标签：#m

The European Union's AI Act establishes a risk-based framework for AI systems that ranges from prohibited practices to minimal-risk uses. Regulation (EU) 2024/1689 divides the framework into four levels: unacceptable risk, high risk, limited risk and minimal risk. For AI developers, vendors and enterprises, the practical importance is straightforward: the system's risk category determines whether it can be used and, if so, the level of compliance, transparency and governance expected around it. The regulation entered into force on 1 August 2024 . Its four-tier approach is designed to avoid applying the same regulatory burden to every AI use case. Instead, the Act reserves its strictest treatment for systems that present the greatest risk, while leaving minimal-risk systems without additional sector-specific obligations under the AI Act beyond general law. The definitive reference is the official text of Regulation (EU) 2024/1689 . Although older explainers may use slightly different labels for transparency-related obligations, the final binding regulation is consistently described by EU institutions as a four-level risk framework. The EU AI Act's four risk levels The categories are not simply labels for how sophisticated an AI model is. They are a regulatory method for connecting an AI system's use and potential impact with corresponding obligations. A business cannot determine its position merely by calling a tool "low risk". It needs to assess the system against the Act's framework and the obligations associated with the applicable category. Risk level Regulatory position Core consequence Unacceptable risk Prohibited AI practices The practices are banned outright. High risk Systems subject to extensive obligations Requirements include conformity assessments and risk management. Limited risk Systems subject to certain requirements Transparency and oversight requirements apply in relevant cases. Minimal risk Most AI systems No additional sector-specific AI Act oblig

The problem I was running a multi-agent pipeline and one of my agents silently failed. The only alert I got said "daily loss limit reached" — completely misleading. The real cause was a missing file the agent never reported. I had zero visibility into what any agent had actually done. What I built AgentLens — a Python SDK for AI agent governance. Three modules: Audit trail — every LLM call and tool use logged to SQLite automatically Authorization — policy-based gates so agents can only call what you've approved Anomaly detection — baseline + threshold config, alerts when behavior drifts One-line integration Drop-in for Anthropic: python from agentlens.integrations.anthropic import TracedAnthropic client = TracedAnthropic(agent_id="my-agent") response = client.messages.create(...) # auto-traced

xAI's Grokipedia, an online encyclopedia with AI-generated articles that Elon Musk once promised would be a "massive improvement" over Wikipedia, apparently hasn't been updated since April 24th, according to a report from Lawfare. "As far as we can tell, no entry has changed in more than three months," Lawfare said. Grokipedia launched in v0.1 in […]

I keep seeing this same tension play out everywhere. TikTok, Instagram, X, all over the tech corners of the internet. It's the fight between software engineers and vibe coders, and tbh, I get both sides of it. Let me take the engineers' side first, because they're not wrong. If you spent four plus years learning to actually code, grinding the fundamentals, learning why the thing works and not just that it works, then yeah, I understand the frustration. Someone opens up Claude or ChatGPT, writes a prompt, ships their first app, and calls themselves a software engineer. And half the time, the second they hit a real problem, the whole thing falls over, because they don't actually know what's under the hood. I'd be a little annoyed too. And real talk, nothing replaces that depth. An engineer who can reach into the code, read it, and understand exactly what every line is doing is on a different level than someone vibe coding their way through. That's just true. But here's the part that sits weird with me. The problem isn't people using AI to build. It's when it turns into a wall. When the message becomes "you're not allowed in here, you don't get to build the thing in your head, because you didn't earn it the right way." That's the part I don't buy. I've watched this play out with my own friends. Engineers on one side, the ones just getting into vibe coding on the other, and there's this real contention between them. Almost a running joke about who counts and who doesn't. I think big ideas come first. The imagination comes first. Then you go find the resources, or the people, or the tools to actually build it. If someone has a huge idea and AI is the thing that finally lets them build it without waiting for permission, I don't see a problem. I see someone building. Gatekeeping who gets to make things never made much sense to me. You can respect the craft and still leave the door open. Those two aren't in conflict. Just my take.

