AI 资讯

French President Macron and Indian PM Modi raised alarms at the G7 summit that the U.S. could cut off access to American AI overnight — a fear the Anthropic blackout just made real.

In the previous series, when we optimized our neural network, we had to write quite a bit of training code ourselves. First, we created an optimizer object that used Stochastic Gradient Descent (SGD) to optimize final_bias . Then we wrote loops to calculate the derivatives required for gradient descent. We trained the model for up to 100 epochs . For each training example, we: Ran the input through the neural network to get a prediction. Calculated the loss. Calculated the derivatives of the loss function. After processing all three training points, we used: optimizer . step () to take a small step toward a better value for final_bias . Then we used: optimizer . zero_grad () to clear the accumulated gradients before starting the next epoch. All of this required a considerable amount of training code. Let's see how Lightning helps simplify this process. Organizing Training Logic with Lightning Previously, we created a class to store the weights, biases, and the forward() function. The optimization-related code was written separately outside the class. With Lightning, we can keep all of this logic in one place. We start by creating the class as usual, and then add a few new methods. Configuring the Optimizer The first method is configure_optimizers() . def configure_optimizers ( self ): return SGD ( self . parameters (), lr = self . learning_rate ) This method tells Lightning how the neural network should be optimized. The learning rate is stored in the self.learning_rate variable that we defined earlier. Defining a Training Step Next, we add a method called training_step() . def training_step ( self , batch , batch_idx ): input_i , label_i = batch output_i = self . forward ( input_i ) loss = ( output_i - label_i ) ** 2 return loss This method receives: A batch of training data from the DataLoader. The index of that batch. Inside the method, we: Extract the input and label from the batch. Run the input through the neural network. Calculate the loss using the squared r

The average knowledge worker switches between apps 1,200 times per day, according to a 2024 Harvard Business Review analysis. Each switch is small. The cumulative cost is not. For freelancers managing their own tool stack, the problem is both a productivity drain and a billing leak. What the Research Actually Says The most cited figure comes from Gloria Mark at the University of California, Irvine: it takes an average of 23 minutes and 15 seconds to fully refocus after an interruption. That number gets quoted a lot, but the context matters. Not every app switch is a full context switch. Checking Slack for two seconds is different from switching from deep coding work to a client call. A more useful framing comes from the American Psychological Association, which distinguishes between task switching (changing what you are working on) and tool switching (changing which app you are using for the same task). Both have costs, but tool switching is uniquely wasteful because it does not change the work -- only the interface. You are still working on the same problem but spending cognitive effort navigating a different app. For freelancers, the most expensive switches are the ones between a task manager and a time tracker, between a calendar and a task list, and between a project view and a communication tool. These happen multiple times per hour during active work, and each one breaks the low-level focus that produces billable output. How to Audit Your Current Tool Stack Before consolidating tools, figure out what you actually use. For one week, keep a simple log: every time you open an app to do work (not social media or entertainment), note it. At the end of the week, tally the list. Most freelancers find they use 6-10 tools daily. The typical list looks something like this: Task manager (Todoist, Asana, Notion) Time tracker (Toggl, Clockify, Harvest) Calendar (Google Calendar, Outlook) Communication (Slack, email) File storage (Google Drive, Dropbox) Invoicing (FreshBook

