找到 2651 篇相关文章
GitHub热门项目 | 🎬 Fully automated YouTube channel management with AI agents. Creates, optimizes & publishes videos 24/7. Works with FREE Gemini API or OpenAI. No coding required! | Stars: 1,937 | 118 stars today | 语言: JavaScript
GitHub热门项目 | An open-source, GPU-accelerated physics simulation engine built upon NVIDIA Warp, specifically targeting roboticists and simulation researchers. | Stars: 5,341 | 15 stars today | 语言: Python
GitHub热门项目 | Community-contributed instructions, agents, skills, and configurations to help you make the most of GitHub Copilot. | Stars: 37,813 | 47 stars today | 语言: Python
Floating labels. Inline validation. Custom checkboxes, radios, and a toggle switch. A gradient button with a press-down micro-interaction. Every bit of it below is CSS — no form library, no useState , no event listener wiring up a class toggle. That's form.fscss — the module in the FSCSS ecosystem. Same philosophy each time: solve the hard visual problem once, ship it as importable mixins, let the browser do the actual work. <script src= "https://cdn.jsdelivr.net/npm/fscss@1.1.24/exec.min.js" defer ></script> <style> @import (( * ) from form ) @ form-root () @ form-group (. form-group ) @ form-input (. form-input ) @ form-label (. form-label ) @ form-float (. form-group , . form-input , . form-label ) @ form-checkbox (. form-checkbox ) @ form-btn (. form-btn ) @ form-btn-primary (. form-btn-primary ) </style> <div class= "form-group" > <input class= "form-input" type= "text" placeholder= " " > <label class= "form-label" > Full name </label> </div> <label class= "form-checkbox" > <input type= "checkbox" checked ><span></span> I agree to the Terms </label> <button class= "form-btn form-btn-primary" > Create account </button> The two tricks doing all the work Forms feel like they need JavaScript because most tutorials reach for it immediately. Two native CSS mechanisms cover almost everything a "modern" form needs. Floating labels run entirely on :placeholder-shown . Give the input placeholder=" " — a literal space, not empty — and the browser now knows, purely in CSS, whether the field is empty and unfocused: .form-input :focus + .form-label , .form-input :not ( :placeholder-shown ) + .form-label { top : -9px ; font-size : 11px ; color : var ( --form-accent ); } No state, no class toggling on keyup. The label just reacts to what the browser already knows about the input. Checkboxes, radios, and the switch all use the classic checkbox-hack: the real <input> stays in the DOM (so it keeps native keyboard support and form submission) but is visually hidden, and a sibling
This is a submission for DEV's Summer Bug Smash: Clear the Lineup powered by Sentry . The bug When you call @google/genai in streaming mode and the model asks to run a tool, Sentry's JavaScript SDK records that tool call to the span twice. One tool call in, two entries out. The attribute that carries them is gen_ai.response.tool_calls . It should hold one object per call. For a single streamed controlLight call it held two. Worse, the two did not even agree on their shape. Here is a real capture, which I come back to at the end: [ { "id" : "call_2079699" , "args" :{ "colorTemperature" : "warm" , "brightness" : 30 }, "name" : "controlLight" }, { "type" : "function" , "id" : "call_2079699" , "name" : "controlLight" , "arguments" :{ "colorTemperature" : "warm" , "brightness" : 30 }} ] Same id, same call, listed twice. One entry keys the parameters under args , the other under arguments . Anything reading this later sees two tool invocations where the model made one. Following the value The streaming instrumentation lives in packages/server-utils/src/ai/google-genai/streaming.ts . Every chunk of the stream runs through handleCandidateContent . That function wrote tool calls from two places: function handleCandidateContent ( chunk , state , recordOutputs ) { if ( Array . isArray ( chunk . functionCalls )) { state . toolCalls . push (... chunk . functionCalls ); // push #1 } for ( const candidate of chunk . candidates ?? []) { // ...finish reasons... for ( const part of candidate ?. content ?. parts ?? []) { if ( recordOutputs && part . text ) state . responseTexts . push ( part . text ); if ( part . functionCall ) { state . toolCalls . push ({ // push #2 type : ' function ' , id : part . functionCall . id , name : part . functionCall . name , arguments : part . functionCall . args , }); } } } } Push #1 spreads chunk.functionCalls into the accumulator. Push #2 walks candidate.content.parts and pushes every functionCall it finds. They look like two different sources. They
