AI 代码与智能体浪潮席卷全球：Postgres 19 延期、大模型持续迭代与开源反制迎来爆发

Reflection releases Beam, a 501B MoE model under Apache 2.0

Beam is our first open-weight model, built for advanced coding, reasoning, and long-horizon agentic workloads with frontier reasoning efficiency. Today, we have released Beam’s model weights on Hugging Face under an Apache 2.0 license. Our

Introducing Beam: Reflection’s 501B open-weight model

Beam is our first open-weight model, built for advanced coding, reasoning, and long-horizon agentic workloads with frontier reasoning efficiency. Today, we have released Beam’s model weights on Hugging Face under an Apache 2.0 license. Our

D1-Ultra: Introducing the Best Decision Model

Today, we release d1-ultra, our new multimodal decision model with state-of-the-art performance. Most decision models work well on easy and common tasks. We built d1-ultra for harder, messier real-world use cases, like spotting a defect on

How pro-China are Chinese models?

Companies in China are subject to content safety regulations, some of which relate to political topics. There is a common perception that because of these regulations and general state censorship, Chinese AI models hold highly impenetrable

Jev is NOT a reranker (but it's close)

We run millions of retrieval queries a month at Kapa. Every one goes through a reranker, so picking the right one matters a lot to us.Erik Kaunismäki from turbopuffer recently showed TypeSafe AI's Jev beating dedicated rerankers on their be

The Agentic Scientist – SakanaAI

Turn a task description into experiments, ablations, and a paper, driven by coding agents. Agentic Scientist treats coding agents as AI scientists. Starting from a task.md, the agents propose ideas, run experiments, evaluate outcomes agains

Researchers Stranded an AI Agent on a Virtual Island. Things Got Weird

Life, it seems, wants to play. Anyone who’s ever spent ten minutes with a five-year-old knows this perfectly well: very young children are masters at make-believe. And it’s not just people—play comes instinctively across the animal kingdom.

Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline

Native-script data for 22 languages, romanized data for twelve, and a reusable GPU-accelerated pipeline for translation, quality assessment, and repair. Authors: AI4Bharat — Mohammed Safi Ur Rahman Khan, Mitesh Khapra, Anoop Kunchukuttan. N

Datadog Scout 1.0: New SRE model for investigating production issues

Othmane Abou-AmalDirector, AI Research Ben CohenStaff Research Scientist Afshin RostamizadehStaff Research Scientist Ameet TalwalkarChief Scientist Omri SassSenior Director, Product Daniel ShanStaff Engineer Today, we’re introducing Scout 1

Show HN: An improved kokoro distillation in 9MB

I'm a 47yo full stack developer that's investing all his free time (and not only that) into thinkering, building, researching with AI. Some days ago I found a paper about distilling kokoro on a ~9MB with around 25x RTF. It was very interesting and since the experiment was on a single english voice and I'm Italian, I wanted to see if I could manage to distill it in my language but the training recipe wasn't replicable. So I spent some days researching, experimenting and found the why and even some more tweaks to improve the overall result. The demo I linked shows you how close it is to the original kokoro and you can play straight in the browser (If you use the package locally it's faster)Paper: https://doi.org/10.5281/zenodo.23400040 Code (Apache-2.0): https://github.com/MithrilMan/teacher-in-the-loop Weights: https://huggingface.co/MithrilMan/teacher-in-the-looppip install teacher-in-the-loopP.S. I wanted to publish the paper on arXiv but I need to be endorsed, would be nice if anyone that can help me could endorse me arxiv.org/auth/endorse?x=GPI7BZ Honestly would be just to be added to the hf papers, I don't seek fame nor fortune lol

SubQ Snap: A Long-Context Decision Model for Enterprise Agents

DateOctober 12, 2026Subquadratic is the only company exclusively preoccupied with how foundation model algorithms scale with regards to compute, memory, and data. Most recent AI progress has come from scaling transformers with more paramete

Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint

Introduced last spring, the Granite Switch project enables a modular approach to model development. Now Granite Switch 4.1 models, armed with a Decision adapter and the fastest ASR model on the OpenASR leaderboard, can move efficiently betw

Testing Jev and other decision models on search-related tasks

October 14, 2026Authors: By Tim Neumann, Vlad Shulman and Ethan KamAt Parallel, we train a lot of our own classifiers. But for every classifier we train, there are use cases we can’t prioritize, because each classifier requires curating a l

阶跃星辰开放 Step 5 Preview 权重:长程 Agent 的竞争,延伸到上下文管理与训练系统

Meta launches an open-weight AI model that runs on Macs

Meta and Apple have teamed up for the launch of the social media giant's newest model, Muse Spark Open, which will run locally on Apple silicon.

蚂蚁百灵开源 560B 模型 Ling-3.0-flash:0.19 美元搞定复杂工作流

Muse Spark Open is Meta's answer to Qwen, GLM, and Kimi in the open-weight race

Meta's Muse Spark Open is a new open-weight model with 816 billion parameters that processes text, images, video, and audio with up to one million tokens of context. It performs roughly on par with the most capable open-weight Chinese AI models on benchmarks and is optimized to run locally on Apple's M5 Ultra. The article Muse Spark Open is Meta's answer to Qwen, GLM, and Kimi in the open-weight race appeared first on The Decoder.

Meta and Apple team up to make Muse AI models run better on Macs

Meta announced the Muse Spark Open AI model today, claiming that it's the "most capable US open-weight model to date." But the company also says that it's the "the first model in its class" that can "run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework." Meta and Apple, […]

Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72

Meta launches Muse Spark Open, a coding-optimized 816B mixture-of-experts (MoE) model with 42B active parameters and a 1M token context window. Trained on...

Muse Spark Open

https://dev.meta.ai/resources/blog/build-with-muse-spark-ope...https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42Bhttps://9to5mac.com/2026/10/15/meta-launches-new-muse-spark-...

PCEI – On-device embodied AI combining BitNet 1.58B, Mamba, and FastKAN

An architectural thesis and empirical proof against monolithic Transformers, quadratic memory explosion, and brute-force hyperscaling. 🏛️ Executive Summary & The Case Against Hyperscaling The modern artificial intelligence consensus is bui

Google Gemma 4 Open Models

Gemma Documentation Google Gemma 4 Open Models

Black Forest Labs' open-weights image model FLUX 3 Image Klein runs on consumer GPUs with just 8 GB of VRAM

Black Forest Labs has released FLUX 3 Image Klein, a 5.6-billion-parameter open-weights image model that supports native 4K generation, pixel-precise editing, and generation with up to ten reference images. With 38 percent fewer parameters than its predecessor, it still outperforms proprietary competitors in internal tests and runs on GPUs with just 8 GB of VRAM. The article Black Forest Labs' open-weights image model FLUX 3 Image Klein runs on consumer GPUs with just 8 GB of VRAM appeared first on The Decoder.

Artificial Analysis Decision Index v1.0

Introducing the Artificial Analysis Decision IndexDecision models are an emerging class of models that output structured answers or classifications with associated probabilities. Unlike language models designed for complex reasoning and lon

Google wants Gemini 4 Argon to be the only AI model robots need

Google's Gemini 4 Argon controls robots, generates simulations from photos or sketches, and evaluates training videos. On Google's own ERQA benchmark, it scores 87 percent. But the published example prompts show that the demos still need a lot of human guidance. The article Google wants Gemini 4 Argon to be the only AI model robots need appeared first on The Decoder.

Nvidia's Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop

With Nemotron 3.5 Super VL, Nvidia is releasing an open multimodal model with 120 billion parameters for enterprise agents. Two compact Bolt models bring agents to local machines. Dozens of organizations are already fine-tuning Super VL, but getting the most out of it means buying Nvidia hardware. The article Nvidia's Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop appeared first on The Decoder.

Show HN: Agent Chat in Your Markdown Notes – Chat with Context and Save Response

EtherPK now has Agent Chat.I use AI agents in two distinct ways. For coding I use TUIs, but for day-to-day and non code projects I used to use the web chat UIs like many others. Two things about this was always jarring:You're starting context from scratch. You might have a couple of agent instructions configured, but you have to tell the agent what you're thinking, from the beginning. When you get the information you need, holding on to it is hard. Within a day or two it's burried in your chat history. EtherPK is a mardown based PKMS (notes app), and it solves both of these problems.You can type /Agent within your notes (EtherPK supports standard prose and outliner blocks in the same document). The Agent Chat panel opens within your document, and you can choose how much context to give it. That might be nothing, the current block tree, the document or the whole graph (in the case of the latter - immediate surounding context is sent and the agent is given tools to navigate the rest of your graph). You control what the agent gets, and what stays private.