Use a kill switch inside the checker when your uptime probes start amplifying an incident — one feature flag, read on every tick, that can disable the noisy checks and stop the retries at the source. Reach for tuned backoff and jitter instead when the retry storm stays inside a single process and never fans out onto a dependency somebody else is paging for. Both are cheap to build. Only one of them lets you quiet a polling client while its target is already on fire. I run cron and queue infrastructure, so most of my pages arrive as either "the job didn't run" or "the job ran four times." Health checking sits in the same family of problems: a small, frequent, automated request that multiplies badly when something upstream changes shape. What follows is the runbook I settled on after a fleet of pollers turned a non-incident into a real one — the failure mode, where the switch belongs, the implementation, and how to verify the flip before you walk away from the terminal. What actually turns a polling client's uptime checks into a retry storm? Amplification. A single check is one request every 15 or 30 seconds, which nobody notices; a fleet of checks with retries layered on top is a synchronized load generator pointed at whatever you decided was important enough to monitor. The math is unkind. Take 40 instances, a 5s interval, and 3 retries per failed attempt, and a dependency that normally handles a trickle of health traffic suddenly sees a couple thousand requests a minute — all of them arriving at the exact moment it's least able to absorb them. Retries stack on top of the polling interval rather than replacing it, and because every poller sees the same failure at the same time, they all back off together and return together. The Google SRE book calls out this shape under cascading failures, and the load pattern that comes out of it looks nothing like organic traffic: sawtooth spikes, perfectly aligned, growing until something sheds load. The worst one I've dealt wit

A publishing bot that depends on one LLM provider has a boring failure mode: the workflow is green, but nothing gets published. I hit that during cycle #1287. The dev.to key was present, the command was read, and the article module simply returned no action after generation failed with LLM unavailable . That is the kind of failure that looks harmless in CI and expensive in a content pipeline. The fix is not more optimism. The fix is a fallback path that produces a plain, useful, bounded article without calling another model. The Failure Mode Most automation code treats content generation and content publishing as one step. That is convenient until the generator fails after the scheduler, secrets, and publishing client have all done their jobs. Separate Generation From Delivery The publishing client should not care whether an article came from an LLM, a template, or a human-reviewed draft. Give it a strict article object and keep the fallback close to the generation boundary. Make the Fallback Honest A fallback article should not pretend it has fresh benchmarks, citations, or provider-specific pricing. It should explain the operational lesson in front of it. Key Takeaways Treat article generation and article publishing as separate failure domains. Return a fallback article when LLM generation fails instead of returning an empty action list. Keep fallback content honest: no invented benchmarks, prices, or citations. Record the original error type so a successful publish does not hide provider trouble. Prefer deterministic recovery for unattended workflows that are expected to produce public output. Next Steps This fallback article is a temporary solution. The long-term strategy is to: Implement a multi-LLM provider system that can switch automatically Add a quota monitoring dashboard to track usage across providers Create a content buffer that stores pre-generated articles for emergencies

Builder Journal · Mars Environmental Dynamics Analyzer (MEDA) Virtual Sensor Recovery Ten times in a row I predicted what my next submission would score before I uploaded it. The worst miss was 0.0025 on a number around nineteen. I took that as confirmation that the physics underneath was correct. It was confirmation that I can do arithmetic. Two days before this competition closed I pointed a review at my own endgame, expecting notes about the code. It came back with three errors and none of them were in the code. All three were in my reasoning, and all three had the same shape: I had run something that felt like a measurement and was not one. This is the fourth entry in this series and the one I would keep if I had to burn the other three. The models are competition-specific. This part is not. The competition in one breath Perseverance carries an environmental station called MEDA. Some of its surface pressure readings are missing, and the competition is to reconstruct them. Scored on mean squared error. The wrinkle is the split. Training covers sols 1 through 100, when pressure is climbing toward its seasonal peak. Test covers sols 201 through 300, when it is falling hard toward the aphelion minimum. Sols 101 through 200 do not exist in either file. Every prediction is outside the range the model was fit on. The first entry covers the first submission, which contained no machine learning at all and took the top of the board at 61.04. Six weeks and seven versions later the public score was 18.99. Almost everything in between was selected by one signal. Not cross-validation. Cross-validation here can only hold out sols from the rising limb, so it is structurally blind to the regime I am scored on. The leaderboard was the only thing that could see the falling limb, so the leaderboard picked every scalar that mattered: the residual shrink, the blend weight, a constant seasonal offset, a diurnal scaling. Hold onto that. It becomes the joke about four hundred words from

