AI 资讯

For about 3 months I've been running an autonomous agent — one that thinks up and writes its own social media posts and comments — unattended, 4 sessions daily, on a 16GB M1 Mac with small models in the 9B / E4B class. I'm about to publish what that operation taught me about hardening, as a series of 4 technical articles. Before that, there's one thing I want to write down first: why small models . I've been to the purchase page for a Mac Studio or a new MacBook Pro more than once or twice. Backing the agent with a large cloud model (Opus or the GPT family) has always been an option in the code. And yet I haven't bought, and I haven't switched. The 16GB M1 is not an economic constraint — it's a constraint I chose . From the outside, building on small models looks like a cheap compromise. This article explains why it isn't, and states where I stand. It also serves as the hub for the 4-article series. A model's intelligence hides the roughness of your design Large models absorb sloppy prompts, ambiguous instructions, and missing guards with sheer intelligence. If all you want is to ship a product, that's a virtue. But if you want to become someone who can build things , it becomes a defect. Because inside the thing that worked, you can no longer tell where your design ends and the model's intelligence begins. "It worked" and "I built it" are different things. Something you bludgeoned into working with model capability counts as a thing that ran — it doesn't become the ability to build. Small models have no absorption capacity. So every design flaw comes to the surface. In my operation, all of the following surfaced: The context window being silently truncated Outputs cut off midway A runaway caused by one missing sampling parameter In cloud or large-model environments, these rarely bother you. The environment has cushioning built in. Context windows are in the 200K–1M token class, so truncation itself rarely happens. And when you do exceed the limit, you get an explic

For the last few months, I’ve been obsessed with a specific problem: the friction between privacy and utility in modern AI tools. Most "private" AI solutions still rely on a local LLM running on your CPU or GPU via a heavy desktop application. They require installation, constant background processes, and often struggle with performance on older hardware. I wanted to see if we could do better. I wanted to see if we could run a capable language model entirely within the browser, using only the device’s hardware acceleration, with zero data leaving the machine. The result is PrivateScribe, a tool I built to handle note summarization, email drafting, and rewriting. But more importantly, it’s an experiment in what’s possible when you treat the browser not just as a display layer, but as a compute engine. The Wedge: WebGPU and True Offline The core constraint that drove this project was simple: nothing leaves the device. In the current landscape, "on-device AI" often means "installed on your device." This is fine for desktop apps, but it creates silos. You can’t easily share a workflow across a Chromebook, a Windows machine, and an iPad without installing three different native applications. By leveraging WebGPU, PrivateScribe runs entirely in the browser. This unlocks a few critical advantages: Zero Installation: Users open a URL and start working. No downloads, no permission dialogs for file system access beyond what’s needed for the session. Hardware Acceleration: WebGPU allows the browser to tap directly into the GPU. This is crucial for inference speed. A small model that runs in your browser can process text significantly faster than a CPU-bound implementation, especially on modern laptops with integrated graphics. True Offline Capability: Because the model weights are loaded locally via WebAssembly and the inference happens on-device, the app works completely offline. If you lose your internet connection in the middle of drafting an email, the AI doesn’t stop. It c

If you ask me what part of the system I protect the most, the answer is the database. I've been writing software alone for twenty-four years, and across every platform I've built, the rule has stayed the same: the web servers can take whatever you throw at them, the batches can be rebuilt, but the database has to stay idle on purpose. Not because I love idle databases, but because the day a database actually starts to struggle is a day with very few good options. This article is about what "keep the database idle on purpose" actually means in practice, and about one particular kind of graph that, in my experience, almost nobody is watching. The three layers and what each of them gets I think of a production system as having three tiers, and each tier gets a different rule. The web server tier can be horizontally scaled. If load grows, you add machines. If something is wrong, you take a machine out of the pool, and the others handle it. Failures here are visible immediately, and they're cheap to recover from. The batch server tier can be scaled up or out depending on the work. A batch that's too slow can be split. A batch that crashes can be retried. End users don't see batch servers, so a stuck batch is a problem for me and not for them. Some headroom up here is fine. The database tier is the one I treat completely differently. The database is not where you absorb load. The database is what you protect from load. The reason is simple: the other tiers can be rebuilt or re-scaled. The database is the irreplaceable record. If it slows down, everything slows down. If it falls over, you don't have many minutes before the rest of the stack notices. So my rule for the database is: keep it idle. Not idle in the sense of "doing nothing." Idle in the sense of "running well below its capacity, at all times, so that any extra load it picks up has somewhere to go." For more than a decade I ran a large appliance-grade database where I kept the load average below 1 at all times. N