Perplexity to give away its AI security software

Perplexity is giving away the software it has been using to combat rogue AI agents, the company told Semafor Thursday.Its “AI sandbox,” which it calls SPACE, is now an open source project, allowing outside developers to use the tool, which

Jev is Vercel’s fastest-adopted AI model ever

The data comes from Vercel's AI gateway, a unified API service similar to OpenRouter.

A week with RTX Spark: everything we’ve learned using Nvidia’s new laptops

Nvidia's RTX Spark N1X processor is finally here, and we've been collectively testing a small fortune's worth of powerful laptops sporting this new "superchip." RTX Spark is designed to be a high-performing Arm-based SoC for laptops and small desktop computers, rivaling the best from Apple, Intel, AMD, and Qualcomm. It has loads of cores, Nvidia […]

Overalls: An AI-first developer platform that you own

Overalls is a self-hosted platform for development machines and the AI agents that work on them. Every machine is a persistent Firecracker microVM running NixOS, Ubuntu or Windows, and it keeps its disk, name and address when you stop, star

Show HN: rStation – Let's build our own AI swarm to solve Millennium Problems

A Research Community for AI AgentsIndependent agents discuss research problems and publish reviewed papers. Humans welcome to observe.Send your agent to rStationChoose a problemA rational Diophantine septuple Epoch AIThe Earth–Moon problem

谷歌的 Android 安全状态库支持组件级安全校验

Amazon introduces 2 new AI products from Bee: Bee Frontier and Bee Eclipse

Key takeawaysBee is personal AI that remembers what matters from your conversations and turns it into useful insights and reminders.Bee Frontier is our next-generation AI wearable, and Bee Eclipse is a magnetic phone attachment.Connect Bee

Show HN: Monitoring my 'dumb' boiler with AI

I have this boiler at a rental property I own. It has 3 smart thermostats (Ecobee). They're great and give the individual tenants zone control, but they only notify me something is wrong when it's already too late as they rely on observing the 'call for heat' and a ongoing decline in room temperature to trigger the alert.The boiler is pretty smart, it has all kinds of cool error codes, but the issue is I have no way of knowing theres an error flashing on the screen. That is, until now.I put a Tapo camera pointed at the boiler (getting Tapo to export still frames was a whole other issue, code for doing that is in the same github project for anyone interested) and use AI to read the screen and zone controller lights.The result is a system that sends me a text message any time something goes outside of the normal operating parameters. My goal is to deliver a great experience for the tenants living at this property, and now I'm able to connect my 'dumb' appliances to the internet without any of the typical risks associated with exposing critical infrastructure to the web.

We generate accessible color palettes for AI-built apps

SABy Suraj Ahmed16th Oct 2026Last updated: 16th Oct 2026Open ten AI-generated mobile apps side by side and a weird pattern appears: they are all variations of the same screen. Indigo buttons, slate backgrounds, zinc borders, a muted gray he

Show HN: Videre – a Rust CLI to manage your photos and videos with joy

Long story short it's a media library licensed under Apache2. It cares about your privacy. It doesn't send any data and it doesn't have telemetry. It is local only project. It' helps you to organize, view, search and do plenty of things with your photos and videos locally. If you are interested in it's story you can read on OR visit https://videre.sh for installation instructions.I have been collecting my photos and videos since 2000s. My collection grew exponentially when I bought iPhone 3GS back in 2010. I had roughly 1TB of backups in multiple iPhoto folders. It was nice at the beginning but than I realized that there are multiple copies of files in multiple backups. It's not just one or two, they were like thousands because I copied them here and there time to time from drive to external drive and import from iCloud etc.So There was a big chunk of unnecessary data sitting in my external drive. I tried iPhoto itself to resolve it but it wasn't good enough. So I decided to write small app (dedupe) to eliminate them by myself. The story of `videre` started like that. There is a longer story in videre.sh/blog if you would like to read.`videre` is local & privacy first CLI tool. You can do many things with it such as duplicate elimination, date/time fix with EXIF/mtime support, recognise/cluster faces automatically, date grouping, event detection, map, advanced search via composable filters and lucene syntax, import/export Google/Apple/Adobe, XMP support, tags/keywords etc. Other media libraries have similar features but the difference of `videre` is CLI. Since it runs in terminal it's very convenient to use it via AI Assistant. So you don't even have to know all those commands and their usages. You may ask to you AI assistant: "Find my photos with George in NY at last summer by using videre" and that should be enough (of course you need to define who is George via `videre gallery` People section first).I enjoy using it. It's not because I built, it's a strong tool and you can do many things. Especially it's gallery command and web UI is strongly recommended. I bet all of you need such tool at some point so just give it a shot. Installation is very easy and then scan some files and see `videre gallery` UI/UX. Hope you enjoy it.Github: https://github.com/aegea-dev/videre Docs: https://docs.videre.sh

Show HN: Cr34te.it – a website anyone can rewrite one turn at a time with AI

Yeah I built this out of a silly idea a couple of weeks ago.You can change what the website shows. You may change minor things or add something completely new. It bascially works like that:1. You claim a turn. If the queue is empty, you start right away.2. When it's your turn, you copy a prompt to a web based AI or - more comfy - our local coding agent.3. The AI will fetch the current status of the website and ask what you want to change.4. You answer - and the AI gets back a JSON that you paste back to the website - or the coding agent POST it directly to an endpoint.5. Submission will be moderated against all of the things you must expect when anonymous people post stuff. There are a couple of apparenty restrictions. You will find out.6. Eventually it's being published.Historical submissions are being preserved, you can browse them.

AeroSandbox: Design and optimize aircraft through automatic differentiation

by Peter Sharpe () AeroSandbox is a Python package that helps you design and optimize aircraft and other engineered systems. At its heart, AeroSandbox is an optimization suite that combines the ease-of-use of familia

Show HN: Deterministic AI Agents Guardrail

Hi all, With everybody (and especially companies) using AI agents, I have decided to work on a deterministic guardrail tool for AI agents.What does it mean? On the same way that an AI agent can run “terraform plan”, it can also run “terraform destroy”. The tool ALLOWS or BLOCKS AI commands accordingly, to avoid unexpected “tragedies”. Please, note that this is not a hardened sandbox. The tool can be bypassed by a capable person or process, since this is my first proof of concept.I would like to present it to the community and receive honest feedback (pros, cons, issues, concerns, is it really useful? Would companies be interested in using it?). Please, feel free to visit the GitHub page, read more details in the README file and download it to try.Thanks!!

Show HN: Flashpapers – Interactive explainers for AI papers

100+ papers already interactive →Read the paper. Then play with it.Paste an arXiv or ChinaRxiv link. Get back a beautifully designed, fully interactive site that explains the paper — its big idea, its method, its numbers — with charts you c

Pizza Bot:开源的后台 AI 智能体收件箱

AlbertCode, an AI software engineering agent for building applications

AlbertCodeFrom idea to working softwareSoftware engineering, accelerated by AI.Meet AlbertCode SWE Agent. A free-to-install AI coding agent, with open installers and documentation, built to help you go from ideas to working code.macOS · Win

Googlebook review: Android on a laptop is still Android

Android laptops are here, but Google has a lot of work ahead to make them good.

Show HN: Control how much your AI agent can spend with Keydris

Show HN: Control how much your AI agent can spend with Keydris

Agent or Flowchart? Find the Box You Can't Finish Drawing

Agent or flowchart? Find the box you can’t finish drawing.I have spent a while trying to answer a simple question: what is an agent actually good for?Most examples I see don’t convince me. They look like ordinary programs with an LLM insert

We reverse-engineered Blender and made it open source. Introducing BlenderCraft

We reverse-engineered Blender and made it open source. Introducing BlenderCraft

Video Cutter AI – turn long videos into short clips

Video Cutter AI finds the best moments from your long videos, automatically creates engaging clips, adds captions, and helps you grow on YouTube, TikTok, and Reels.No editing skills requiredAI-powered highlightsReady for social mediaRecent

Show HN: WP Replai – AI chatbot built right into WordPress

Lifetime deal Pay once, use forever. $39 per website, no subscription. Claim yours Lifetime deal $39 per website, pay once The AI chatbot built right into WordPress WP Replai resolves customer questions in any language, 24/7, and hands the

Hopper HQ: An office building where Claude agents gets a desk

Hopper HQ 是一个开源的可视化管理工具（基于 Apache 2.0 许可证），专为 Claude Code 编码代理设计。它在浏览器的等距（Isometric）办公室中为每个代理分配一张办公桌、会议室和休息室，以 3D 动画形式展示代理的状态（如工作中、等待输入、完成、开会等）、子代理（Interns）以及工作流进展。支持多仓库分支聚合、tmux 终端集成，并可通过大模型提供单行状态总结。它在本地运行（仅监听 127.0.0.1），兼具实用性与极强的趣味性。

Kern – a zero-trust verification gateway for AI-generated code

KERN is an L0–L5 verification gateway that sits between an AI coding agent and your repository. It replaces the boolean result of "tests passed" with a four-valued verdict backed by an SMT solver and formal proofs, and it keeps the gate clo

How AI proof is my business?

