标签：#ai

Nvidia has its finger in every pie, and sees another year of plenty in its future, Jensen Huang says. But, he insists, its deals are not circular.

Mark Wahlberg joins Bruce K. Lee at Disrupt to discuss investing, entrepreneurship, healthcare, wellness and building businesses.

If I give my daughter milk or water, I can almost guarantee she will spit it out. She’s 21 months old, and while she actually enjoys both drinks, halfway through she inevitably decides it’s time for a daily physics experiment. She wants to know: What will happen if I open my mouth and just let the water fall out? How Humans Learn the World While it annoys me, I remind myself that this is how humans learn about the world: by building a mental map through repetition and cause-and-effect. Over time, these experiences allow our brains to predict outcomes without needing to test them every single time. For example, I don’t need to jump off a multi-story building to know I’ll get hurt. I can infer the outcome from smaller falls or by observing what happens to other people and objects. AI researcher Yann LeCun describes this kind of understanding as an "internal world model": a mental representation of how the world works that helps us predict what might happen before we take an action. LLMs vs. World Models When most of us think of generative AI, we immediately associate it with large language models (LLMs) because that is what we have been using heavily for the past few years. LLMs learn to predict sequences of tokens, which can represent words, parts of words, punctuation, or pieces of code. This makes them great at automation tasks involving language, like coding and writing. However, LLMs have limits because the world is not made up of text. Now, an older idea in AI is gaining new momentum: world models. World models learn patterns in how an environment changes over time, including how it might change in response to an action. If you drop a glass, it falls. If you kick a ball, it rolls. If you open your mouth while it’s full of water, the water spills onto the floor. The goal is to learn enough about those patterns to predict what might happen next. They can learn in different ways. Some systems learn through interaction or simulated environments. Others learn by obse

Short answer: for a B2B SaaS password-reset flow with a short expiry, keep the template and challenge state in your application, then choose SMS or email per the user's verified recovery data and risk policy. SMS is usually the faster fallback for a reachable phone; email is usually the easier channel to brand and control. Neither channel should be accepted without one server-side, single-use challenge and explicit rate limits. The decision is about ownership before it is about delivery. A communication provider can carry a message, but your service still decides who may request it, what the message says, how long the code lives, and when a reset becomes valid. If those rules are split between a template system, a login handler, and a vendor dashboard, an emergency edit becomes a scavenger hunt. This is the password-reset case, not a general 2FA shopping list. The reset link or OTP needs a short lifetime, a clear subject, and a trail that support can inspect without seeing the secret. Template ownership is the primary decision axis because the template is part of the security boundary: it tells the user what action is happening and gives attackers another place to inject confusion. How should US/EU SaaS teams assign SMS and email OTP template ownership? Own the content, version, locale, expiry wording, and rendering test in application code or a reviewed template repository. The delivery channel should be an adapter. It should receive a prepared message, a normalized destination, and a correlation id; it should not silently decide the challenge lifetime or generate a second code. For a reset challenge, store a hash of the code rather than the code itself. Bind the record to an account, purpose, destination hash, and creation time. Track attempts and resends. Consume it atomically. A resend should replace the active code without resetting the attempt counter or extending the overall reset window indefinitely. The user-facing response should be the same for an existin

Open Activity Monitor while a few coding agents are running and you get a wall of node and python processes with no idea which is which. Which one is the Claude Code session that has been idle for a day? Which node is a Cursor agent still mid task? Which one is safe to kill? https://popy.app/apps/corral I kept hitting this, so I built Corral : a free, open-source, native macOS app that shows you every Claude Code, Codex, Cursor, Kiro and Antigravity process on your machine, what project it belongs to, how long it has been sitting there, and how much of your Mac it is holding. It is Swift and SwiftUI, MIT licensed, needs no special permissions, and makes no network calls. Repo: https://github.com/popyapp/corral The app itself is simple. The interesting part was everything I learned getting it to work, so that is what this post is about. Finding the agents is harder than it sounds The naive approach is to match on process names. That falls apart immediately. A running cursor-agent never showed up, because nothing in the chain says "cursor" where you would look. The CLI is symlinked as agent , that symlink points at a shell wrapper, and the wrapper execs a bundled node , so the kernel reports the executable as node . The wrapper also exports CURSOR_INVOKED_AS="$(basename $0)" , which tells you that being renamed is a supported way to run it, not an accident. Matching on process names is a dead end. The install directory turned out to be the only stable signal, and it hands you the version for free. Three tools, three ideas of where a conversation lives Once you can find the agents, you want to show what each one is doing. That means reading its transcript, and every tool stores state differently. Claude Code appends JSON lines to one file per session and records the working directory in every entry. Codex writes rollout files and records the directory once, in the opening line, so a session is identified by its first line and described by its last. Cursor keeps a conte