Every generative-AI app I worked on hit the same wall. Pick a backend — OpenAI for images, Volcengine for video, Mureka for music — and before long the app is full of that provider's SDK quirks: its field names, its sync-vs-async polling loop, its error shapes. Then if want to swap one backend, or add a second one for failover, and it's a rewrite. mm-gateway is an open-source Python gateway that sits in front of that mess. One provider-neutral contract — over 13 backends : OpenAI · Google · xAI · DashScope · Volcengine · Flux · Stability · ElevenLabs · MiniMax · Mureka · ACE-Step · OpenRouter · UdioAPI The idea is simple: provider wire formats never appear in application code. Every request goes through a strict, modality-specific envelope — an ordered list of typed input parts plus provider-neutral parameters — and each backend adapter translates that to its native SDK or REST shape. For more information, visit https://github.com/sloth-os/mm-gateway
As someone who is constantly exploring ways to make AI applications faster and cheaper, I found...
A hands-on test of BGE-M3 + Qwen3 (RAG vs. direct-context answering) on a real research paper and a full-length book including a retrieval bug hiding in a footnote, and one surprisingly good model behavior. I wanted to answer a simple question: when you feed a document to an AI model, is it actually reading it or just pattern-matching to whatever text happens to look similar to your question? So I built a small open-source pipeline to test this directly. For any document and question, it generates two separate answers: RAG answer: BGE-M3 finds the most relevant chunks of the document, and Qwen3 answers using only those chunks. Direct answer: Qwen3 reads the raw document text directly, no retrieval involved. Both run on a free Google Colab GPU. I kept the retrieval side deliberately "vanilla" fixed-size chunking, plain cosine similarity, no reranking, no fancy tricks so I could see exactly where the basic version breaks before adding any fixes. Before running my first real test, I already knew one thing to guard against: reference lists. Early experimentation (not covered here) showed that a paper's bibliography, once chunked like any other text, can get retrieved as if it were real content a citation for a paper about "text embeddings" can look deceptively similar to a generic question about a document's topic. So going in, my pipeline already strips everything after a References/Bibliography heading before chunking. With that fix in place, I ran two real tests. Test 1: A research paper on Nepali legal machine translation First document: a SIGUL 2024 workshop paper on a bidirectional English-Nepali machine translation system for the legal domain. Question: "What is this paper about?" RAG answer: This paper presents the first transformer-based bidirectional machine translation system for the English-Nepali legal domain, using a custom-built parallel corpus of 125,000 sentences. It achieves encouraging BLEU scores and addresses the scarcity of domain-specific legal tr
A couple of weeks ago I dropped the CRA text (the EU's cybersecurity regulation for IoT devices) into ChatGPT and asked when the main requirements actually kick in. The answer was confident and wrong - it mixed up the date the regulation entered into force (2024) with the date the requirements actually apply (2027). Three years off, stated like an obvious fact. My team (Platanor, embedded security for IoT) has been building an internal reference on CRA/RED/NIS2/CSA for a few months now, and this is exactly the kind of mix-up we kept running into whenever we just threw the regulation PDF at a model. The problem isn't the model. It's how the source is laid out: dates are scattered across different articles with no explicit link between them, token-based chunking cuts sentences off mid-article, and the model has no way to tell how fresh the text is. When we rebuilt the base as a public repository, we fixed this with file structure, not prompting. Cut by article headings, not by tokens: ### Article 13 Obligations of manufacturers 1. When placing a product... ### Article 14 Reporting obligations... ### Article N is a natural boundary. Each chunk stays whole - the article never gets split mid-sentence. Source priority, written into the file itself, not the prompt: primary source > official related documents > third-party summaries > our own analysis. The model sees this right next to the content, not as an instruction that's easy to lose in a long chat. A verification date on every file: > Last verified: 2026-08-10. > Annex I application deadline: 11 December 2027 (not to be confused with the entry-into-force date - 10 December 2024). That one line is what removed the exact error I opened with. llms.txt at the repo root - an index of every file, so an agent can pick what to load instead of reading the whole repository. The same questions now get answered correctly - not because the model got smarter, but because the source stopped being one continuous wall of text. We pac
How to rescue abandoned open-source projects, modernize build systems, and generate multi-architecture Docker images (x86_64, ARM64) in a single afternoon with Antigravity.
