AI 资讯

Fizz has expanded its lawsuit against rival Sidechat, alleging that a Maveron VC shared its confidential information obtained during a fundraising meeting with the competing startup.

The movie-centric social network is said to be seeking new ownership.

As the AI boom boosts demand for RAM, SK Hynix - one of the world's biggest suppliers of memory chips - launched on Wall Street Friday. The South Korean chipmaker opened at $170 per share and raised $26.5 billion, surpassing Alibaba's record as the largest debut of a foreign company, according to reports from The […]

The AI chip boom just produced its biggest Wall Street moment yet. Now SK Hynix and Samsung are being asked to build U.S. factories.

HyperTexting's new app aims to make the open web feel more like social media by turning websites, blogs, newsletters, and podcasts into a scrollable feed, while also making it easier to post to your own website.

Spotify is giving listeners control to fine-tune what gets surfaced for them in Release Radar - one of its most popular weekly playlists. The new options allow you to narrow the playlist to a specific genre, focus on artists that are new to you, and more. Listeners can choose from up to five options like […]

Affected users will have to configure their lights and settings all over again.

Netflix has shows and movies. And video games. And live sports. And podcasts. And also, apparently, YouTube videos? For a company that used to seem like the next big thing in TV, it all feels a little frenetic, and maybe a tad desperate. For a company that sees sleep as its primary competitor, it might […]

If you were to look up into the night sky, what would you see? Countless points of light, scattered in every direction. Most of what you're looking at are stars. But some of those points are whole galaxies—vast collections of stars, spread across incomprehensible distances, compressed by that distance into a single pinprick of light. And what you can see with the naked eye is only a small fraction of what's actually out there. I want to use this as a way to offer you a way of thinking about how large language models work. Just an analogy, not literally what's happening inside the mathematics—that's not my forte. My hope is that it captures something true about the mechanics, and more importantly, it gives you a mental model you can actually use when you're working with these systems. About two years ago, I was wrestling with finding a way of explaining what an LLM does. My first analogy was that of a dictionary. The naive view was that a dictionary uses words to define other words, and an LLM holds a matrix of words with weights that describe their relationships to each other. So the parallel seemed natural: both systems work through relational structure. However, a dictionary gives you denotation—the surface-level meaning. It's a lookup tool for individual words, not a model of language itself. And critically, you have to already understand language before a dictionary is useful to you at all. The analogy didn't capture what was actually happening in the weight relationships—the distributional semantics, the contextual patterns that let an LLM generate coherent text. Ok, so back to galaxies, when you look up at the night sky, you're not seeing distance—you're seeing direction. That galaxy over there, the one that looks like a point of light, could be millions of light-years away, but what matters for our analogy isn't how far it is. It's which way you're looking. And when you point yourself in that direction and venture toward it, you discover it's not a point at a

Note: Google Cloud credits are provided for this project. In this blog post, I want to demonstrate how I use Antigravity CLI to build an image analysis demo using Angular, Firebase Hybrid & On-device Inference Web SDK, and Gemini models. Users upload an image and use a Gemini model to analyze it to generate a few alternative texts, tags, recommendations, and CSS tips to enhance the image quality. When the demo is running on Chrome 148+, the Hybrid & On-device SDK leverages the Prompt API of the on-device Gemini Nano model to perform the image-to-text tasks, and the token usage is 0. When other browsers such as Safari or Firefox executes the same tasks on the demo, the SDK falls back to Cloud AI (Gemini 3.5 Flash model), and the token usage is greater than 0. Next, I will describe how I installed the skills in my Angular project, and registered the Stitch MCP server in the Antigravity CLI to develop the infrastructure, services, and UI design of my demo. 1. Skills I installed grill-with-docs , angular , and firebase skills in my project for the following reasons: grill-with-docs: Conduct a rigid Q&A session to generate a specification for a feature, refactor or a critical fix. AI is responsible for performing a thorough analysis and putting in more effort to generate code to achieve the task. Angular: Provide the best practices of Modern Angular architecture, such as using signals and signal forms. Firebase: Provide the skill for Firebase AI Logic, Firebase Remote, etc. Resources Firebase Hybrid & On-device Image Analysis App Firebase Hybrid & On-device Inference Chrome Built-in Prompt API Stitch Stitch MCP Server grill-with-docs Angular skill Firebase skill

Anthropic Built a Fix That Proves the Problem 🔍 Right, so Anthropic dropped Tool Search on November 24th alongside Claude Opus 4.5, and I need you to sit with the implications for a second because they're brutal for MCP. Tool Search lets you mark tools as defer_loading: true . When you do that, the tool's schema never enters your context window until Claude actually needs it. Claude gets a name and nothing else. When a task comes in that requires the tool, Claude calls the Tool Search tool, pulls the schema, and only then does it have the full definition loaded. Lazy loading for AI tools. Sounds boring. It's not. Here's the number that should make you spit your tea out: accuracy went from 49% to 74% on Opus 4. On Opus 4.5, it climbed from 79.5% to 88.1%. Same model. Same tools. The only difference is not loading the tool definitions upfront . I'll say it again. The mere presence of tool schemas in context was tanking accuracy by 25 percentage points. Twenty-five. Not because the model got confused by irrelevant tools (though it did). Because the context window was full of JSON nobody was using, and the model was drowning in it. This is an admission dressed up as a feature I've been banging on about MCP's context cost since July. The Playwright MCP server alone burns 12.8k tokens just sitting there. Load four or five MCP servers and you've torched half your context window before asking the model to do anything useful. Tool Search is Anthropic's answer. And the answer is: don't load the tools. Think about that for a second. The company that created MCP built a feature whose entire purpose is avoiding the cost of loading MCP tool definitions. If that isn't an admission that the protocol has a context problem, I don't know what is. They didn't say "we've made MCP more efficient." They said "here's a way to not load MCP schemas until the last possible moment." The fix is the diagnosis. And the timing is proper interesting. Three weeks earlier, on November 5th, mcporter d