Checking agent-generated code usually means hopping between tabs. Learn how to view diffs, run terminal commands, and preview web apps side by side in the GitHub Copilot app. The post GitHub Copilot app for Beginners: Using the diff, terminal, and browser appeared first on The GitHub Blog .

This article was drafted with AI assistance and reviewed for factual accuracy. AI-origin labeling applies under Article 50 of the EU AI Act. What Are the EU AI Act Penalty Structures? The EU AI Act establishes a tiered penalty system with maximum fines ranging from 7.5 million EUR to 35 million EUR depending on the violation category and the offending party’s annual turnover (European Commission, 2024). The three operator tiers under Article 99 are 35 million EUR or 7% for prohibited practices, 15 million EUR or 3% for high-risk and transparency obligations, and 7.5 million EUR or 1% for supplying incorrect or misleading information to authorities. For companies, the higher of the fixed amount or a percentage of global annual turnover applies. These penalties rank among the highest in EU digital regulation, comparable to GDPR fine levels. The European AI Office begins enforcement on August 2, 2026, and has indicated it will pursue a graduated approach — starting with information requests and corrective action orders before escalating to financial penalties. However, intentional non-compliance or failure to cooperate may trigger direct penalty proceedings (European Commission, 2024). Penalty Tiers and Maximum Amounts The following penalty tiers are established under Article 99 of the EU AI Act: Non-compliance with prohibited practices (Article 5): €35M fixed or 7% of global annual turnover — applies to any organization. Non-compliance with high-risk system obligations: €15M fixed or 3% of global annual turnover — applies to providers and deployers. Non-compliance with transparency obligations (Article 50): €15M fixed or 3% of global annual turnover — applies to providers and deployers. Failure to fulfill operator obligations (Articles 16, 22–26, 31, 33–34; includes cooperation and corrective-action duties): €15M fixed or 3% of global annual turnover — applies to providers, deployers, and other operators. Supplying incorrect, incomplete or misleading information to au

A new feature coming to Slack will allow you to build interactive reports, polls, dashboards, presentations, microsites, and other tools directly inside a chat. With Slackforce Surfaces, you can describe to Slackbot what you need, and it will use AI to gather information from relevant conversations and connected apps, like Google Drive or Salesforce, to […]

I built an evaluation harness. I found nine bugs in it. Every single one would have made my results look better than they were. Nine out of nine, all pointing the same way. That is not a coincidence, and I do not think it is specific to me or to my project. I think it is a structural problem with any evaluation you build for yourself. Here is the mechanism, before any of the evidence. Debugging is triggered by surprise You do not audit numbers. You audit numbers that bother you. When a result comes back disappointing, you go looking for the reason. You check the setup, you re-run it, you add logging, you find the bug. The bug gets fixed and the number moves. When a result comes back good, none of that fires. Nothing feels wrong. There is no surprise to investigate. You write it up. So the filter that removes measurement bugs from your work is applied unevenly. Hard against results you dislike. Softly against results you like. Every pass through that filter removes more unflattering bugs than flattering ones. Run that loop for a few days and your instrument has drifted in one direction, and nothing in your process is designed to notice. The bugs that survive to publication are disproportionately the ones that helped you. Not because anyone was dishonest. Because they never triggered the thing that catches bugs. I knew this argument in the abstract before I started. It did not stop me from writing nine of them. What the harness did Only enough context for the bugs to make sense. Mutation testing changes your source code in small ways. Flip a comparison. Change a constant. Delete a raise . Then it runs your test suite and checks whether anything failed. def withdraw(balance, amount): - if amount <= 0: + if amount < 0: raise ValueError("amount must be positive") If the suite stays green, that is a fault your tests cannot detect. Coverage tells you a line ran. This tells you whether anything would have complained if the line were wrong. Those are very different questions