En una gestoría del automotor, consultar las multas de un auto era entrar a 33 sistemas distintos (Provincia, CABA, municipios), cada uno con su captcha y sus caídas. Lo automatizamos con una sola API, la de Multita , y comparto cómo quedó porque sirve a cualquiera que arme herramientas para el rubro automotor o fintech en Argentina. El problema Las infracciones de tránsito en Argentina no viven en un solo lugar. Hay sistemas provinciales (Buenos Aires, Santa Fe, Entre Ríos, Misiones, Chaco, Salta, Mendoza) y municipales (decenas). Ninguno habla con el otro. Consultar a mano son 15 a 20 minutos por vehículo. La solución: una request, todas las jurisdicciones La API de Multita recibe una patente, un DNI o un CUIT y devuelve, en JSON, las actas de cada jurisdicción con su monto y su estado. curl -X POST https://multita.com.ar/api \ -H "X-Api-Key: TU_KEY" \ -H "Content-Type: application/json" \ -d '{"tipo":"patente","valor":"AB123CD","jurisdicciones":"todas"}' { "resultados" : [ { "jurisdiccion" : "pba" , "nombre" : "Provincia de Buenos Aires" , "cantidad_actas" : 2 , "total_oficial" : 418500 }, { "jurisdiccion" : "caba" , "nombre" : "CABA" , "cantidad_actas" : 1 , "total_oficial" : 95000 } ], "resumen" : { "cantidad_actas" : 3 , "total_oficial" : 513500 } } Lo que nos ahorró Pasamos de 15-20 minutos por auto a segundos, y de cuatro ventanas abiertas a una sola llamada. Para una gestoría que cotiza decenas de carteras por día, es la diferencia entre atender 10 clientes o 30. Datos clave Cubre 33 jurisdicciones argentinas (provinciales y municipales), por patente (dominio), DNI o CUIT. Respuesta en JSON al instante; opcional, el total ya cotizado con tu pricing. Hay también una consulta web gratis para probar sin integrar nada. Si tenés una gestoría o estudio y querés esto andando sin programar, escribinos a BA Gestoría y te lo dejamos listo (y un descuento si venís de este post). Docs de la API: https://multita.com.ar/api

Here's the thing nobody warns you about when you put Supabase behind a "real" backend. My stack is React + FastAPI + Supabase Postgres. Every write goes through FastAPI. Every endpoint checks the user, the role, the ownership. I audited that backend HARD — rate limits, JWT validation, RLS, the whole thing. I was proud of it. And none of it mattered for the two holes I actually shipped. Because the Supabase anon key lives in the browser. It HAS to — that's how supabase-js talks to your project. Which means every logged-in user is holding a key that talks to Postgres directly . Not through my FastAPI. Around it. That anon key is a SECOND API. And I'd spent months hardening the first one while the second one sat there, wide open, the whole time. Hole #1 — the answers were just... readable Quiz questions live in quiz_options , one is_correct boolean per option. My backend never sends is_correct to a student before they submit. Obviously. But the browser doesn't have to ask my backend. // any logged-in student, straight from the console: const { data } = await supabase . from ( ' quiz_options ' ) . select ( ' question_id, label, is_correct ' ) // <- the answer key. all of it. The RLS policy said "authenticated users can read quiz_options ." Totally true for the rows. It just also handed back the column that decides the grade. The answer key. To anyone with a login and ten seconds of curiosity. Fix: column-level REVOKE SELECT from the client role, and let the backend be the only thing that ever reads is_correct . (PR #775.) Hole #2 — they could WRITE things they shouldn't Same class of bug, bigger blast radius. The default Postgres grants let the client role insert/update far more than I'd realized — including a path toward forging a certificate. Nobody did it. But "nobody did it yet" is not a security model! So I stopped patching table by table and flipped the whole thing: -- kill the client's entire write surface, then grant back the ONE thing it needs ALTER DEFAULT PRI

I use Neovim btw... Unfortunately, I'm also a C# developer. Anyone who has used .NET in Neovim knows that the language support is tenuous. I've spent hours tweaking my configuration to get the roslyn LSP to interact with my code editor, and I could never get the limited feature set that I wanted to feel smooth and native the way other LSPs do in Neovim... especially razor. That is... until now. A few weeks ago with this commit , the developers of the roslyn.nvim plugin made it so much easier to get up and running with the roslyn server that's officially distributed by Microsoft. Early this year, Microsoft put out a prerelease of a new dotnet CLI tool roslyn-language-server . Before this release, developers got access to the server by downloading the binaries from an obscure azure feed and pointing the Neovim plugin to the proper dll . Some adventurous and patient developers even got razor support in this method, but I was not one of them. With an officially (pre-) released tool from Microsoft, downloading roslyn is now as easy as running: dotnet tool install -g roslyn-language-server --prerelease After a month or two, the developers of roslyn.nvim updated the plugin to support the new roslyn-language-server command to setup the plugin. They even set the plugin to search for it on a fallback by default, so no configuration changes are required. My configuration for roslyn.nvim used to be about 20 lines of Lua. Here it is now: { 'seblyng/roslyn.nvim' , ---@module 'roslyn.config' ---@type RoslynNvimConfig opts = {}, }, It's Simple. It works. Large projects initialize relatively quickly, and all of the standard language server operations, like semantic highlighting, renaming, jump to definition/implementation, in-editor diagnostics, etc. are available out of the box. I'm not exactly sure what magic the developers pulled here, but as I've mentioned, I've never gotten razor syntax support. With the recent updates, razor is supported! I've been using this setup for about a

Anthropic has joined the Frontier coalition, which received another $915M in pledges to fund carbon removal projects.

