; head / tail are i32.load , empty is i32.eqz ). A list sums and folds by recursion, inside wasm. Sum types — (variant Tag e) becomes a tagged cell [tag-id | payload] ; match loads the tag, compares, binds the payload, branches. You can watch it: the live playground has a Compile → WAT button and WASM · fib / list-sum / match examples. Type a program, see it become real assembly, in your browser. Honest scope: ints, let , integer lists and sum types compile to wasm today. Records, closures and effects are the next frontiers (closures are the hard one — a func

TMX: The open standard AI agent memory has been waiting for The problem no one talks about: your agent's memories are prisoners. If you build an AI agent today using Mem0, your memories are locked in Mem0. Switch to Zep? You lose everything. Move to a new framework? Start from zero. This is exactly the problem email had in 1970. Every system had its own format. You couldn't send an email from one system to another. Then SMTP was invented. And email became universal. Today I'm publishing TMX v0.1 — the SMTP of AI agent memory. What is TMX? TMX (Truvem Memory eXchange) is an open, model-agnostic JSON format for storing, exporting, and importing AI agent memories across any platform, framework, or provider. It looks like this: { "tmx_version" : "0.1" , "exported_at" : "2026-06-26T20:00:00Z" , "source" : "truvem" , "agent_id" : "my-agent" , "memories" : [ { "id" : "550e8400-e29b-41d4-a716-446655440000" , "content" : "User prefers dark mode and concise responses" , "created_at" : "2026-06-01T08:30:00Z" , "updated_at" : "2026-06-01T08:30:00Z" , "expires_at" : null , "tags" : [ "preference" , "ui" ], "source_model" : "gpt-4o" , "metadata" : {} } ] } That's it. Plain JSON. Human-readable. Portable. Why this matters Right now, the AI agent ecosystem is exploding. Every week there's a new memory provider, a new framework, a new cloud service. But every one of them uses a proprietary format. This means: Developers are locked to their first choice forever Agent memories can't travel between clouds Switching providers = losing everything your agent learned This is the biggest hidden tax in the agentic AI stack. TMX fixes it with a single open spec that anyone can implement — for free, with no approval needed. The 5 core principles 1. Open — No license required. Implement TMX in any product, commercial or otherwise. 2. Model-agnostic — Works with GPT-4, Claude, Gemini, Mistral, Llama, or any future model. 3. Framework-agnostic — LangChain, CrewAI, Mastra, AutoGen — doesn't matter

NYT shifts OpenAI/Microsoft copyright claims after SCOTUS ruling against Sony.

Everyone is selling AI to small businesses right now. Most of it is hype. But some of it is genuinely useful — and knowing the difference can save you thousands in wasted tooling. I run a small agency in Stuttgart that builds websites and automations for local service businesses: coaches, doctors, beauty studios, consultants. Here's what actually moves the needle for them in 2025. What "AI Automation" Actually Means for Small Businesses Forget the generic pitch. For a local service business, AI automation is useful in exactly three places: Client communication at scale — responding to inquiries 24/7 without hiring a receptionist Reducing admin time — intake forms, follow-ups, reminders, invoicing triggers Content creation — but only as a speed boost, not a replacement for your voice Anything beyond that is usually overkill for a business under 10 employees. The One Automation Every Service Business Should Have Automated follow-up after initial contact. Here's the typical flow without automation: Client fills out contact form You see it 4 hours later You write a reply If you're busy, it takes a day Client has already booked elsewhere With automation: Client fills out form Immediate confirmation email ("Got your message, here's how to book a slot") Link to booking calendar You're notified. If they don't book in 48h, a follow-up email goes out automatically This alone converts 20-40% more inquiries into booked clients. No AI model needed — just a simple workflow in n8n, Make, or Zapier. Where LLMs Actually Help Language models (ChatGPT, Claude, etc.) are genuinely useful for small businesses in these areas: Intake Forms → Personalized Responses A coaching client fills out a detailed intake form. Normally, you'd spend 20 minutes reading it and writing a personalized welcome email. With a simple LLM integration: Intake form submitted Webhook fires to n8n LLM reads the form, generates a personalized summary + welcome You review it in 30 seconds and hit send Same personal