Nearly every AI agent benchmark you read is unfalsifiable. Not wrong, necessarily - unfalsifiable. There's a blog post with a bar chart, a claim that framework A beat framework B, and no way for you to check it. No run count. No model version. No raw output. Often no cost. You are asked to trust a summary statistic produced by people with an interest in the result. We publish agent benchmarks, so this is our problem too. This post is about the evidence bundle we settled on, and how you can pull one down and take it apart in about two minutes. Every command below is one I actually ran while writing this, with its real output pasted in. The claim we're going to try to break From one of our pilot runs: LangGraph 1.2.9 and Pydantic AI 2.13.0 both completed 20 of 20 tasks under gpt-4o , at a total spend of $0.094275. That's the sort of sentence you'd normally have to take on faith. Let's not. Two minutes to verify it yourself The bundle is a directory in a public repo. Pull it: BASE = "https://raw.githubusercontent.com/benchclawio/harness/main/results/gpt-4o-vs-gpt-4o-mini-tool-calling-2026-07-24" for f in SHA256SUMS README.md gpt4o-pilot-manifest-v0.4.0.json \ scored-pilot-gpt4o-raw-2026-07-24.jsonl \ scored-pilot-raw-2026-07-24.jsonl \ scored-pilot-analysis-2026-07-24.json \ scored-pilot-gpt4o-analysis-2026-07-24.json \ real-pilot-status-manifest-v0.3.0.json ; do curl -sfO " $BASE / $f " done First question: is this the same data we published, or has something drifted? sha256sum -c SHA256SUMS README.md: OK gpt4o-pilot-manifest-v0.4.0.json: OK real-pilot-status-manifest-v0.3.0.json: OK scored-pilot-analysis-2026-07-24.json: OK scored-pilot-gpt4o-analysis-2026-07-24.json: OK scored-pilot-gpt4o-raw-2026-07-24.jsonl: OK scored-pilot-raw-2026-07-24.jsonl: OK That's the cheapest integrity control there is and almost nobody ships it. It costs one line in your run script and it means a reader can tell the difference between the file you published and a file someone edited afte
We just shipped EverShop 2.2.1 — the largest release since 2.0. It folds in the React 19 work that had been sitting in an unpublished 2.1.3 branch and stacks four months of development on top of it: a visual page builder, a blog module, entity custom fields, a multi-language storefront with a translated admin, a rebuilt shipping and fulfillment stack, built-in cloud storage, product recommendations, and a serious security and performance pass. If you're upgrading an existing store, one number to keep in mind: 31 database migrations across 10 modules run automatically on first start. Several of them transform data and drop legacy tables, so back up your database first and read the breaking-changes section below. This release also patches several security vulnerabilities, so upgrading promptly is the right move. Here's a tour of what's new, and what you'll need to change if you maintain themes or extensions. Visual Page Builder The headline feature is a drag-and-drop editor for the storefront, living at /admin/page-builder . You edit any storefront route — plus CMS pages and landing pages — by composing widgets into your theme's areas, with layout-aware drag/drop. The workflow is draft-based: changes accumulate in a per-admin, per-theme draft changeset with per-widget auto-save. When you're ready you can publish immediately, or schedule a rollout for later — and those rollout plans stay editable and cancelable right up until they run. There's inline editing on the canvas (text and images edited in place, with an image picker that understands cloud storage), a layers panel, a "Globals" view for site-wide areas, and per-widget styling controls. Link fields resolve products, categories, CMS pages, and blog posts through a single unified link resolver. Because it's touching public-facing content, the whole editor pipeline went through a dedicated security-hardening pass and ships with an end-to-end test suite. Blog module EverShop now has a first-class blog core module: p