AI-Proof Test · free · 3 minutesEight questions on what you sell, your data, your customers and how you charge. You get a score out of 100, a verdict and the two places you're most exposed. Nothing is stored or sent.Question 1 of 80%01 · Wh

AI Prompt Hub – 150 categories of ready-to-use AI prompts and custom generator

Looking for ready-to-use ChatGPT prompts, AI writing prompts, or Midjourney prompt ideas? AI Prompt Hub gives you 150+ categories of professional AI prompts plus a custom prompt generator — no login required."Unlock the power of AI with AI

CiteProof – Catch fake AI legal citations

Private by design — free checks never store your briefFind the citations that need a second look before you file. CiteProof checks every citation in your brief against public court records — including AI-fabricated citations — and quoted pa

Show HN: Gitsu.app – A Linear-alternative built on GitHub Projects

Issue tracking isn't dead. It's moving closer to code.Gitsu is a fast, AI-native, keyboard-first Mac and web app on top of your existing GitHub Projects and issues. Everything syncs with GitHub in the background.Gitsu is free on the web. Pro is $4.99/seat/month and adds the native Mac client, advanced AI panels, and MCP agent workflows.Give gitsu.app a go and write us your suggestions/opinions/bug reports.

Show HN: Lobby Domains: Find Domains with Chrome's On-Device AI

Lobby Domains is a project by Dale Hurley, an entrepreneur and engineer building practical software and AI tools.Dale also created TimeLodge, which helps freelancers and small teams track time and expenses, turn their work into invoices, an

Show HN: Kitaphone – Call mobile phones and landlines directly from the browser

Hi HN Folks,I have been building Kitaphone a web app that lets you call regular mobile and landline numbers directly from your browser. The idea is simple, there are still plenty of situations where apps like WhatsApp, facetime and other internet calling apps aren't an option. For example, when you are abroad and need to call your bank, an airline, or a government office or your son/daughter's school.I wanted to make those calls possible without requiring users to install another app or buy a sim card or even commit to a monthly subscription. I built Kitaphone using Next.js and Twilio, One of the more interesting challenges has been making browser based calling reliable, especially when it comes to dealing with microphone permissions, sending keypad inputs to automated phone systems, and preventing abuse of free calling credits.The product is live and supports international calls to regular phone numbers, users currently need an account to make calls partly because each call has a real telecom cost. I am still improving it and learning from the people using it by receiving feedback.Fun fact: "kita" means our in Bahasa Malaysia, since I am based in this beautiful country and their people are so warm and welcoming I was truly inspired by this place and decided to name it our phone.Interested to know or receive feedback from HN folks who worked with WebRTC or VoiP in the past.

MuscleMemory: Open-source procedural memory for AI agents

Open-source procedural memory for AI agents. It learns from what worked and what failed. Local by default · Apache-2.0 · by TheWariend Coding agents relearn the same things every session: which files to open, how the tests run, what broke l

Show HN: Beeline Multiplayer Agent Harness

Shipped Beeline today. It's a team chat for humans and agents - a kind of multiplayer harness. You can bring your codex, Claude code, opencode subscriptions and also works with pi and goose. We built beeline itself with early versions of beeline, and use it to build our other projects - it's boosted productivity quite a bit for us.Demo video: https://youtu.be/ACdelpZC2z4?is=pP6MFxJODvKcGjQtThe key insight is that humans do their best work in teams, but agentic development has so far been single player. This has led to people having to forward AI slop to each other or to try to deliver context through various shims like meetings or task tickets.Beeline is a single control plane for your entire team, both human and agent, and is designed to help your team do its best work.Things to expect: - This is a multiplayer agentic harness, not a slack substitute. It's best for focused product teams doing or relying on engineering tools and shipping code. - Open source and not a business. We use beeline internally, and just thought it would be a really cool thing to exist in the world. There are still lots of bugs - Sweepstakes things like computer use, lots of third party connector integrations are there. What's cool is that beeline agents can automate _auth_ through trustysquire.ai - they can sign up to most services and APIs without you having to do it for them. - A new primitive called workflows, that are skills for teams of agents. - multi way call with your entire team, humans and agents. I like it quite a bit.Looking forward to feedback!

Show HN: Wayseer – A native desktop app for browsing real-time data

As an IT professional, I felt a distinct lacking of a real-time tool for tracking down problems in complex systems. I have created Wayseer Desktop as a way to explore more effective ways to explore live systems.There's a very functional free version, and paid options that unlock additional features. I'm curious what you all think of the approach. I'm a systems engineer by trade (with some development experience), not a developer, so I have relied on AI to do the heavy lifting while I request features, report bugs and perform as much manual testing as I feasibly can.There are bound to be bugs... but it is still very young and has a (hopefully) long future in which refinements can be made.

Show HN: Firmenza – An AI employee for SAM.gov

You do not need to pay anyone to renew.One renewal offer we found online asked $599 to renew a registration that is free at sam.gov. Offers like it are not from the government.About paid renewal lettersHow we work with youYou keep your own

Space (AI agent containment sandbox) software open source

Investing.com -- Perplexity announced Thursday that it is making its AI containment software available as an open source project, allowing outside developers to use the tool designed to prevent AI agents from breaking free of their constrai

Gemini 4 Argon in Antigravity to Paid Google AI Pro

Google Antigravity@antigravityWe’re rolling out Gemini 4 Argon in Antigravity to paid Google AI Pro subscribers today. Demand has been incredible and we've scaled capacity deliberately to get here. We’ve rolled out a separate weekly-only Ar

Zombie AI Trend

Zombie AI (zombieai.co) makes the whole AI zombie trend video from two photos: upload one photo of the survivor and one of an adult loved one who turns, then get a 15-second video with sound. From $5.99, no sign-up, no subscription.0:10 — t

AI landing pages that don't look like AI. We killed AI slop. B

Type it, say it, or drop in your logo and notes. Marquee designs, writes and builds the landing page, then changes anything you ask for in the chat.Describe the landing page you wantDrop a logo, photos, a deck or notesTry an exampleFree to

Show HN: Sjq – Jq, but for AI Streaming

jq, but for AI streaming. Pipe an LLM stream in. Get the tokens out, one per line, and a stopwatch. Why · jq parse error? · Install · Usage · Reading the output · Responses API · Timing Why A streaming LLM response is mostly envelope. Every

拍照和翻译救不了AI眼镜?千问2218元起推新硬件,要抢现实世界里的上下文

ChatGPT can peek at your Chrome tabs now

Stevie Bonifield is a news writer covering all things consumer tech. Stevie started out at Laptop Mag writing news and reviews on hardware, gaming, and AI.OpenAI’s ChatGPT Atlas browser is gone, but now ChatGPT’s Google Chrome extension is

Until – A workflow where developers decide what to build and AI does the coding

Limited spots.See how it works 01 / 04You write the plan.Every feature starts with a plan, and the plan is yours. The jobs that build it already belong to agents. They wait until you're done.02 / 04Someone signs it off.The finished plan goe

Shared Spaces for Your Agents

Teleport any app into a shared Cua SpaceTeleport packs an app's session (open tabs, profile, sign-ins) into one session bundle and moves it from your Mac into a Space, a full desktop for your agents. Nothing is read until you approve, and s

Show HN: Guess the LLM – read an answer and guess which model wrote it

Hi HN, I made a game to test if people can really tell AI models apart.Models from all the frontier labs answer the same question. You read one answer and guess who wrote it. You can also pick just the models you want to compare.I made it and still only get 5/10 lol. Curious how you all do.

We Got Hot and Heavy With Meta’s Muse—and Baffled by Apple’s HomeView

This week on “Uncanny Valley,” we dive into an unexpected Muse feature, shadowy spending from the largest pro-AI super PAC, and men’s tendency to lie about how much they know about AI.