Anthropic has spent much of this week fighting to get its newest AI models back online after the Trump administration abruptly ordered the company to cut access for all foreign nationals, including users inside the US and its own employees, forcing Anthropic to block access to Fable 5 and Mythos 5 for everyone. "To my […]

By Sergio Colque Ponce — Software Engineering, Universidad Privada de Tacna. Full source code: github.com/srg-cp/design-principles-go When people say "this code is well designed" , they rarely mean it has clever tricks. They usually mean it is easy to change . New requirements arrive every week, and good design is what lets you absorb them without rewriting half the project. In this article I take a small, very common requirement — "send a reminder to the user" — and I show how four classic design principles turn a fragile module into one that is open to change and easy to test. Everything is written in Go , and you can run it yourself from the repository linked above. The requirement We are building the backend of a bank appointment system. When an appointment is created, the user should get a reminder. Today it goes by email . Next month, product wants SMS too. After that, WhatsApp . The pattern is obvious: the list of channels will keep growing. A first (bad) attempt The fastest thing to write is one function that does everything: func SendReminder ( channel , recipient , body string ) error { if channel == "email" { // ... open SMTP, format the email, send it } else if channel == "sms" { // ... call the SMS provider } else if channel == "whatsapp" { // ... call the WhatsApp API } return nil } It works on Monday. But look at what it costs us: Every new channel means editing this function and risking the ones that already work. The function knows about SMTP, SMS providers and HTTP clients all at once: it has many reasons to change . To test the email path you need a real (or faked) SMTP server, because the logic is glued to the transport. This is the design we want to avoid. Let's fix it one principle at a time. 1. Single Responsibility Principle (SRP) A piece of code should have one reason to change . Instead of one function that knows every channel, we give each channel its own type that only knows how to deliver through that channel. Here is the email one: // E

By Sergio Colque Ponce — Software Engineering, Universidad Privada de Tacna. Full source code: github.com/srg-cp/spec-driven-development If you have used an AI coding agent — Copilot, Claude Code, Gemini CLI — you have probably lived this moment: you describe a feature, the agent produces code that compiles and looks right, and then it quietly does the wrong thing. The agent is not weak; the input was ambiguous. We have been treating coding agents like search engines when they behave more like very literal pair programmers. Spec-Driven Development (SDD) is the answer to that problem: instead of jumping straight to code, you write down what you want and why , refine it, and only then let the implementation follow. The specification — not the code — becomes the center of the project. What Spec-Driven Development actually is The idea is old (anyone who has written a Product Requirements Document will recognize it), but it has become practical again thanks to tools like GitHub's open-source Spec Kit . Spec Kit organizes the work into a small set of Markdown artifacts, each feeding the next: Constitution — the non-negotiable principles of the project (security rules, coding standards, architectural constraints). Spec — what you are building and why , with no implementation detail. Plan — the technical blueprint derived from the spec (stack, structure, decisions). Tasks — the plan broken into small, ordered, verifiable steps. Implement — the agent (or you) builds the tasks, with the previous artifacts as structured context. The workflow is usually summarized as Spec → Plan → Tasks → Implement , and the same process is meant to work regardless of language, framework, or which of the 30+ supported agents you use. The real shift is not "more documents." It is this: when requirements change, you update the spec, regenerate the plan, and let the implementation follow — instead of patching code and hoping the intent survives. The spec is a living artifact, not a dusty Word file

In Shenzhen, workers at IO-AI Tech control humanoid robots using a VR rig reminiscent of Ready Player One.