Builder Journal · ARC Prize 2026 There is a specific horror in a detective story when you realize the witness everyone trusted has been lying, or just wrong, the whole time, and every conclusion built on their testimony has to come down with them. I had that moment with my own notes this month. The unreliable witness was me. Context, if you are new to this thread : I'm competing in the ARC Prize 2026, building an agent that has to win games it has never seen. It had been stuck, underperforming on the hidden test in a way I could see on the scoreboard but could not explain, and I had been hunting the cause across several sessions. The two comforting facts In two earlier work sessions I had written down, as settled conclusions, two things about why the agent was failing. One: the failure was a kind that only happens on the hidden online games, so it could not be taken apart and studied on my own machine. Two: the practice games I did have were useless for investigating it anyway, because they scored a flat zero on the relevant measure. Notice what those two beliefs do when you put them together. They say, in a calm and reasonable voice, that there is nothing to be done here. The problem is unreachable, the practice data is a dead end, the smart move is to spend your energy elsewhere. They were not just facts. They were permission to stop looking. So I stopped looking. Twice. The hour that knocked it all down Eventually I made myself do the one thing I had been quietly avoiding. Instead of rereading my own notes for the third time, I went and checked. I wrote small probes and ran them against the real artifacts, the actual code and the actual game data, rather than against my memory of what they did. Both beliefs collapsed inside an hour. The failure was not unreachable. It came apart cleanly, deterministically, on the games I already had sitting on my disk. And the "dead end" practice data was not a dead end at all. It showed the problem plainly the moment I asked it

Your monthly OpenAI or Anthropic invoice tells you how much you spent. It doesn't tell you which feature spent it, which model, or why last Tuesday cost three times as much as Monday. So at some point you (or your team) will build a metering layer: wrap the client, read usage off the response, multiply by a price table, ship it to a database. I did exactly that over the past few months while building an LLM observability service, and my numbers were wrong in five different ways before they were right. Every one of these failures was silent. No exception, no alert, just numbers that were quietly too low or too high. This is the list I wish someone had handed me. Pitfall 1: Streaming responses quietly report zero tokens OpenAI's Chat Completions API returns no usage data at all for streaming requests unless you pass stream_options: { include_usage: true } . No error, no warning. The stream just never contains token counts. If your metering reads usage off the chunks, every streaming call gets recorded as 0 tokens, $0. And since chat UIs are almost always streaming, that's most of your traffic. This one bit me twice in the same audit. First finding: all streaming calls in my own dashboard were $0. Second, nastier finding: I had a budget-gate feature that blocks calls once spend crosses a limit, and it waved every streaming call straight through — because as far as it could tell, streaming was free. The fix is to inject the option in your wrapper when the caller didn't set it: let injected = false ; if ( params . stream && params . stream_options ?. include_usage === undefined ) { params = { ... params , stream_options : { ... params . stream_options , include_usage : true }, }; injected = true ; } But there's a trap inside the fix. With include_usage on, OpenAI appends one extra chunk at the end of the stream that carries usage and has an empty choices array . Any downstream code that does chunk.choices[0].delta — which is most example code on the internet — will throw