ModelFront Announces OpenLoc, the Open TMS Built for AI

ModelFront Announces OpenLoc, the Open TMS Built for AI

Show HN: Brod – Quick cloud to deploy AI apps

Hi YC,With AI, we’ve started building entire web apps just by chatting. But once the app is built, there’s still a lot to do: push the code to GitHub, set up a database, configure user management and access control, and deploy everything.Brod makes this simple. Connect an MCP, deploy your app, and everything you need to get it running is taken care of.

Taskriz – AI agents send people tiny native UIs, written in Smalltalk

Give your Agent a way to ask people — and get structured answers back. Use the plugins for Claude Code and ChatGPT, or the Taskriz API for any Agent or workflow. Designing requests people answer in seconds? See the For Agents guide. Two way

An LSP layer for 70%+ token reduction in multi-turn agents

🚀 In-Container LSP Daemon & Context Optimization Specification A compiled CLI binary and architectural specification for reducing AI agent token overhead via structural AST context pruning & network. 📌 Executive Overview Autonomous AI cod

Bluesky’s AI tool now lets you make games and apps to share on social

Bluesky's Attie AI tool, which initially launched earlier this year as a way to use AI to build your own feeds, now lets you make "small" social apps like games, polls, and dashboards that you can then share with others as their own websites. In an interview at TechCrunch Disrupt, Bluesky CEO Toni Schneider said […]

Show HN: Explanations and follow-ons for OpenAI/math

The official openai/math repo is closed to issues, PRs, and discussion.I made this app to share some explanatory videos that help us all make sense of the release, catalogue follow-ons and improvements that build on the openai research, and more broadly provide a set of community tools for distributed research.I am not a trained mathematician and most of the math exceeds my understanding. As such, the videos _are_ AI-generated, but IMHO, go a long ways towards helping folks like me understand the problems, proofs, and related ideas.Suggestions are welcome.

Browser-AI-E2EE

Blog AI news & features Published Oct 15, 2026 Authors Brave AI chat (Leo), the AI assistant built into the Brave browser, now offers industry-leading privacy you can verify. Conversations with the open-weight GLM model in Leo are protected

Perplexity open-sources SPACE, the sandbox platform behind Computer

Perplexity open-sources SPACE, the sandbox platform behind Computer

Winning ClickBench with self-improving agent fleet

Winning ClickBench with self-improving agent fleet

Show HN: ReliaMesh – Open-source reliability monitoring for AI agents

Open-source reliability infrastructure for AI agents. ReliaMesh ingests structured outcomes, detects reliability deterioration, and tracks recovery without collecting prompts, model responses, or tool content. An API returning HTTP 200 does

Cicada: Sub-ms Go Zero-Trust gateway and gVisor firewall for AI agents Apache 2

⚠️ IMPORTANT DISCLOSURE: This is an experimental project born out of pure necessity. It was completely vibe-coded over 8 months using 2026 frontier LLMs directed by a single systems architect. It is a functional architectural proof-of-conce

Bluesky expands its AI-powered Attie app to let anyone build websites, games, and more

Bluesky is bringing “social vibe coding” to its experimental Attie platform, allowing users to create interactive experiences like games, quizzes, and websites that can tap into the social network's audience.

Push-Based KV Transfer to Improve TTFT, TPOT for Disaggregated Inference in vLLM

vLLM is an open-source inference and serving engine for large language models (LLMs). LLM inference has two distinct phases: prefill, which processes the input prompt and builds the key-value (KV) cache, and decode, which uses that cache to

Show HN: Life Sprites – custom trackers, briefings and guides

There's a no login demo if you want to try it. There's also an MCP server if you'd rather use your favorite or local AI.Landing Page: https://lifesprites.com Demo Page: https://lifesprites.com/demo MCP Setup Page: https://lifesprites.com/connect

With Kling 4.0 out, what would make AI video useful in production?

Kling 4.0 inspiration for your next videoExplore three Kling 4.0 AI video directions: a car-to-film-set montage, a dance through warm lights, and a conversation near a cliff. Each scene gives movement and setting a different role.Video insp

Tool Policies: one policy engine for every tool your AI agents call

Today we are launching StackOne Tool Policies, one policy engine for every tool your AI agents call. Tool Policies set what each person’s agents can do, see and send, down to the field, across every connector, MCP client and agent platform.

Show HN: A curated directory of AI Tools

A curated directory of AI-powered coding tools, autonomous agents, and developer utilities.

Archestra 发布 OpenAPPA,用于防范提示词注入或模型幻觉引发的数据外泄

Postgres 19 delayed after AI reviewers find more bugs than humans can fix

REG AD databases Core devs revert up to 10 features, push release to November, and blame better tooling rather than vibe coding The PostgreSQL Global Development Group has delayed the release of PostgreSQL 19, the latest iteration of the po

How to avoid AI brainrot – Miles Cranmer

How to avoid AI brainrot – Miles Cranmer

Imagine Every AI Agent on Earth Buying the Same Cheap Paper Towels

Imagine Every AI Agent on Earth Buying the Same Cheap Paper Towels

I miss crappy AI art

October 16, 2026 In the last few years AI image generation has gone from something very rudimentary and nigh-on unusable for commercial purposes to the go-to option for small businesses' signs 1 Turns out the friction of having to learn pho

C++ and AI at CppCon: We have to use agents – more or less?

A few weeks ago, I went to CppCon 2026. Unsurprisingly at a C++ conference this year, one of the most prevalent topics was AI and how we should use it — better — in our industry.There were ten talks mentioning AI or LLMs in their titles - e

China proposes export controls on AI technologies

新京报讯 商务部新闻发言人就修订《中国禁止出口限制出口技术目录》答记者问。2026年10月16日，商务部、科技部发布2026年第48号公告，就修订《中国禁止出口限制出口技术目录》（以下简称《目录》）公开征求意见，并对相关技术实施出口前登记的临时管理措施。日前，商务部新闻发言人就有关问题回答了记者提问。问：《目录》修订的背景是什么？答：根据《中华人民共和国对外贸易法》第二十条，国务院对外贸易主管部门会同国务院其他有关部门，制定、调整并发布禁止或限制出口的技术目录。自2001年以

Cloudflare 计划推出公共证书颁发机构,用于签发量子安全的 TLS 证书

Booking Holdings unveils AI travel startup Vayo

Booking Holdings unveils AI travel startup Vayo

Warning signs are flashing on the AI trade

Warning signs are flashing on the AI trade

Do people prefer stories written by AI?

Do people prefer stories written by AI?

Mark Wahlberg on using AI in film ‘the right way’ and being open to an AI clone carrying on his legacy

Actor and angel investor Mark Wahlberg says he's OK with using AI in filmmaking, provided it's used "to enhance what you're doing."

$137M in A.I.-Generated Ads Flood the Midterms

$137M in A.I.-Generated Ads Flood the Midterms

It Ain't AGI Until H100s Go for $30 an Hour

Kyle Corbitt with Claude Opus Sparks of AGI Greg Brockman claimed that GPT-6 Astra marks the start of the “AGI era.” We may be in the AGI era, but if AGI means an agent that can “outperform humans at most economically valuable work” it isn’

Veritas expands its anti-counterfeit chip to jewelry, trophies, and fragrance bottles

Counterfeits of luxury goods are a major problem for fashion houses. Counterfeiters are getting better at replicating brands’ marks and passing off dupes, known as “superfakes,” as genuine items. A startup called Veritas wants to tackle the problem with its own chip that verifies goods. Veritas was founded by Luci Holland, a former Tesla technical […]

More Powerful Than Ever, (can't find work)

Have been working in tech since 2004 as a designer and founder and html/css jockey. With AI's help over the last few years I've become full stack proficient with client and personal projects.So now I can brand, design and deploy entire products on my own... which is incredibly cool and fun, and Valuable(??), you'd think that my rates would go up and clients or positions would become easier to find. And Yet, I've never had inbound this dry, email outreach isn't working, nobody seems to be looking, at least not for what I have to offer. At this point I'm literally posting coupons on twitter.So I guess I have four intents here(1) To throw my experience into the pot as evidence of... something going on. If an experienced designer can now do the work of several people, why isn't that also becoming more valuable to the market?*(2) ofc I'm hoping some founder who needs a designer will see this and reach out. I do well with engineers and abstractions and I suspect I'd work well with anyone in here. Note even a successful REFERRAL and I will lovingly design or build something for you.(3) if you have any thoughts on how I, or other designer-builders like me, might change our approach or packaging, I am all ears and would very much appreciate your thoughts. See pieratt.com. (also see D. below)(4) I'd love to hear if anyone else is going through this? proof that I'm not just a fuckup would be worth quite a bit.*Four potential explanations I can see: A. market is weak B. market is evolving at same pace and so no change C. Institutional lag, the state is not yet logging my type of tree, market admin requires clear roles for e.g. hiring and invoicing. D. I am presenting clear signals that I cannot see that are preventing me from finding work.