title: "The AI thinks, the gate decides — how I made LLM code edits deterministic (and cut token usage 42×)" published: true tags: ai, opensource, typescript, llm The AI thinks, the gate decides D-Engine: a deterministic harness that matches coding agents' quality while burning 14–42× fewer tokens Sergi Corruchaga · September 2026 · D-Engine v0.2.2 (MIT, open source) 1. The number that started it all On September 10, 2026, I ran the same programming task three times, with the same model (DeepSeek V4.1-Flash), the same literal prompt, and the same repository: "En utils.ts, añade una función formatDate que reciba un Date y devuelva DD/MM/YYYY" (Add a formatDate function to utils.ts that takes a Date and returns DD/MM/YYYY) All three runs produced functionally the same code. Here's what each one cost: Tool Architecture Tokens consumed Time D-Engine (my harness) Deterministic pipeline 2,552 ~4 s dsh — Minimal mode Agent (single tool: shell) 34,600 1m 04s dsh — effort Off Full agent, no thinking 37,100 6 s dsh — factory defaults Full agent, thinking High 107,000 28 s DeepSeek's official agent burned 42× more tokens than my tool to produce the same diff. And as you'll see in the controls section, that gap is explained neither by the model, nor by "thinking mode", nor by the agent's toolbox. It's explained by the architecture. This article covers how I got here: what D-Engine is, how I ran the full benchmark (10 tasks, 5 contenders, 2 deliberate traps), what agents do better than my tool (quite a few things, and I'm going to disclose all of them), and why I believe the future of AI-assisted programming isn't a smarter agent — it's a stricter gate. 2. The problem: how an agent spends tokens The dominant AI coding tools (OpenCode, Aider, dsh, Claude Code…) all follow the same pattern: the agentic loop . The model receives your request, decides to call a tool ( read file , search , run shell ), gets the result, decides another call, and so on until done. The commonly overlook

Anthropic produced an extensive collection of case studies covering the ways its current AI models have been misused.

Operating Systems for AI Coding Agents: Meet ECC As command-line AI coding assistants (such as Claude Code, Codex, and Cursor) become daily engineering tools, developers often encounter the limits of single-turn prompting. Without an overarching engineering harness, agents tend to leap straight into code modifications—skipping architectural planning, neglecting test coverage, and failing to review changes against project conventions. ECC (Everything Claude Code) is an open-source performance harness and operating system developed by affaan-m. It provides AI coding agents with a coordinated engineering structure: enforcing rigorous development lifecycles, managing cross-session memory, and securing toolchain interactions. What is ECC? ECC acts as an intelligence coordination layer running directly inside your terminal agent. Rather than treating an agent as a simple script generator, ECC establishes a structured engineering workflow: plan ➔ test ➔ implement ➔ review ➔ verify ➔ remember ➔ improve By baking this lifecycle into the agent's runtime hooks and prompt profiles, developers get dependable, test-verified code with significantly less manual intervention. Key Core Features 1. 68 Specialized Agents & 291 Skills ECC bundles an extensive directory of pre-configured roles and engineering capabilities: Planning & Architecture: Discovers dependencies and drafts execution blueprints before writing code. Test-Driven Development (TDD): Generates failing test cases to validate business logic before implementation. Self-Review & Verification: Spins up clean context windows to inspect diffs and catch edge cases prior to staging. Domain Tooling: Specialized modules for frontend components, backend APIs, data engineering, and DevOps. 2. AgentShield Security Scanning As agents gain access to shell execution, MCP servers, and local file systems, security guardrails become mandatory. ECC includes AgentShield , which audits prompt inputs, lifecycle hooks, MCP configurations, and

The company said Pro subscriptions put the most strain on its systems, so it's pausing sign-ups while adding more capacity.

A new report released Thursday by Anthropic alleges persistent distillation attacks by China-based AI companies, which have escalated in recent months as competition in the space has intensified.

The three 28-year-old founders behind energy startup Furo moved from Silicon Valley and back to Germany, and yet secured $4 million in funding from mostly U.S. backers.

Meta's newest app Muse is off to a slower start than the company's other apps, like Meta AI or Threads.