The Man Behind the Minimalism I met Peter Steinberger at Claude Anonymous in London sometime in June, and the presentation was properly fun. He's got that energy where you can tell he's not performing, he's just genuinely excited about the stuff he's building. Maintains a 300k line TypeScript React ecosystem. Web app, Chrome extension, CLI tool, Tauri desktop client, Expo mobile app. All maintained by one person with AI agents. And his setup is basically: open 3-8 parallel Codex instances, write short prompts (often 1-2 sentences plus a screenshot), let them go. He calls everything else "charade." I've been following his work since and I reckon he's one of the most practical voices in the agentic coding space. No hype. No frameworks for the sake of frameworks. Just a bloke who ships a lot of code with AI and has opinions about how to do it well. His Principles (and Where I Stand) Here's the full list from his talk and blog posts. I agree with most of it. Where I don't, I'll say so. Things I violently agree with: Use tmux to run CLIs persistently. Absolutely. This is foundational. If you're not doing this, start here. Use ast-grep as a pre-commit hook. Proper structural linting that catches things regex can't. Brilliant addition to any agent workflow. Ask for options before executing. I built a whole /interview skill around this. Get the model to present choices rather than guessing. Way better outcomes. Don't reset context. Cold start wastes time + tokens. Agreed, but I'd go further. Don't reset, /handover instead. Summarise what was accomplished, carry it to the next session. Cold start is a waste. Context reset without transfer is nearly as bad. Give examples. Multi-shot prompting is always better than zero-shot. If you want the model to output something specific, show it what that looks like first. Massively improves one-shot accuracy. Use voice input. Absolutely. I started with Superwhisper and it changed how I work. Speaking your intent while pacing around the

What I'm Reading and Watching This Week 📚 [Article] " Deep Work in the Age of AI " by Cal Newport. Proper interesting take on how AI coding tools change our relationship with focused work. Made me rethink a few things about how I structure my own deep work blocks. [Video] "The Art of Code Review" by ThePrimeagen. Some good bits on making code reviews actually useful instead of the usual rubber-stamping faff. [Book] Currently reading "Staff Engineer" by Will Larson. If you're a senior dev wondering what comes next, this is the one. Side Projects and AI Dev Tools I'm Building 🛠️ stevengonsalvez.com 🟢 Active Next.js 15 + MDX blog platform with the byte-sized banter section you're reading right now. Foundation's done, blog section coming together. Next up is deploying to Vercel and sorting the domain. Personal MCP Server 🟡 On Hold Custom MCP server for personal productivity tools. Paused while I wait for Claude Desktop to get its MCP support properly sorted. What Changed This Week 🗞️ Migrated the blog from dev.to-only to a self-hosted Next.js site. Launched this new weekly banter format. Still experimenting with how often I actually want to publish, reckon weekly is about right but we'll see. Developer Tips and Tricks 💡 Quick git alias for better logs: git config --global alias.lg "log --graph --oneline --all --decorate" This creates a beautiful visual git history that's way more readable than the default log. Claude Desktop optimization tip: Keep conversations small and restart often. The message limit resets every 5 hours, but shorter conversations use fewer tokens per message. You can get 2-3x more usage by starting fresh chats instead of continuing long threads. (If you're weighing up Claude Code versus other terminal tools, I did a proper comparison of Claude Code vs Warp AI a while back.)

Skills are what MCP should have been 📝 Anthropic announced Agent Skills on October 16th, and I reckon this is one of those quiet releases that ends up mattering more than the flashy model drops. A skill is a markdown file with YAML frontmatter. That's it. No server process. No JSON-RPC. No WebSocket connections. No Docker containers. A markdown file. And the clever bit is how it loads. Level 1 is metadata. Name, description, trigger conditions. Costs you about 30 to 100 tokens and it's always in context. Level 2 is instructions. The actual "how to do the thing" content. Under 5,000 tokens. Only gets loaded when the model decides it's relevant to your task. Level 3 is resources. Referenced files, example code, whatever. Only pulled in when the instructions explicitly reference them. So you've got a system where 100 skills can sit in your context at a cost of maybe 3,000 to 10,000 tokens total. The metadata layer alone. The model reads the names and descriptions, figures out which ones matter, and loads only what it needs. Now compare that to MCP. Four or five MCP servers and you're looking at 40,000 to 60,000 tokens of JSON schemas. Loaded upfront. All of them. Whether you need them or not. Sitting in your context window like furniture in a flat you never use, taking up space and making the place harder to navigate. Most MCP servers were never about live data Here's the thing I keep coming back to. What were people actually using MCP for? Some of it was legitimate live data connections. Database queries. API calls. Fetching real-time information the model can't have in its training data. Fair enough. That's a genuine use case and skills can't replace it. But a huge chunk of the MCP ecosystem was procedural knowledge. How to deploy this thing. How to format commits. How to run the test suite. How to interact with Jira. Step-by-step instructions wrapped in a protocol layer and loaded as tool definitions. Tens of thousands of tokens to tell the model "when someone asks