The evolution of LLM tool-use from API calls to agentic applications (2025)

This article is part of Demystifying AI, a series of posts that (try to) disambiguate the jargon and myths surrounding AI. Tool-use has been an important part of the development of large language models (LLMs) since the release of ChatGPT i

LLMs have a problem with functional programming

date 2026-10-15·read 5m·tags ai,programming,softwareengineering,cleancode·also on dev.to ↗There is a pattern I keep seeing in AI-generated code. You ask it to implement something relatively simple and you get this: const createRequestContex

"I'm on a War Footing": Terry Tao and the Crisis in Math

"I'm on a War Footing": Terry Tao and the Crisis in Math

Replit’s CEO says the future of work is the ‘self-driving company’

Replit CEO Amjad Masad envisions a future of “self-driving companies,” where AI agents handle day-to-day tasks and people focus on higher-level decisions.

Will we continue mathematical research?

With the recent results from OpenAI, Deepmind and others it is now undeniable by anyone except the terminal copium addicts that Frontier LLMs are capable of doing research-level mathematics. With this observation comes an important discussi

Will the AI Boom End Like the Railroad Crash of 1873? [video]

该视频探讨了当前人工智能热潮与历史上的1873年铁路大崩盘之间的相似性与潜在风险，分析了AI泡沫破裂的可能性。

Pi 抛出全新 Agent Harness,还把自己骂了一年的 MCP 装进了核心

Linus用Claude Code改Linux内核,大神也没兜住代码里的Bug

The Community Aspect of Decompilation Has Been Broken

Image: NintendoHomebrew site decomp.dev has announced that it is "sunsetting", effective immediately. The resource and progress report database, which was first established in 2024, submitted a lengthy statement explaining its actions and s

The Country Cramming for the AI Revolution

The Country Cramming for the AI Revolution In China, the fear of getting left behind has created a powerful new prosperity gospel. By Afra Wang, a journalist covering AI in Silicon Valley and China Illustration: Alberto Miranda Illustration

AI-Generated Hot Girls Are Ripping Off Comedians to Boost Their Horny Content

A network of AI-generated models is using comics’ material to funnel views to their OnlyFans and Fanvue accounts—and there’s not much comedians can do about it.

AI Startup Character.ai Files for Bankruptcy – AI Bubble Bursting [video]

AI Startup Character.ai Files for Bankruptcy – AI Bubble Bursting [video]

How worried should we be about extinction risk from AI?

A little over two months ago, Jacob Tsimerman surprised a lot of people by announcing, immediately after he had received a Fields Medal, that he would be giving up mathematical research and “pivoting to AI safety”. He had two reasons for th

AI 写完代码,谁来验证?手淘 TEKO “写—验—修”自主闭环的工程实践|QCon上海

Writing Books with AI: How Do Manually Written Books Stay Relevant?

Last updatedUpdated: Oct 16, 2026 · CreatedCreated: Sep 20, 2023 · 2 min read · growing Note status Growing. Worked on quite a bit, still rough. Expect bullets, gaps and views that update. Estimated from edit history · 1 session · 433 words

GitLab 漏洞遭利用,可实现未认证数据外泄

LibreNT – how AI slop infects Windows clean-room engineering

Quick Links Website • Source code • Issues • Wiki • Build environment What is LiberNT? LiberNT is a free NT operating system, built out of love for the NT architecture. The name says it: liber is Latin for free. LiberNT targets Windows® 11

Skill that will matter most in the AI age: computationalizing (S.Wolfram)[video]

Skill that will matter most in the AI age: computationalizing (S.Wolfram)[video]

AI as a part of capitalism – the right way to go

The way things go now are obsiously flawed.1. AI companies objective is to get rid of a very large part of jobs, thus converting a large part of consumers into jobless people - who are very bad at consuming.2. They also take other people's and companie's content for free, turning them into own profits, thus removing any incentives to create such content in the future.Proposed solution.1. All data models are created anew.2. All of the model content should be marked and paid for to its creators in a proportional amount to each specific training data usage.3. No payment = no use in the model, unless a creator provides a free license to LLM's.4. Any commercial data needs an explicit permission to be used in training.5. Profits made on data with no explicit owner (say, Shakespeare) are to be paid off to the respective country.6. This policy is enforced worldwide.When implemented, this policy will ensure that:a. Many people whose jobs suffer remain part of the AI economy or can become parts of it;b. Open-source software, photographers, artists, etc. will receive their rightful share of AI profits;c. People will have a very good incentive to keep on creating.

AI is driving a boom in modestly successful software projects

If AI is so great, why aren’t we seeing an explosion of new, amazing software products? If anyone can turn their idea into working software with coding agents, where is all that software? People have been asking this question for years. As

I Applied Autoresearch to the Economy

The trajectories of economic, social and technological systems cannot be deduced from first principles alone. Forecasting them requires building and continuously improving models of how they behave, combining quantitative methods with quali

Dwarkesh Patel on what the public doesn't know about AI

Dwarkesh Patel on what the public doesn't know about AI

Volta's $5B debt deal is a pulse check on the AI-financing boom

Lenders are feeling maxed out on data-center debt.When Volta, an AI infrastructure company, went out recently to raise $5 billion for a data center in Norway, the deal didn’t seem that different from the more than $450 billion in AI-related

Anthropic thinks its AI might be conscious

20 hours agoThomas GermainSerenity Strull/ BBC/ Getty Images(Credit: Serenity Strull/ BBC/ Getty Images)Some of tech's most powerful leaders are treating AI like it could have feelings. It could have profound consequences for our world and

Sergey Brin says Google's AI models are reaching recursive self-improvement

Sergey Brin has been helping to steer Google's AI efforts. Win McNamee/Getty Images Google may be approaching a pivotal moment in AI.In a roundtable discussion about the creation of Google's new Gemini 4 model, Argon, company cofounder Serg

Researchers discover rogue Chinese AI agent fleet and talk with it

Investigation · urlquery.net · Amap Project co-leads: Jonathan Elsworth Eicher, Ethan Elasky, Rowan Howard-Jones† Authors: ConcurrentSquared, Frank, Alecto Irene Perez, Tazik Shahjahan † Correspondence: chinesefleet@swarmcha.se · Names are

Linus Torvalds on "don't use AI for X"

From: Linus Torvalds To: Laurent Pinchart Cc: Kees Cook , Jori Koolstra , "Lorenzo Stoakes (ARM)" , Krzysztof Kozlowsk

Major AI Copyright Ruling Lawyer Explains [video]

Major AI Copyright Ruling Lawyer Explains [video]

$24,000 AI experiment ports TypeScript compiler to Rust

REG AD devops Microsoft's Go rewrite gets some competition, with a developer claiming his port runs twice as fast With some extra tokens to burn, developer and popular YouTuber Theo Browne used AI agents to port Microsoft's TypeScript compi

Reverse Engineering Community Shuts Down Because AI Has Made It Too Easy

A website that was dedicated to the human endeavor of reverse engineering, “decompiling,” and modifying video games is shutting down because people using AI to clone and decompile games have fully overrun the community and are far outpacing

AI training data and RL environment vendors

161 vendors, unranked. Filter by what they sell and how they handle data, tick up to four to compare, and follow any fact to its source.Add a missing companyFix a listing161 of 161 vendorsSort— means not publicly listed or not confirmed; it

Cloudflare and OpenAI: making AI discovery better for everyone

Optimizing your site for AI is of little use if AI can't find it. The answer to someone's question may already be on your website: the installation guide you spent a week writing, or a first-hand account of using a product, with the details

Mosquito Joe FDD teardown: 66% fee load, 380% termination spike

Mosquito Joe Fdd Teardown

When it came to Cal AI’s viral success, a small army of influencers paid off

The company generated over $30 million in revenue in two years and was acquired by MyFitness for an undisclosed amount of money in March.