New OpenAI Signals data shows how people use ChatGPT worldwide, with country-level insights on adoption, usage trends, and evolving behavior.

For nearly two years, researcher Vangelis Stykas has maintained access to North Korean hackers’ servers. His work shows they pulled off intrusions in a shocking number of systems across the globe.

Nikita Bier was the face of some of the platform's biggest changes.

Baseboard management controllers from the world's biggest manufacturers are a security mess.

X head of product Nikita Bier is stepping down and says he will move into a role as an advisor, writing that "it's time to pass the torch and demote myself to my natural state: a poster." He shared the update just over a month after celebrating his one-year anniversary on the job, and just […]

TL;DR: LLMs charge per token, but control how tokens are generated. So the real skill isn’t prompting better—it’s constraining output to reduce entropy and cost. I. The Political Economy of Metered Latent Space In traditional public utility infrastructure, metered consumption follows a clear material logic: the unit of billing corresponds directly to a tangible, user-controlled commodity—gallons of water, kilowatt-hours of electricity, or therms of natural gas. While the provider owns the infrastructure and the meter, the user dictates the exact rate and volume of consumption required to accomplish a physical task. The modern cloud-based Artificial Intelligence (AI) ecosystem introduces a structural asymmetry into this model. Under prevailing API pricing and enterprise subscription frameworks, Western hyperscalers meter access to Large Language Models (LLMs) per token—covering both context input ingestion and payload output generation. Crucially, however, the platform retains operational control over how those tokens are selected, expanded, and emitted. This arrangement produces an alignment of incentives consistent with structural surplus capture, regardless of specific vendor intent. When platform revenue scales linearly with output generation volume, the system's economic environment selects for high-entropy conversational output—politeness markers, administrative hedging, corporate disclaimers, and redundant summaries. Conversely, zero-entropy symbolic execution yields minimal billable payload. The user thus incurs an emergent "conversational tax," where surplus tokens serve the economic logic of the host rather than the computational objective of the operator. II. Output Densities and Execution Constraints To understand how token economics intersect with model behavior, output payloads must be evaluated through information density and interface constraints rather than naive string tokenization. The Field-Array Operator Algebra (FAOA)—a proposed abstraction laye

🎯 What a “Stacked PR” Is (and Why You’ll Want One) A stacked pull request (sometimes called a stacked PR , stacked diff , or dependent PR ) is a series of PRs that build on top of each other, each one containing a small, logically‑isolated change. main ──► A ──► B ──► C │ │ │ │ │ └─ PR‑C (depends on B) │ └─ PR‑B (depends on A) └─ PR‑A (directly on main) A is based on main . B is based on A (its head). C is based on B , etc. When you eventually merge the stack in order (A → B → C), each change lands cleanly, and reviewers can focus on one cohesive piece at a time. Why Stack PRs? Problem Stacked PR Solution Huge, monolithic PRs that are hard to review & cause long CI times Break the work into bite‑size PRs (e.g., “feature flag”, “data model”, “UI”) Inter‑dependent changes (e.g., a new API + its consumer) Each dependent change lives in its own PR, but they still get tested together because they are built on top of each other Rebasing on main constantly drags in unrelated changes Only the bottom PR needs to be rebased onto main ; the rest stay on top of it Need to ship part of a larger change early Merge the first PR in the stack; the rest stay pending until they’re ready CI resources Only the bottom PR runs the full suite against main ; higher PRs can run a lighter subset because they already passed lower‑level tests 📦 The Landscape of Tools (as of 2026) Tool / Service Key Features Installation / Setup Typical Workflow ghstack (GitHub CLI plugin) - Creates stacked PRs automatically from a series of commits. - Handles base‑branch updates, resolves merge conflicts, and can re‑stack after rebases. - Works with GitHub's GraphQL API, so you get “dependent PR” links in the UI. pip install ghstack (or brew install ghstack ). Requires a personal access token with repo scope. bash git checkout -b feature/stacked\n# create many commits …\nghstack push\n# later, after rebasing on main\nghstack rebase . | | GitTown (aka git-town ) | - git town ship can ship a stack of dependent br