From Amazon's “Off Campus” to Netflix’s upcoming “Icebreakers,” the recent spate of hetero hockey romances shows Hollywood learned the wrong lessons from “Heated Rivalry.”

Your AI agent remembered the user's name. Then it forgot what it was doing. Here's the setup: User asks the agent: book the cheapest flight to NYC, search hotels under $150/night, then compare total trip cost. By step 3, the agent calls the LLM with 8,000 tokens of raw conversation history — and still answers as if it's turn 1. You need a memory architecture before this ships. Which one do you pick? A) In-context window only — full conversation stays in the system prompt. Simple. Breaks at ~15 turns or 8K tokens, whichever comes first. B) Vector memory store — embed past turns, retrieve the top-k by semantic similarity at query time. Works great until "NYC flight" pulls a memory about a past NYC trip instead of the current task. C) Episodic memory with summarization — compress old turns into structured event summaries, inject the relevant ones per request. More complex to build. Much harder to confuse. D) Redis session state — structured key-value store, explicit agent reads/writes. Deterministic. Requires the agent to know what to store and when. One of these collapses past 15 turns. One retrieves the wrong context at exactly the wrong moment. One is the right answer for task-oriented agents. Pick A, B, C, or D — and tell me where you've hit this in production. Full breakdown in the comments.

How GPU scheduling complexity and MLOps integration are forcing platform teams to rearchitect Kubernetes clusters before operational debt becomes insurmountable. As AI workloads consume roughly 40% of enterprise Kubernetes clusters by 2026, the platform's default scheduler is proving fundamentally mismatched with the topology-aware, gang-scheduled demands of GPU-intensive training and inference. Platform engineering teams that invest now in purpose-built GPU scheduling layers, multi-tenant partitioning, and FinOps-driven autoscaling will separate themselves from organizations drowning in 30-45% GPU utilization rates and mounting infrastructure costs. Why the Default Kubernetes Scheduler Fails GPU Workloads Kubernetes was designed for stateless, CPU-bound services, and its pod-by-pod bin-packing scheduler has no native awareness of GPU topology, NUMA boundaries, or NVLink interconnect bandwidth. This becomes a critical failure point with NVIDIA H100 SXM5 nodes, where achieving full-bandwidth tensor parallelism requires all 8 GPUs on a node to be scheduled as a single atomic unit. The default scheduler cannot guarantee this co-placement, meaning distributed PyTorch FSDP or MPI training jobs frequently land on suboptimal node configurations, wasting expensive NVLink bandwidth and forcing teams to over-provision GPU capacity. Idle GPU memory stranded across partially-utilized nodes is the primary driver behind the 30-45% utilization rates reported in 2025 surveys by Gradient Dissent and Weights and Biases, representing millions of dollars in annual wasted spend for mid-to-large enterprises running mixed AI workloads. Building the GPU Scheduling Stack: Volcano, KAI Scheduler, and MIG Platform teams are converging on a layered scheduling architecture that replaces or augments the default Kubernetes scheduler with GPU-aware primitives. Volcano has become the dominant choice for distributed training workloads, using its PodGroup abstraction to enforce gang scheduling across

A common way to design an application is to begin with its nouns: User Product Order Payment Then each noun receives the standard architectural starter pack: UserController UserService UserRepository The controller receives users, the service services them, and the repository stores them somewhere responsible. This is noun-oriented architecture : treating every important thing in the domain as if it were automatically a useful software boundary. It works for simple CRUD systems. Unfortunately, most applications eventually do something. The noun becomes a drawer Consider a typical UserService : register() findByEmail() resetPassword() changeAddress() disableAccount() mergeAccounts() assignRole() calculateDiscount() These operations all involve a user. That is approximately where their similarity ends. They have different rules, dependencies, side effects, security concerns, owners, and reasons to change. They live together because User was the nearest available noun when the folders were created. As more behaviour accumulates, UserService becomes the official location for anything vaguely user-shaped. Other components depend on it. It gradually depends on authentication, email, permissions, billing, auditing, and several services added during incidents nobody wishes to revisit. The noun becomes both a dependency of everything and a consumer of everything. The folder remains impressively tidy. Name the capability, not the material A better starting question is not: What things exist in this system? It is: What must this system be capable of doing? That leads to components such as: UserRegistrar PasswordResetter AccountMerger OrderPlacer PaymentRefunder SubscriptionCanceller These are agentive names . They name the component responsible for performing a capability. Compare: UserService with: PasswordResetter UserService tells us which noun is nearby. PasswordResetter tells us what the component is for. That difference produces better architectural questions: What rules