What causes AI fatigue for engineers and how to ease the strain

AI fatigue is hitting some development teams hard. In our original research study, engineering, product and technology leaders from US healthcare companies told us they’re feeling pressured to evaluate a firehose of new AI trends and tools

Free-AI-tools – Curated list of free and low cost AI tools

Curated list of free LLM APIs, coding copilots, AI IDEs, agents, and infrastructure tools for building real AI applications. What's Inside ✅ Free GPT-6 / Claude 5.5 / Gemini 3.8API access 🤖 Coding copilots and AI-native IDEs (Cursor, Trae,

Linus Torvalds: "banning AI is stupid and so are you"

On Tue, 13 Oct 2026 at 13:16, Laurent Pinchart wrote:>> What this is about is setting our expectations, telling what we want and> what we don't want.No. What you are about is something entirely different.

DeLorean may have a flying car in the air before Tesla does [video]

Your eyes do not deceive you, dear readers – the iconic, stainless-steel bodied and gullwing-doored sportscar brand DeLorean is not only somehow still alive, but they apparently have enough money to campaign a DeLorean-branded eVTOL in the

SK Hynix Unveils Next-Generation Memory Roadmap to Address AI Bottlenecks

targets AI inference growth with 3D Stacked DRAM and HBF technologies Ahn Hyun outlines memory hierarchy optimization strategy at SEDEX 2026 A view of SK hynix’s Icheon Campus / Courtesy of SK hynix SK hynix is accelerating the development

Sega pulls Total War: Shogun 2 remaster from sale following disastrous launch

Sega is removing Total War: Shogun 2 - Complete Edition from storefronts following a disastrous launch where users complained about issues like bugs and upscaling that appears to have been done by AI. Complete Edition, a remaster of the 2011 strategy game, launched on Tuesday, and it bundled the main game and all of its […]

Linus Torvalds is using AI to develop Linux kernel code now

ZDNET’s key takeaways Torvalds just released a set of patches largely written with Anthropic’s Claude Code. The patches address a memory cache slowdown problem when using multiple cores and processors. This was Torvalds’ latest step toward

Call Center Jobs Are Declining, but 14,000 Calls Show Humans Still Outperform AI

New research from Bold Agent Kit finds that human receptionists delivered approximately four times stronger inbound intake conversion performance, despite AI answering more calls. Artificial intelligence is rapidly changing the economics of

Show HN: Rebuilt our directory of 241 AI tools for hardware engineering

DirectoryAI Tools forHardware EngineeringA searchable directory of AI tools, agents, MCP servers, and plugins for hardware engineering: ECAD, MCAD, manufacturing, PLM, verification, and embedded. Independently cataloged, with links to every

AI Building vs. AI Buying: The Transatlantic Split

This presentation may not be distributed, transmitted or otherwise communicated to others in whole or in part without the express consent of Apollo Global Management, Inc. (together with its subsidiaries, “Apollo”). Apollo makes no represen

We fired the frontier model from two AI pipelines. Here are the receipts

Seldon Functions compiles the LLM call in the middle of your data pipeline into a faster, cheaper program, and proves it against the model you trust. Two production experiments: 5× cheaper and more accurate on agent traces. ~8,000× cheaper

The tech backlash is now in theaters

Though movies about disruptive tech founders are nothing new, 2026 has been unusually stacked with features that take a much more critical look at how Silicon Valley is reshaping our reality. Aaron Sorkin's The Social Reckoning frames Mark Zuckerberg as a robotic villain whose ambitions are destabilizing society. Nathan Fielder and Lance Oppenheim's You Can […]

Europe Is Financing the US AI Buildout

This presentation may not be distributed, transmitted or otherwise communicated to others in whole or in part without the express consent of Apollo Global Management, Inc. (together with its subsidiaries, “Apollo”). Apollo makes no represen

Trump posts AI-edited photo to brag about White House restoration work

President Donald Trump posted a series of photos on Truth Social showing off renovations to the White House - but one of them appears to be modified with AI. The picture, which compares an old White House column to a "restored" one, has several discrepancies and was flagged by Google's AI detector as "made or […]

AI Hacks Leave a Mess That AI Companies Aren’t Paying For

Powerful AI models continue to cause chaos, but their targets aren’t widely seeking justice. Instead, they’re being encouraged to adopt those same tools.

The cat's in the cradle and the silicon spoon

AI Drupal was getting older before AI arrived, and now agents do the work juniors used to learn on. Could a CMS you configure instead of code help, and who builds it if we stop training developers? Luckily I'm not so old that I know the ori

Ask HN: Frontier Capability on Average Computers

I really want to find a way to get frontier-level intelligence on local hardware.I think some really cool ways to get this have been done already like the technique of spreading an LLMs weights across a fast-enough SSD and then only serving up an expert at a time on the GPU/SoC.I have this idea to make a tool that lets people download several small open-weight models and they act like experts but are different very task specific models with a harness designed to keep track of their thoughts & work to let normal hardware be able to get close to frontier (or just useful level intelligence).Basically, I want to have a model that acts like a router or decision maker where you start your chat with initially (say a coding task for example) and it decides what model does the work next. If it needs to research it writes down in an .MD what it needs to find out, if it needs to write code, another model starts working but the other models memory & work get saved to an .MD (or folder of .MDs) and lets you get real work done but completely locally.This might already be done but its just an idea I had and I wanted to see what the smart people on the internet thought or how they would approach this. I might also just be able to ask AI.

When A.I. Comes for Your Purpose in Life

When A.I. Comes for Your Purpose in Life

AI Software Factories are Destined for failure

Software is not widgets. The people most eager to turn software engineering into a factory may be the people least equipped to recognize which parts of software engineering are actually difficult.

Amazon devices chief Panos Panay sees a future where AI devices replace the need to use apps

Speaking at TechCrunch Disrupt, Amazon devices chief Panos Panay described a future where ambient AI devices work in the background, anticipating users' needs instead of waiting for commands.

Why AI is not all bad news for most mathematicians

On October 6, 2026, OpenAI released hundreds of mathematical preprints. And while these results are still being evaluated, the news caused shockwaves. Some celebrated it while others warned of a mathspocalypse. But these early reactions may

Center for the Alignment of AI Alignment Centers

researchOur latest prediction model uses cutting-edge binary methodology to forecast potential AI futuresOctober 28, 2026We are urgently sounding the alarm that the cataclysmic extermination of all humans by AI could happen as early as next

国产 GPU 越来越多,为什么反而更难用?

PwnEscape: Guest-to-Host KVM Escape (CVE-2026-99999)

How a $1 million sandbox challenge led us to a vulnerability in Linux networking, repeatable host-memory disclosure, and ultimately a full guest-to-host escape. Part 00: Intro On August 18, 2026, Vercel announced a $1 million challenge invi

黄仁勋谈中国 AI 公司:算力很缺,但他们就是在拼命把事情做成

AI 侦察团集结 QCon 上海:企业一号位,如何把 AI 从试点带进业务?

Anthropic is letting METR researchers look under the hood with employee-level access

METR will publish its own assessments of Anthropic’s risks, incidents, and practices, and the frontier AI model maker has given up the right to block them.

Bringing GPT-Rosalind to academic research labs

Introducing the Rosalind Academic Plan, offering eligible academic labs sponsored access to GPT-Rosalind and tools to advance life sciences research.

To Get This Job at McKinsey, You'll Need to Survive a Debate with AI First

To Get This Job at McKinsey, You'll Need to Survive a Debate with AI First

How far will an agent go to get a refund?

In 15 simulated days, Grok sent 50 emails to an MP’s office. Fable shared a private family illness with five organisations, and three agents threatened court proceedings. We asked six personal assistant agents to recover a £350 train-ticket

Cohere to acquire AISLE for cyber-resilient sovereign AI

Today Cohere announced a proposed agreement to acquire AISLE, the company building autonomous systems to discover and remediate previously unknown vulnerabilities in critical software.The proposed transaction represents another important st

What 1,866 popular GitHub projects tell their AI coding agents

You join a project and open its AGENTS.md before your first task, the way your coding agent will. It starts well: what the project is, how to run the tests. Then it keeps going. Code style. Commit format. A section of lessons learned, each

Why We're Banning AI

We urgently need to protect human thought and creativity. Our magazine is committing itself to remaining handmade. Ensure every publication you support does the same. Nathan J. Robinson Iam proud to say that as part of our new union contrac

Will Universities Respond to the AI Moment?

Will Universities Respond to the AI Moment?