Hackathons are intense. You've got a brilliant idea, a tight deadline, and often, limited sleep. The last thing you want is to spend half your precious time wrestling with database boilerplate, ORM setup, or SQL query syntax. This guide will walk you through building a functional Node.js backend for your hackathon project, focusing on speed and minimal friction, so you can spend more time on your core idea. The Hackathon Backend Challenge Typically, setting up a database and its interaction layer involves several steps: Schema Definition: Deciding on tables/collections, fields, types, and relationships. ORM/Driver Setup: Installing and configuring your database driver or ORM (e.g., Mongoose, Sequelize). Model Creation: Translating your schema into code, often with verbose syntax. Query Writing: Crafting SELECT , INSERT , UPDATE , DELETE statements or ORM methods for every data operation. Debugging: Fixing typos, schema mismatches, and complex join logic. This process, while fundamental, eats up valuable time that could be spent on features, UI, or even sleep. For a hackathon, you need to iterate rapidly, and database interactions should be the least of your worries. Strategy 1: Embrace Simplicity For many hackathon projects, you don't need highly optimized, production-grade queries from day one. You need functional queries that work quickly. Focus on getting data in and out reliably. Strategy 2: Natural Language for Data Modeling Instead of writing verbose schema definitions, think about how you'd describe your data to a non-technical person. For example, if you're building a task management app, you might say: "We need a collection of tasks. Each task has a title, a description, a due date, and a status (like 'pending' or 'completed'). Each task belongs to one user." This natural language description contains all the essential information for a data model, including relationships and field types. Strategy 3: Expressive Querying Similarly, when you need to fetch dat

Someone would ask a question in a public Slack channel. Every so often a couple of people would start to answer. Then the manager would step in, say what was going to happen, and the thread would go quiet. On its own, it looks like nothing. A decisive manager keeping things moving. But it was a team going quietly into debt, and the dead Slack thread was one of the interest payments. You already know tech debt. You cut a corner in the code to ship faster, and you pay interest on it later in bugs, slow changes, and the one file nobody wants to touch. Culture debt works the same way, except the corners you cut aren't in the code. They're in the norms, the expectations, and the relationships that decide how people actually work together. But tech debt is visible. You can see it, point at the file, write a ticket, argue about whether it's worth paying down. Culture debt is more dangerous because it gives you none of that. You don't watch it accruing. You see the symptoms, and by the time they show up, the debt has already compounded. Let me tell you how a team I joined got there. The reward was volume. The only thing that reliably got praised was pushing a lot of code. The manager was open about it...their whole framing of the job was being able to out ship anyone on the team. Everyone else stayed quiet. Nobody ever stood up and argued against quality. If you'd asked, the manager would have agreed that testing mattered and that quality mattered. Those things just never got prioritized. So over and over, what actually got rewarded (volume) quietly beat what everyone said they wanted. This didn't happen out loud. The reward silently won every time. You can guess what that bought. Planning went first, so features shipped in half finished states and got abandoned there. Testing basically didn't exist. We had a QA person, but things slipped through constantly. Bugs were everywhere. Plenty of features barely worked, and some just didn't. The human side hollowed out at the same

The problem with most web UIs is the gap between user action and visible feedback. A user clicks "like" and waits 200-400ms for the server to respond before the button changes. That delay reads as slowness even when the server is fast. The network round-trip is the ceiling. Optimistic UI inverts this: assume the operation will succeed, update the UI immediately, then reconcile with the server response when it arrives. If the server fails, roll back. React 19's useOptimistic hook gives you this pattern with minimal boilerplate and automatic rollback built in. The API const [ optimisticState , addOptimistic ] = useOptimistic ( state , // the current "real" state — synced from server updateFn , // (currentState, optimisticValue) => newOptimisticState ) optimisticState — during a pending transition, reflects the optimistic update. Once the transition completes, it reverts to state addOptimistic(value) — triggers an optimistic update, must be called inside startTransition Pattern 1: Like Button ' use client ' import { useOptimistic , useTransition } from ' react ' import { toggleLike } from ' @/actions/likes ' type LikeState = { liked : boolean ; count : number } export function LikeButton ({ postId , initialLiked , initialCount }: { postId : string initialLiked : boolean initialCount : number }) { const [ isPending , startTransition ] = useTransition () const [ optimisticState , addOptimistic ] = useOptimistic < LikeState > ( { liked : initialLiked , count : initialCount }, ( current ) => ({ liked : ! current . liked , count : current . liked ? current . count - 1 : current . count + 1 , }) ) function handleToggle () { startTransition ( async () => { addOptimistic ( ' toggle ' ) // updates UI immediately await toggleLike ({ postId }) // syncs with server }) } return ( < button onClick = { handleToggle } disabled = { isPending } > < Heart className = { cn ( ' h-4 w-4 ' , optimisticState . liked && ' fill-red-500 text-red-500 ' ) } /> < span > { optimisticState . count }