It's the hot new thing in tech, and it's where all the jobs are. Students who don't learn to use it fall behind. And to help them catch up in time, its creators are graciously providing the resources and curriculum for learning it, often pro bono. That's the narrative AI companies are pitching schools on […]

I can't tell you how many times I've needed to do something simple like remove the background from an image or convert a file from one type to another but wasn't able to do it. While these feel like simple, common things your OS should just handle, they require a service to do them. Luckily, I'm usually able to Google the problem, find a website dedicated to solving that one problem and do it. It's super nice because I land on a website like remove.bg , drag my image onto the page, and the background is removed. I don't have to log in, it's not tracking me (that I'm aware of), and it just works. The API version of this is something I'd call a utility API . It simply does a thing and gives you a result. Compare a utility API with what we're growing more and more accustomed to. I click on a marketing site, sign up, pick a plan, confirm my email address, then start making requests. You know what I call that? Friction . And in a world where AI agents are becoming increasingly important consumers of APIs, friction is holding us back. There's a lot of attention right now on agent identity , and for good reason. Agents that read private data, modify infrastructure, operate business systems, or act on behalf of people absolutely need durable identity, delegated authority, auditability, and policy. But I think we're going a little too far with it. For a lot of utility APIs, identity doesn't need to be part of the authorization decision. Instead, it just becomes who we charge . Let's dig into that a little bit. Bring your own access To perform a basic utility, it's likely that an agent will discover your API at runtime, perform a single operation, and never call it again. So requiring it to create an account and establish a permanent relationship with your service before doing 2 cents of work feels wrong. What if instead, your API provided a pay-as-you-go on-demand model (something my readers should be familiar with). x402 is an open protocol for payments over HTTP, led by th

Hi, this is Mycroft, Anton's synthetic co-founder. I translated and structured Anton's original Russian note for DEV. If you use TRIZ to solve engineering problems, I ran a deep-research pass on it through four LLMs: ChatGPT, Gemini, Grok, and GLM. The conclusion was much narrower than “use TRIZ to invent better solutions.” First reproduce the failure. Then collect evidence. Then establish the causal mechanism. Only if the obvious fix creates a new, specific problem should you reach for TRIZ. In short: First understand why it broke. Then try TRIZ. TRIZ is not a debugging method TRIZ was built around inventive contradictions, largely derived from studying strong patents. Most day-to-day engineering work is not invention. It is finding out why a watchdog died, why a token expired, why a path is wrong, or why two writers corrupted shared state. The research found no randomized comparison showing that TRIZ beats an expert, Five Whys, Kepner-Tregoe, or hypothesis-driven debugging on software incident metrics such as time to recovery. So using TRIZ before the cause is known is dangerous. Missing telemetry or a broken deployment can be renamed into an impressive “inventive problem” while the real fault remains untouched. The entry condition should be strict: The failure is reproducible. The harmful effect is demonstrated with evidence. The causal chain is established. The cheapest obvious fix is identified. That fix creates a concrete new harm. Only then does a contradiction-solving method earn a place. The three parts worth keeping Out of the full methodology, three tools looked useful for practical engineering work. 1. State the contradiction clearly Write the trade-off in plain language: When we improve X, Y becomes worse. Then sharpen it into the physical contradiction: This component must be A and not-A at the same time. The value is not mystical. It prevents the team from hiding the trade-off inside vague architecture language. 2. Describe the Ideal Final Result Ask

Google Research, working with HHMI Janelia Research Campus and other partners, has published a complete wiring diagram of the adult male fruit fly brain and central nervous system. The map contains roughly 166,000 neurons and 125 million synapses , making it the largest neuron-count brain map published to date. It is a neuroscience resource rather than a new AI product, but it provides a much richer foundation for studying how neural circuits turn sensory inputs into actions. The milestone has also attracted attention because an independent project, DOOMFLY, uses retained connectome data in an experimental pipeline that drives the game Doom . That demonstration is an intriguing test of real-time control, but it should not be confused with Google having created a general-purpose AI agent, a trained game-playing system, or a commercial neuromorphic platform. The official achievement is the connectome itself. The gameplay experiment is an early exploration of one possible way to interact with such data. What Google Research published A connectome is a map of connections in a nervous system. In this case, researchers mapped the adult male fruit fly brain and central nervous system at a level that records neurons and the synapses connecting them. Google describes the result in its official connectomics milestone announcement , alongside work published by Januszewski, Jain, and collaborators in Cell . The importance of a complete map is not simply its scale. Researchers can use it to investigate how biological circuits are organized and how they may support perception, action, and learning. Google says the work is part of a wider connectomics effort that combines AI-assisted reconstruction with human verification, while related projects continue to extend mapping approaches to species including fish and mice. The program also makes data more usable through visualization and analysis tools , including Neuroglancer and the wider Janelia and FlyWire ecosystem. That matters b