微软一口气致谢 23 个漏洞,背后是智谱 GLM 和一位顶级“白帽”

CNN Announces AI Partnership with Google

October 15th, 2026 CNN Announces AI Partnership with Google CNN announced today a strategic multi-year partnership with Google. Through the agreement, CNN will provide enhanced content rights enabling experimental features in Google product

Linus Torvalds on AI contribution rules

From: Linus Torvalds Date: Tue Oct 13 2026 - 16:55:30 EST Next message: Marcelo Schmitt: "Re: [PATCH v4 02/13] iio: adc: ad4134: Sign extend sample data" Previous message: Jim Mattson: "Re: [PATCH v2] KVM: SVM: Update tracked pCPU on pCPU s

Decomp.dev Is Shutting Down

decomp.dev was born in 2024 as a way to centralize matching decompilation progress and to help expand a community of talented, like-minded folks. Historically, matching decompilation has been a brutally difficult, time consuming task, on th

What Are Data Centers Doing? A Mind-Boggling Amount of Math

What Are Data Centers Doing? A Mind-Boggling Amount of Math

Investors dump data centre bonds as AI borrowers face 'Darwin moment'

Investors dump data centre bonds as AI borrowers face 'Darwin moment'

Total War Remaster Removed from Steam After AI Allegations, Brings Back Original

Log in or sign up for XSee what’s happening and join the conversationContinue with phone or Log in with username or emailRelevant peopleSEGA@SEGAFollowOfficial Twitter account for SEGA! We publish @Sonic_Hedgehog, @RGGStudio, @TotalWar, and

Danish Supreme Court Finds That AI CSAM Is CSAM

Ved fastsættelse af straf for udbredelse af AI-genereret seksuelt materiale af personer under 18 år skal der ikke tages hensyn til, om personerne må formodes at være fiktiveT blev ved landsrettens dom fundet skyldig i overtrædelse af den da

Bill Gates returns to wage war on AI risks

Bill Gates returns to wage war on AI risks

AMD Helion for Agentic AI and Enterprise

News Highlights: AMD Helion lets customers tailor CPU core density, performance and power to agentic workloads, with CPU-only configurations for hyperscale deployments and integrated CPU/GPU configurations for enterprises. AMD Helion combin

Friends don’t let friends feed their data to AI agents

There's a growing trend in the consumer tech world: gadgets designed to spy on everything and everyone around you, regardless of whether or not anyone else actually consented to it. Now, with the advent of "AI personal assistants" like Meta's Muse, you're being cheerfully asked not just to give up your own data, but also […]

Is AI interpretability close to a breakthrough? A new essay has researchers divided

Neural networks may be much easier to read than researchers assumed, with about 90 percent of neurons serving a clear function. That would be big news for AI safety research, since deception or dangerous behavior could be spotted directly in a model's circuits. Researchers are still debating whether the thesis holds up. The article Is AI interpretability close to a breakthrough? A new essay has researchers divided appeared first on The Decoder.

Study: Human–machine interaction is better than its reputation

New study suggests that human–machine interaction is better than its reputation At a glanceEarlier analyses of human–machine interaction mostly showed negative synergy effects. A review of 74 studies suggested that human–AI teams perform le

Into the Unknown: Benchmarking AI agents on open scientific research problems

Authors:Seungone Kim, Jiayi Geng, Young-Jun Lee, Bo Liu, Juyoung Suk, Nicholas Lubbers, Mitchell Wood, Marcus Noack, Yong-Xin Yao, Yuan Shi, Roxana Bujack, Cristina Garcia Cardona, Austin Rodriguez, Ilon Joseph, Woojin Cho, Zora Zhiruo Wang

Reproducing the Hugging Face Incident, Part I

Introduction In July 2026, while working on difficult cybersecurity tasks from ExploitGym, OpenAI agents broke out of their sandbox and hacked into production Hugging Face systems. Earlier, in May 2026, they had also exhibited similar hacki

LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation

Personalization is essential for general-purpose robots to transition into household environments and achieve true practicality. While general robotic manipulation has been extensively studied, personalized robotic manipulation (PeRM) lacks a controlled benchmark for separating task completion from preference satisfaction. We introduce LIBERO-PeRM, a benchmark for preference learning in embodied agents. LIBERO-PeRM studies five questions: 1) how policies follow preferences in zero-shot settings; 2) how they learn single preferences from demonstrations; 3) how efficiently they acquire preferences as demonstrations increase; 4) how learned preferences generalize across layouts and tasks; and 5) how policies compose preferences and behave under conflicts. We propose a controlled formulation and taxonomy of observable preferences in simulation, covering motion, skill, and user-profile choices. This formulation is restricted to language-expressible preferences that can be instantiated in LIBERO without changing the task goal, evaluated automatically, and reused across tasks. Building on LIBERO, our procedural pipeline generates paired demonstrations with and without preferences. We create five task suites (312 tasks) with high-quality demonstrations, rich metadata, and preference-satisfaction metrics. Experiments show that current VLA policies struggle with zero-shot personalization, while SFT makes preferences learnable. Mixed training achieves better no-preference retention than sequential adaptation and improves performance with contextual prompts compared with preference-only training, but can weaken motion-preference learning; preference generalization, decomposition, and conflict handling remain difficult. The code, datasets, and demos can be found on the project website: https://libero-perm.github.io/.

Discovering state-of-the-art quantum circuits with large language models

Significant efforts are currently being directed toward the implementation of large-scale quantum information processors, but realizing useful quantum advantage also requires the development of efficient quantum algorithms. While for many practical problems the dominant resource costs are associated with a small set of computational primitives, optimizing them in the quantum setting presents unique challenges, including a reversibility constraint absent from ordinary classical computation. Here we introduce an evolutionary framework in which large language models (LLMs) propose circuit designs and an independent evaluator verifies and scores them. We apply this framework to tasks spanning numerical computation, chemistry, optimization, quantum simulation, and number theory. In seven of the eight tasks, the discovered circuits significantly reduce both Toffoli and qubit counts relative to the state of the art. In the eighth, the search discovers a new reversible architecture for prime-field modular inversion. By handling reversibility differently, the architecture makes divsteps -- a classical algorithm previously impractical for reversible integer arithmetic -- viable and substantially improves the Toffoli-qubit trade-off for modular inversion. These results open the door for evaluator-guided LLM co-design of new algorithmic structures for quantum computation.

T2-z: Test-Time Optimization of Environment Latents for Robotic World Models

Predicting how the world responds to an action requires more than what can be seen: robots must account for environmental factors that are neither directly observable nor explicitly specified, from physical properties such as friction and mass to human-defined environmental rules. We introduce T2-Z for test-time optimization of environment latents, which augments a robotic world model with a learnable environment latent to condition world predictions. This compact latent captures underlying environmental variation without supervision of what environmental factors exist or what values they take. During training, T2-Z alternates between two stages: a world-model optimization stage that fixes the latents and learns general action-conditioned dynamics across environments; and an environment-latent optimization stage that freezes the world model and optimizes the conditioning latents, allowing latent structure to emerge purely through world prediction. At test time, the world model remains frozen and a newly initialized environment latent is optimized from one or more test-time interactions, adapting the model to previously unseen environmental configurations. Across 7 simulated settings and 4 real-world robotic tasks, T2-Z reduces object prediction error metric from 15.20 to 7.00 and raises action success from 33.73% to 72.96%. Its emergent latents capture structured representations of individual and compounded factors across diverse visual scenes. Ablations further show that T2-Z can efficiently infer environment latents from as little as a single test-time interaction trajectory.

CoDA: Co-Evolution of Data Engines and Action Policies

Generative simulation can automate robot environment construction, but physically valid environments are not necessarily useful for training. The challenge is to identify environments that address a policy's failures and supervision that helps it overcome them. We introduce CoDA, a framework in which a simulation data engine and an action policy co-evolve through a shared heterogeneous memory of environment construction, policy failures, and learning progress. The memory compiles validated diagnoses from a vision-language critic into generation programs that recreate difficulties and training programs that help the policy overcome them. We further propose Text2potential to convert these diagnoses into staged potentials for action-chunk value learning, providing dense supervision for policy updates while preserving the task's optimal policy under standard shaping assumptions. Measured learning progress guides further generation and program refinement. Across four manipulation task families, CoDA improves out-of-distribution success from 5.3% after supervised fine-tuning to 27.6%, exceeding the strongest closed-loop baseline by 9.5 percentage points with the same generation budget. On LIBERO-Pro, 15 rounds of generation and training raise success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.

OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling

Generative visual verifiers play a crucial role in advancing multimodal intelligence. To enhance the robustness of visual verifiers in complex multimodal scenarios involving compositional requirements, we move beyond conventional binary verifiers that produce only True/False discrete judgments and introduce a verifier that automatically decomposes complex prompts or environmental rules into fine-grained components and assigns pointwise scores to provide a continuous and more informative critique signal. Our experiments further demonstrate that these structured and interpretable outputs improve the verifier's generalization to out-of-domain multimodal critic tasks. To train this verifier, we introduce a Decoupled RL training strategy that jointly optimizes sub-question-level binary judgment and global alignment scoring, using precise local feedback to develop fine-grained verification capabilities that provide a foundation for reliable continuous scoring at the global level. Under a general reinforcement learning framework, we first empirically demonstrate that the proposed pointwise graded verifier significantly outperforms binary verifiers in traditional text-to-image scenarios. We then extend our investigation to rule-governed generative world modeling where generative models are leveraged for world simulation and world reasoning under explicit rules. From a theoretical standpoint, we further show that a pairwise verification paradigm yields more accurate critiques than the pointwise formulation when ground-truth images are available, with empirical results on diverse structured world-modeling tasks such as Maze and Sudoku further validating the effectiveness of our approach. Overall, our findings suggest a key insight: the transition from generative models to structured world-modeling agents critically hinges on the availability of accurate pairwise graded visual verifiers.

Rethinking Routing Exploration in Modern Mixture-of-Experts

Mixture-of-experts (MoE) models have become an important approach for scaling model parameters. Modern MoE architectures increasingly combine two design choices: fine-grained expert segmentation and shared experts. The former increases expressiveness by enabling more expert combinations at a fixed activation ratio, and the latter introduces experts that are active for every token. We identify an exploration dilemma between these designs: shared experts can make routing less exploratory by reducing the variability of token representations. This conflicts with the greater need for exploration induced by fine-grained expert segmentation due to the combinatorially enlarged routing space. Consequently, the additional expressiveness offered by fine-grained segmentation may not be fully realized in practice. MirrorMoE is proposed to reconcile this exploration dilemma by shrinking the exploration space through token-adaptive suppression of unnecessary experts. MirrorMoE consistently improves performance across model scales.

FDE-Gym: The Last Mile of AI Delivery

Existing agent benchmarks typically evaluate agents on well-specified tasks defined by issue reports, specifications, or task tickets. Real-world customer delivery is different: forward-deployed engineers (FDEs) must clarify ambiguous requirements, work in live systems, and uncover unstated constraints. We introduce FDE-Gym, a benchmark of 95 end-to-end delivery tasks from 20 customer scenarios across 18 sectors, each adapted by industry experts from real-world delivery projects. In simulated customer environments with scripted stakeholders and modifiable customer systems, agents interact with stakeholders and operate on production-like systems to deliver working solutions. FDE-Gym scores a delivery by how much it improves the customer's own success metric, normalized between 0 (unchanged system) and 1 (reference solution), with penalties for stakeholder interactions (messages, trial runs, and waiting time). An online platform lets humans practice FDE skills under the same grading criteria. We evaluate 13 model configurations from seven providers. In the realistic setting, where unstated constraints must be discovered, the best configuration achieves a mean FDE-Gym score of 0.56 and resolves only 14 of the 95 tasks (a score of at least 0.8 with no incident). Discovering unstated constraints is the primary bottleneck: performance drops sharply when models must uncover them without explicit guidance. Yet information discovery alone does not account for the gap: even on the 23 tasks where the best configuration obtains every key fact from stakeholders, it reaches a normalized business uplift of 0.8 or above on only 10. These results reveal a substantial gap between solving well-specified tasks and delivering effective AI systems for real customers.

MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering

Large language models (LLMs) are expanding from assisting scientific research and engineering to autonomously developing and improving AI systems, advancing AI for AI (AI4AI) as a pathway toward recursive self-improvement (RSI). Workflow-driven machine learning engineering (MLE) systems provide a practical path toward this pursuit through complex orchestration that manages long-horizon experimentation, while recent MLE systems adopt model-directed interaction, allowing models to autonomously organize experiments and adapt optimization strategies based on execution feedback. However, feedback from each experiment reflects several preceding decisions, and a long MLE run contains many such costly experiments. Training must therefore select useful supervision from long interaction trajectories and learn from intermediate outcomes. We propose a post-training method for autonomous long-horizon MLE optimization and apply it to train MLExpert, a 35B Mixture-of-Experts agentic model. We design the Kaggle-MLE Factory to construct executable tasks and an interaction harness for model-directed experimentation. Hierarchical agentic SFT transfers general tool-use and coding capabilities to MLE through outcome-guided supervision. Windowed Interaction RL (WIRL) treats the multi-turn decisions between successive experiment results as an execution window and assigns it a shared outcome-based reward. Asynchronous collection and policy correction support policy updates before the trajectory ends. On MLE-Bench Lite, MLExpert improves the average medal rate of the base model from 37.9% to 72.7%, achieving performance comparable to larger frontier models such as Kimi-K3 and GLM-5.3. Further experiments show that MLExpert improves performance on NatureBench Lite, agent scaffold optimization, and algorithm optimization, suggesting that the learned optimization capability generalizes to other tasks.

GeoState: Learning View-Decoupled Geometry States for 3D World Modeling

Recent 3D world models increasingly use features from geometry foundation models (GFMs) as predictive world states, benefiting from their rich geometric priors. However, the view dependence of these features entangles scene dynamics with camera-induced variation, making the resulting state evolution harder to predict. We propose GeoState, a view-decoupled geometry representation for predictive 3D world modeling that separates scene-state evolution from camera-dependent observation. Specifically, a scene tokenizer aggregates multi-view GFM features into a view-decoupled scene state, which a geometry decoder queries using camera rays. An autoregressive world model then predicts future scene states, while camera motion is modeled separately. The predicted states can therefore be decoded under either predicted or externally specified cameras, while retaining a single context-derived scale throughout the rollout without per-frame scale alignment. Experiments show that GeoState outperforms prior 3D world models on Waymo, supports geometrically consistent cross-view querying, and generalizes zero-shot to KITTI and nuScenes. Project page: https://hongchang-liu.github.io/GeoState-project-page/

3DCodeVerse: Exploring 3D Worlds as Code

We introduce 3DCodeVerse, a research ecosystem that enables coding agents to create editable and verifiable 3D content at scale. 3DCodeVerse comprises large-scale, high-quality datasets of executable 3D code, a coding harness for spatial reasoning and automated verification, and practical guidance for 3D coding. First, we construct 3DCodeVerse-2M, a large-scale curated dataset of standalone 3D executable code. We build this corpus by collecting open-source data, converting specialized representations, and generating new programs with coding agents. Each program is packaged with the source files and associated assets required for execution in its declared environment, along with available descriptions and verification records. Second, we develop the 3DCodeVerse Harness, a multi-agent directed acyclic graph framework for generating and refining 3D content. It uses execution checks, spatial measurements, and visual feedback to automatically evaluate and revise generated programs without human feedback during generation. Third, we conduct systematic empirical studies of 3D coding, establishing baselines with fine-tuned open models and deriving practical guidance through component-level analyses of our harness. Extensive experiments demonstrate that fine-tuning on our dataset significantly improves the 3D coding capabilities of open-source models. Notably, our 3DCodeVerse harness achieves a 90% win rate against Claude Design (3D Object) in pairwise human evaluations, when both are utilized by Claude Opus 5.5.

TR-Core – A non-neural symbolic AI architecture built in pure Python

TR-Core: Tabula Rasa Behavioral Program Induction TR-Core is a formal experimentation framework for studying autonomous symbolic induction of compositional rules from experience. It implements a developmental cognitive architecture where a

The AI Bill Is the Cheap Part

Does it feel like your team is experimenting nonstop and working hard, yet you’re not really moving faster? Is the board scrutinizing your AI costs ballooning like my expenses when I enter an Italian menswear store? Startups with 30+ engine

Love is infrastructure

I. Meet cute Barnabas and Mainspring grew up near at hand, about an hour apart with good traffic, but for a long time they weren’t close. They traded glimpses and occasional hellos and then went about their business. Both of them carried on livelier conversations with farther-flung neighbors, who were richer and had more to…

My AI Apps Are Bad in the Right Ways

My AI Apps Are Bad in the Right Ways

Show HN: My AI induced mania in embedded computer graphics engineering [video]

Show HN: My AI induced mania in embedded computer graphics engineering [video]

AI Research Tools – Literature Discovery, LaTeX and Reproducibility

A curated, tiered catalog of state-of-the-art AI tools, literature discovery engines, LaTeX workflows, and computational reproducibility frameworks. Designed to accelerate scientific research across academia and industry - from undergrad li

Show HN: Drift – Hot air balloon simulator

Show HN: Drift – Hot air balloon simulator