The build-vs-buy question gets answered wrong in both directions. Scrappy teams build things they should have bought, wasting six months reinventing Stripe. Enterprise teams buy things they should have built, ending up with a duct-taped stack of ten SaaS products that cost more than a full-stack engineer. The real answer depends on five questions most decision frameworks don't ask. This guide is a practical walkthrough for anyone trying to figure out the right call for their own business. The Myth That Distorts Every Build-vs-Buy Conversation "Buying is cheaper." This is the default assumption, and it's wrong often enough to be dangerous. Buying looks cheaper because the cost is monthly instead of upfront -- a psychological trick, not an economic one. Run the numbers on any SaaS tool over 5 years and you'll usually find the cost lands within 2x of building custom. Sometimes below. The real cost difference is not price; it's time, flexibility, and ownership. When you buy: You spend less today, more in year 3 You get speed now, rigidity later You trade money for control You own none of the code When you build: You spend more today, less per year You trade speed now for flexibility later You trade money for control You own the code and can change anything Both are rational trades. The question is which one matches the stage and strategy of your business. When Buying Wins Start with the easy case. Buy off-the-shelf when: 1. The problem is generic and solved. Email hosting, payment processing, accounting, HR payroll, customer support tickets, video conferencing, file storage. These are solved problems. Building your own is nearly always the wrong call. 2. The space has mature competitive options. If there are 5 reputable companies competing on price and features, you benefit from that competition. Building custom takes you out of it. 3. Your process is standard. If you do exactly what every other company in your vertical does, a tool built for every company in your verti

Last month my OpenClaw agent kept making the same mistake: it would run a health check, the script would fail silently, and the agent would report "all systems operational" with total confidence. It wasn't broken. It was just doing what it was built to do — execute tasks — without any mechanism to learn from the outcome. So I built it a self-improvement loop. Every night at 2 AM, an isolated OpenClaw session wakes up, reads the previous day's execution logs, identifies patterns in what went wrong, and updates the agent's memory files. No human in the loop. No re-deployment. Just... learning. Here's what I built, what broke, and what actually works. Why Self-Improvement Is Hard for Personal Agents Enterprise AI labs solve this with massive infrastructure: reinforcement learning pipelines, full fine-tuning jobs, A/B testing frameworks that run for weeks. For a personal agent running on a cron job, that's not an option. The self-improvement loop for a personal OpenClaw setup has to be lightweight. It has to run in seconds, not hours. It has to write to plain text files that the next session will actually read. And critically, it has to avoid the feedback loop problem — an agent that rewrites its own improvement logic can spiral into nonsense if there's no anchor. The key architectural decision I made: separate the executor from the critic . Your main agent runs tasks. A separate isolated session reviews what happened and recommends changes. The main agent applies them on the next run. No single session is both judge and executioner. The Nightly Cron: What Actually Runs This is the cron I have running at 2 AM ET every morning: { "name" : "nightly-self-improvement" , "schedule" : { "kind" : "cron" , "expr" : "0 2 * * *" , "tz" : "America/New_York" }, "sessionTarget" : "isolated" , "payload" : { "kind" : "agentTurn" , "message" : "Review the last 24 hours of OpenClaw execution. Read memory/$(date +%Y-%m-%d).md and memory/yesterday.md. Identify 3 patterns where the agent u

I am maintaining an open-source project (deterministic open source duplicate-code detector) and a user asked for a feature which I don’t have a clear answer on how to implement. This seems a very hard problem to solve: -Tests repeat the same scenario. For a structural detector, this flags as repetition (duplication). However, tests are not something people want to delete from the codebases. -The repetitions from tests (on purpose) end up looking like undesired code duplication and the tools cann

AI models with advanced hacking capabilities will soon be the norm.

World models are the next big thing in AI beyond LLMs and, with this round, Odyssey has cemented itself as one of the startups to watch.