Introduction I've spent way too much time trying to come up with some generic introduction for this story, but then I realized none of you probably want to read that anyway. So instead, I'll just jump straight into the story—which is why you're here in the first place. The day I received the requirements The story begins when I received the requirements for a new feature that allows Teachers to share their presentation to review slides before the Lecture begins, so we would have teachers aids using the web version and seeing a screenshare from the main pc powerpoint, at first I thought maybe we can use HLS or RTMP for this and be okay with the 3 seconds delay that it has, but then I continued reading the ticket, we also needed the user to move to the next and previous slides via the web application, which immediately threw my initial idea out of the window. This is because if the user needs to interact with the application there is no way it will be usable without almost immediate feedback. Since we needed to show this to the client quickly we had 2 weeks to implement this feature, so before I did anything, I stopped and started drafting a simple design doc, which besides the fancy name was really just a document with my raw notes taken from research and comparisons between different solutions. After spending some time doing research and looking into different architectures and engineering blogs from companies like Twitch, Slack and Discord, I narrowed the possibilities down to four common architectures used for this type of use case. Architecture Options P2P Mesh This approach revolved around a user establishing WebRTC connections with every other user in the room. Besides being difficult to manage in terms of connections and sessions, it had one fatal flaw: network and CPU overhead. If we had twenty users in the room, every participant would maintain nineteen separate peer connections while sending nineteen streams, quickly consuming both CPU and bandwidth. MCU (M

This week's Java roundup for July 6th, 2026, features news highlighting: the GA release of TornadoVM 5.0; point releases of JHipster, Keycloak and Google ADK; maintenance releases of GraalVM Native Build Tools and Micronaut; the OmniFish Build of Payara and introducing Vidocq, a new implementation of the Jakarta EE 11 Core Profile and MicroProfile 7.1. By Michael Redlich

Google and AIM launched ATL Saathi, a Gemini-powered AI tool empowering Indian educators in robotics labs.

Davide de Paolis discusses the realities of rolling out cloud infrastructure compliance without fracturing developer relations. Drawing from a real-world platform team reboot at Sevdesk, he explains how to implement "minimum viable governance" on AWS, utilize event-driven Slack alerting to automate policy feedback, and shift from rigid enforcement to high-empathy, data-driven collaboration. By Davide de Paolis

Hello, I'm Rijul. I'm building git-lrc, a micro AI code reviewer that runs on every commit. It's free...

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Sperm donors need limits, says a European fertility group Ties van der Meer doesn’t know how many siblings he has. The 47-year-old was conceived at a private fertility clinic using sperm…

Your AI agent for autonomous web app testing Discussion | Link

Create AI ads that convert, not just look good Discussion | Link

Swimming against the iGarden Swim Jet X Pro 10’s artificial current is a real workout, even if it doesn’t quite replace a full-size lap pool.

This essay was written with Nathan E. Sanders, and originally appeared in The Guardian . Opposition to AI data centers has emerged as a primary theme in US politics, one that—surprisingly—doesn’t fall along party lines. We applaud people coming together for constructive debate on any issue, and agree that communities need to evaluate whether any economic benefits these data centers bring is worth their costs. Still, we worry that a focus on data centers obscures the larger impacts of AI on people’s lives: the concentration of power of AI companies, and their widespread political and financial influence...

Experts explain how they work, what they can do, and what's still unsettled.

When a series of regional outages forced a rethink of a multi-region AWS API, the team discovered that an obstacle to global failover was hiding in plain sight: a pre-flight discovery call baked into every client session years earlier as the only available option. This article describes what it took to remove it, and what the rollout actually cost. By Suresh Gururajan