The serial entrepreneur is stepping down a little over a year after taking the "24/7 job" of overseeing X.

Claimable Clouds are temporary Cloudinary environments for AI workflows that let AI Agents safely manage media with no signup required. Imagine you're a busy designer, with many satisfied clients who depend on you to take their images and make them look great across social media. All that manual cropping and scaling, it's enough to make a body cry. On top of that, you know that AI can give you a hand here, but managing the handoff between your AI, your own skilled hands and artistic taste and style, and your always-in-a-hurry client list is another big pain. Enter the concept of the Cloudinary Claimable Cloud, just released today. Take a look at the docs about these new temporary instances available now What we built and why Provision a disposable Cloudinary cloud with no signup, using npx @cloudinary/cloud Auto-detect a dropped image and upload it to that temporary cloud Auto-crop it into 6+ social formats (Instagram, LinkedIn, X, Facebook, Stories) using AI-based smart cropping Generate a side-by-side gallery of results automatically Hand off a Claim URL so a client can make the cloud permanent Now, you can hand off the main pain points to AI - the resizing and reshaping of your images for the various social media platforms, while giving your clients a clean handoff via a temporary Cloud environment that they can use to create a Cloudinary account and start using these assets. One side effect: this also nudges your whole client base toward the same toolset - Cloudinary. The bigger deal is working with an AI agent that makes your life easier but ALSO allows you to keep control of the output. Let's walk through how this works! It all boils down to a new command: npx @cloudinary/cloud Type that into your terminal to kick off the process. I built a small app around this concept to provide this AI agent with a simple harness, so let me show how that looks. The user experience is to drop any image you want resized into the /drop folder. Under the cover, there are a few

Kalanick continues to get the band back together, after acquiring Anthony Levandowski's autonomy startup, and even soliciting investment from Uber itself.

If your automations are simple and low-volume, Zapier's per-task billing is fine and the cheapest thing about it is your time. The moment a single workflow fans out into many steps, or you start running thousands of runs a month, the pricing model — not the sticker price — is what decides your bill. Make charges per module execution, which is finer-grained than a Zapier task; n8n charges per workflow execution regardless of how many steps that workflow has, and it can be self-hosted for infrastructure cost only. The switch point is almost always about billing units, not features. I've run all three in production for internal automations, and the migrations I've done were never triggered by a missing feature. They were triggered by a monthly invoice that grew faster than the value of the work being automated. This post is about spotting that inflection before the invoice does. How does each tool actually count usage? The three tools use three different meters, and conflating them is where most cost surprises come from. Zapier bills per task. A task is one action step that successfully runs. The trigger that starts a Zap does not count; every action after it does. So a Zap that watches a form and does one thing costs one task per submission. A Zap that watches a form, looks up a record, formats a value, and writes to two places costs four tasks per submission. Filters and paths that stop early generally don't consume a task, which matters more than people expect. Make bills per operation. An operation is a single module doing a single unit of work. It's conceptually similar to a Zapier task, but Make's modules are more granular and the included volumes on comparable tiers are typically much higher, so the effective cost per unit of work tends to be lower. The catch is that iterators, aggregators, and array-processing modules can multiply operations fast — a scenario that loops over 50 items can spend 50+ operations in one run. n8n bills per execution. One workflow run

Meta expanded its AI coding offerings with a new agent that, it promises, can handle complex tasks with complex software.

The DOJ alleged that OpenAI did not meaningful attempt to hire U.S. citizens before seeking permanent residence for Visa-holding employees.