Hello, I'm Shrijith Venkatramana. I'm building git-lrc, an AI code reviewer that runs on every commit. Star Us to help devs discover the project. Do give it a try and share your feedback for improving the product. An AI model writes a paragraph. It sounds fluent. It looks convincing. But how do you know whether it's actually good? This deceptively simple question has occupied researchers for more than two decades. Long before ChatGPT, machine translation researchers faced exactly the same problem. Human evaluation was expensive, inconsistent, and painfully slow. If every new model required thousands of humans to compare translations, research would crawl. That necessity gave rise to BLEU , one of the most influential evaluation metrics in AI history. Years later, as language models became better at paraphrasing and reasoning, BLEU started to show its age. Researchers responded with learned metrics like BLEURT and COMET , which use neural networks to judge language much more like humans do. Interestingly, this mirrors software engineering itself. We first wrote simple unit tests, then integration tests, and today we increasingly rely on sophisticated observability systems. Evaluation metrics for LLMs have undergone a similar evolution. Let's see why. Before BLEU: The Evaluation Bottleneck Imagine you're building Google Translate in 2001. Every time your team improves the model, someone has to read thousands of translated sentences and score them. Suppose a single sentence pair takes only 20 seconds to judge. Evaluating 50,000 sentences would require nearly 280 human-hours . Now imagine dozens of experiments every week. Evaluation—not training—quickly becomes the bottleneck. Researchers at IBM, led by Kishore Papineni , introduced BLEU (Bilingual Evaluation Understudy) in 2002 to automate this process. Their idea was surprisingly simple: If a machine translation resembles what professional translators write, it's probably good. This became one of the most cited papers

“We don’t believe this kind of government access process should become the long-term default,” says OpenAI. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”

How I built an open-source Slack assistant with persistent semantic memory, powered by any LLM and Mem0's managed memory layer — no vector database required. The problem with Slack bots and memory Most AI Slack bots have the memory of a goldfish. Every conversation starts from scratch. You ask it about your sprint goals, it gives a great answer, then three days later you ask a follow-up and it has no idea what you're talking about. You end up re-explaining context constantly. The commercial solution to this is Claude Tag — a Slack integration that maintains genuine conversational continuity. But it's tied to one provider and not open-source. slacktag-oss is our attempt to replicate that experience: a Slack bot with real, semantic, persistent memory that works with any LLM — including ones running entirely on your laptop. What I built A Python Slack bot with: Socket Mode for local dev (no public URL needed), HTTP-ready for prod LangChain to abstract LLM calls across any OpenAI-compatible endpoint Mem0 managed cloud for semantic memory — no Qdrant, no Pinecone, no infra to run Three memory scopes: per-channel, per-thread, per-DM Built-in !clear and !memory commands A clean, extensible architecture you can fork and build on Architecture Before diving into code, here's the full request lifecycle: ┌─────────────────────────────────────────────────────────────┐ │ Slack │ │ @mention in channel ──┐ │ │ DM to bot ──┼──► Slack Events API │ │ Thread reply ──┘ │ │ └───────────────────────────────────│─────────────────────────┘ │ (Socket Mode / HTTP) ▼ ┌─────────────────────────────────────────────────────────────┐ │ slack-bolt (Python) │ │ bot.py ──► router.py ──► handler.py │ │ │ │ │ ┌───────────────┤ │ │ │ │ │ │ ▼ ▼ │ │ Mem0 Client LangChain │ │ (managed) ChatOpenAI │ └────────────────────────────────────────────────────────────-┘ │ ▼ ┌───────────────────────┐ │ Mem0 Managed Cloud │ │ Vector Embeddings │ │ Entity Extraction │ │ Deduplication │ └───────────────────────┘ The ke

The hire marks OpenAI's latest push into India, expanding offices, partnerships and hiring.

One thing I've noticed after using AI for development over the past year is this: The code it generates is usually correct. The architecture slowly isn't. That doesn't happen because AI writes bad code. It happens because architecture rarely erodes all at once. Imagine a modular application with clear boundaries. The billing module talks to the orders module through its public interface. Authentication is isolated. Notifications are independent. Everything is predictable. Now imagine hundreds of AI-assisted commits over the next few months. One suggestion imports an internal class because it already exists. Another bypasses a service layer because it's shorter. A helper gets copied into another module. A database query is duplicated instead of reused. None of those changes are catastrophic. In fact, every pull request probably gets approved. The application still builds. The tests still pass. Customers never notice. Until one day, making a simple change requires touching five different modules because everything has quietly become connected. That's architecture debt. And unlike a failing test, it doesn't show up immediately. One thing I've realized is that our current tooling doesn't really watch for this. Unit tests verify behavior. Integration tests verify interactions. Linters enforce style. Static analysis finds bugs. All of those are important. But none of them are asking questions like: Should this module depend on that one? Did someone bypass a defined boundary? Are we introducing new architectural coupling? Is the overall architecture getting healthier or worse over time? Those questions usually get answered during code review. Or worse, during a production incident. The interesting part is that AI isn't really the problem. If anything, it's doing exactly what we ask it to do. It optimizes for solving the problem in front of it. Architecture, on the other hand, is about protecting the system as a whole. Those are different goals. As AI makes us write code fa