When a language-learning product says it teaches “the sounds of French,” one deceptively simple question appears immediately: how many sounds are there? There is no useful answer without first defining the job the inventory is meant to do. A phonological analysis, a pronunciation dictionary, a speech-recognition system, and a beginner curriculum can all model French sound structure differently without one of them necessarily being careless. They have different users, evidence, and failure costs. Our team encountered this while turning Parle's internal pronunciation inventory into a public CSV. We needed a list that could connect IPA symbols to French spelling patterns, example words, and short mouth cues for English-speaking beginners. We also needed to avoid presenting one product's learning model as the only correct account of French phonology. The result is a bounded dataset of 35 practical sound entries. This article explains the design decisions, the schema, and the limits we published with it. A teaching inventory is a model, not a census The International Phonetic Alphabet gives us a shared notation for describing speech sounds. It does not require every analyst or teacher to draw identical category boundaries for every language variety. The official IPA chart is a notation system; selecting a French inventory still requires linguistic and pedagogical decisions. Counts can change when an inventory treats any of the following differently: a contrast that is maintained by some speakers but merged by others; a marginal or loan sound that appears mainly in borrowed words; schwa, whose realization and deletion depend heavily on context and variety; a historical contrast that remains visible in spelling but not in every speaker's production; a phonetic realization versus a contrastive phoneme; a glide represented separately from its related vowel. For a curriculum, the important question is not “What number wins?” It is “What distinctions and cues help this audienc
TL;DR — I've been building the Social Resource Floor: an open blueprint for coordinating one person's access to basic survival resources — food, housing, energy, healthcare, and more — across many independent providers, so that reaching those resources is grounded in being human rather than in financial access. The first blueprint version is now complete: language-neutral schemas, prose specifications, a reference implementation, and a first adapter. This post is about the engineering choices behind it, and the reasons for each — how it stays a contract rather than a product, how it keeps personal data out of the coordination layer, why it binds to existing standards instead of inventing new ones, and how I check that the contracts are implementation-independent rather than just claiming they are. The problem the Floor is trying to help with Today, for most people, survival routes through financial access. To reach food, housing, energy, or healthcare you generally need money, and to hold or move money you need banking, employment, or purchasing power. Financial access has become the gate standing in front of the resources a person needs to stay alive. The goal of the Social Resource Floor is narrow and specific: to help make it so that financial status is not the condition that determines whether a person can reach the basic resources required to survive. It does not try to abolish money, banks, or markets — money stays a first-class resource and delivery method. It aims at one thing: a floor beneath which no person should fall, defined locally, reachable regardless of financial circumstances. That's the mission. Everything technical below exists to make that mission buildable by the institutions — governments, municipalities, NGOs, cooperatives, community providers — that would actually run it, without asking any of them to give up their own systems or hand over their data. Where the Floor sits The delivery systems for social protection already exist and are stron
OpenAI’s rogue agent hack was a watershed moment for AI safety and cybersecurity. It also sparked internal questions about the culture that led to it.
OpenAI is expanding ChatGPT beyond chat with ChatGPT Work , a cross-platform work environment that includes a desktop agent able to interact with local applications, files and browser content. The change matters because it moves ChatGPT closer to an operational role: not only explaining how to complete a task, but potentially clicking, typing, moving files and staying engaged with a project over time. In OpenAI's official announcement on ChatGPT Work , the company describes a unified experience across web, mobile and desktop. The desktop app combines Chat, Work and Codex, while its built-in browser and local computer capabilities are intended to support more contextual, end-to-end work. OpenAI's terminology centers on ChatGPT Work and Computer Use. "Computer History," the name used in the originating signal, is not the feature name used in the official announcement. The underlying shift is significant for developers and knowledge workers. A chat interface has traditionally depended on users copying information into a prompt, describing where files live, and manually carrying results into the next application. Desktop automation can reduce those handoffs, provided users grant the relevant access and organizations establish appropriate controls. From answers to work across a computer ChatGPT Work is positioned as an agentic layer for work that spans apps and files. OpenAI says the desktop agent can act locally in the background, including interacting with applications, files and browser content. It also highlights plugins, workflows and Scheduled Tasks as ways to connect tools and automate recurring actions across connected apps and local files. That does not mean every task should be delegated without review. The practical value depends on how clearly a workflow can be defined, the permissions it requires, and the consequences of an incorrect action. For example, moving or modifying local files is fundamentally different from drafting a response in a chat window. The
Another OpenAI executive is departing. Denise Dresser, who joined OpenAI as its chief revenue officer in December after serving as CEO of Slack, will be leaving in the "coming weeks" to "pursue other opportunities," she said in a team note posted to LinkedIn. Dali Rajic, president and COO of Wiz, will be taking over the […]
IBM plans to train and certify tens of thousands of consultants on OpenAI's technologies as part of this deal.
I have been an avid privacy advocate for quite some time now. It started with outright rejecting all "Big Brother" tech, and being hyper paranoid with every little detail, willing to sacrifice ease of use, in exchange for added privacy. However, as time went on, I slowly understood what is that I actually consider my "threat model" , and what exactly is my "sweet spot" between privacy and ease-of-use. I'm now back on multiple "Big Brother" tech, with some extra steps, to ensure I get the facilities they provide, while also being wary of my data. However, while I did make this compromise, I was very annoyed I had to make this compromise in the first place. In an ideal world, I would want the tech where everyone actually is, and is the standard for that particular domain, to have privacy features by default, and not be treated as a niche, or a luxury you have to go out of your way to avail. It was this annoyed version of myself, with my strong belief of privacy features and tools being the new norm, I started looking at everything with that lens. And that is how I got concerned about tracking in links and URLs. Try sharing any Instagram post, or YouTube video, by copying its URL, and you will see a bunch of garbage (garbage to you) in the link. Take for example this (fake) link: https://www.instagram.com/p/Cxyz123/?igshid=AbCdEf123456 These links contain something along the lines of utm_* (marketing attribution), or in this case, Ad-Click Identifiers, such as fbclid (Meta), gclid (Google), or igshid (Instagram). These pesky trackers help collect information regarding you, your device, and also help connect you across the internet, mapping your movement as you browse the web. The thing is, while there are good Samaritans who have built tools and websites to get rid of these trackers, and many privacy oriented browsers have introduced a "Copy Clean Link" option while copying the link from the browser, I believe there should be a tool which should not be restricted to a
I’ve been building an open-source project called TraceMotive. It started from a problem I kept running into with AI agents: When an agent run fails, the place where the error appears isn’t always where the execution first started going wrong. That makes debugging agent workflows harder than it looks. So I built TraceMotive, a local-first tracing and debugging tool for AI agent execution. What TraceMotive does The current v0.1 includes: Python SDK canonical traces and spans a local Collector backed by SQLite a React UI for inspecting agent runs optional OpenAI Agents SDK integration TraceMotive is local-first, and content capture is disabled by default. I’m intentionally keeping the first version small. I’m not trying to add replay, automatic root-cause analysis, cloud sync, or support for every agent framework yet. Why? I’d rather get real feedback before adding a lot of features. Right now I want people who actually build AI agents to try it and tell me: where setup is confusing what breaks what information is missing from traces what feels awkward in the API The longer-term direction is: “The causal debugger for AI agents.” Eventually, I want TraceMotive to help identify where an agent execution first started going in the wrong direction, instead of only showing where the final error appeared. But first, I want to make the basic observation and debugging layer solid. Try it PyPI: pip install tracemotive GitHub: https://github.com/doraemonfv-glitch/tracemotive If you build AI agents, I’d really appreciate you trying it for a few minutes and telling me what you run into. Even small feedback is useful.