J'utilise la dictée vocale tous les jours depuis six mois. Pas pour taper moins vite. Pour penser plus vite quand je vibe-code avec Claude Code et Cursor. Pis j'ai fini par construire mon propre outil parce que les outils existants me tapaient sur les nerfs d'une façon très précise. Le problème réel Quand tu travailles en tech au Québec, tes phrases ressemblent à ça : "OK fa que je fais un useState pour le component pis je passe le handler en props" Ça, c'est une phrase normale. Personne en tech QC ne parle autrement. Pas parce qu'on est négligents avec la langue. Parce que le vocabulaire technique vient de l'anglais et qu'on le soude naturellement au français au fil de la pensée. Ça s'appelle le code-switching. Et c'est là que la plupart des outils de dictée craquent. Ce que les outils mainstream font mal Dragon NaturallySpeaking Dragon, c'est le vieux standard. Médical, juridique, corporate. Ça coûte environ 500$ en une shot. C'est lourd à installer et à entraîner. Et sa gestion du français québécois avec des termes tech intercalés... c'est en gros zéro. "useState" devient "usé état". "Fa que" devient "faque" parfois, "fake" d'autres fois. C'est aléatoire. T'as intérêt à corriger après chaque phrase. Wispr Flow Wispr Flow est plus moderne. UX propre, cross-platform, et leur gestion du français s'est améliorée. Leur plan Pro coûte 15$/mois, soit environ 144$/an. Mais il y a un problème structurel que leur propre doc admet : la détection de langue se fait par session, pas par mot. Autrement dit : Wispr détecte la langue une fois au début de la session. Si tu commences en français, il reste en mode français jusqu'à la fin. Les mots anglais qui arrivent dans la phrase, il tente de les translittérer en français. "Handler" peut devenir "andler" ou "ender", "props" survit parfois, parfois pas. C'est variable. Pour une phrase de temps en temps avec un mot anglais, ça passe. Pour un vibe-coder québécois qui switch constamment dans la même phrase, ça ne passe pas. Pourquoi

Nvidia has dominated the AI chip market for years, but the era of total dependence might be ending. OpenAI just shared its plans to spice things up with Jalapeño, its custom inference chip built with Broadcom, joining Google, Apple, and SpaceX in a growing list of companies building their way out of single-supplier risk. The goal is less of a […]

The White House asked OpenAI to delay the rollout of its GPT-5.6 AI models, two weeks after Anthropic had to take its most advanced AI models offline.

Less than 24 hours after news broke that OpenAI would stagger its next model release at the request of the Trump administration, that model, GPT-5.6, is here. On Friday, the company unveiled the limited preview of its new GPT 5.6 model suite: Sol, the flagship; Terra, a medium-tier model for "high-volume work"; and Luna, a […]

AI models have progressed to the point where their capabilities have real political consequences. Dealing with those consequences will require collective action.

You don't have a million labeled images or a GPU farm — and you don't need them. Transfer learning lets you stand on a model someone else trained and reach high accuracy with a few examples in minutes. Here's the idea, visualized. ♻️ Race scratch vs transfer: https://dev48v.infy.uk/dl/day17-transfer-learning.html The insight The early layers of a trained network learn general features — edges, textures, shapes — that are useful for almost any vision task. Only the last layers are task-specific. So why relearn edges from scratch? Two ways to do it Feature extraction: freeze the pretrained backbone, replace the final classifier with a small new "head," and train only the head on your data. Fast, needs little data. Fine-tuning: also unfreeze the top few backbone layers and train them at a low learning rate so you adapt without wrecking what they learned. The demo races two accuracy curves: "from scratch" crawls up and plateaus low (not enough data); "transfer learning" starts high and climbs fast. Tweak the example count and freeze/fine-tune to see them respond. Why it matters now This is exactly why fine-tuning an open LLM works: a foundation model already learned language; you adapt it cheaply. Transfer learning is what makes deep learning practical for the rest of us. 🔨 Full recipe (load pretrained → freeze → new head → train → optionally fine-tune low-LR) on the page: https://dev48v.infy.uk/dl/day17-transfer-learning.html Part of DeepLearningFromZero. 🌐 https://dev48v.infy.uk