AdvertiseWatchSubscribe
Updated , rescanned every two hoursThursday, November 19, 2026

Early Signal: AI news, launches and the projects taking off today

Rewrite apps piece-by-piece in Rust, verified automatically

erkamyaman/rust-anything: preview
Top story▲ 2.3k todayNew · 3d old

It records what your existing Node.js, Python, C, Go, Ruby or Java code returns for generated edge cases, then creates a Rust crate that mimics the same API. The tool replays every recorded case against the Rust code, only allowing the migration to continue when all results match.

· spotted 10 h ago

Today in one paragraph. Today’s theme is AI taking the grunt work out of niche tasks, exemplified by Google’s PaperOrchestra, a suite of multi-agent bots that draft, cite and format research papers with minimal human input. The roundup also covers wireless battery monitors, Rust-first rewrites, and on-device MoE inference.

Top stories

Ad space · FeaturedLaunching something? Put it here.A big card with your own image or video, in the middle of the front page, for a whole day.$39/dayBook a day →

Open source

Open-source projects ranked by stars gained today

ProjectTodayTotal
ersinkoc/phonebox: previewersinkoc/phoneboxNew · 12d oldTurn an old Android phone into a customizable desk panel and touch control deck for Windows, Linux and macOS. 124 widgets and 30 templates for system stats, AI coding, media and automation, in 12 languages. USB or paired Wi-Fi. No cloud required. ▲ 8473.7k
sveltejs/svelte: previewsveltejs/svelteSvelte hits 90k stars, the compiler that rewrites web UISvelte turns your component code into vanilla JavaScript that updates the page without a virtual DOM. The result is smaller bundles and faster runtime performance. ▲ 69289.7k
jingyaogong/minimind: previewjingyaogong/minimindTrain a 64M LLM from scratch in 2 hours for $3MiniMind lets you build a 64-million-parameter language model from zero using only a few dollars of cloud GPU time. The whole pipeline, from tokenizer to RL-fine-tuning, is provided in pure PyTorch. ▲ 68766.7k
DeanChensj/jiangshan-map: previewDeanChensj/jiangshan-mapInteractive Chinese History Atlas Gains 600 Stars in One Dayjiangshan-map is a web-based interactive map that lets you scroll through 2,400 years of Chinese history, showing changing borders, cities, passes and poetic landmarks across 18 dynasties. You can explore each era on a single parchment-style interface or jump to specific locations like Jiankang or the Red Cliffs. ▲ 6238.4k
ag-ui-protocol/ag-ui: previewag-ui-protocol/ag-uiAG-UI protocol standardizes AI agent interaction across appsAG-UI defines a lightweight event stream that lets any AI agent talk to any user-facing application. Agents emit typed events and clients render them, so the same backend can power web, mobile, terminal, or Slack interfaces. ▲ 60220.6k
zhouxiaoka/autoclipzhouxiaoka/autoclip ▲ 56111.1k
getarcaneapp/arcaneArcane simplifies Docker management for developers and opsArcane provides a web UI and CLI to start, stop, and monitor Docker containers without hand-written commands. It aims to be approachable for users who aren't Docker experts. ▲ 54710.1k
mjmirza/app-store-complianceAvoid common App Store and Google Play rejections. Open playbook and pre-submission guard with 117 documented rejection patterns and fixes, privacy manifest and Data safety checks, 95 tracked deadlines (EU AI Act, DSA, COPPA, age laws), and a skill for AI coding agents. ▲ 5421.7k

Models

Trending on Hugging Face

From the labs

Straight from the companies building them

  1. AWS ML
    Introducing enhanced infrastructure observability in Amazon SageMaker HyperPod: preview

    Introducing enhanced infrastructure observability in Amazon SageMaker HyperPod

    Amazon SageMaker HyperPod now offers enhanced observability to help you diagnose GPU training issues faster. It brings together diagnostic and profiling capabilities for training workloads on both Amazon EKS and Slurm orchestrators, pinpointing issues across G

  2. AWS ML
    Using NVIDIA Run:ai GPU Fractioning on Amazon EKS: preview

    Using NVIDIA Run:ai GPU Fractioning on Amazon EKS

    Learn how to run two memory-bounded Kubernetes Pods on a single NVIDIA A10G GPU using open-source NVIDIA Run:ai GPU Fractioning on Amazon EKS. This walkthrough builds a custom Amazon Linux 2023 image, installs KAI Scheduler and the GPU Fractioning Operator, an

  3. AWS ML
    Full Fine-Tuning (FFT) on SageMaker AI Serverless Model Customization: preview

    SageMaker adds Full Fine-Tuning to Serverless Model Customization

    SageMaker AI Serverless Model Customization now lets you run full fine-tuning (FFT) of large language models without provisioning servers. The blog shows FFT beats LoRA on several benchmarks and explains why the weight matrices behave differently.

  4. AWS ML
    Exa boosts embedding throughput by 50% to index billions of documents on AWS AI chips: preview

    Exa speeds up embeddings 50% on AWS AI chips

    Exa rewrote its embedding inference to run on AWS Inferentia2 and Trainium processors, cutting latency by half. The boost lets it embed and index tens of billions of web documents for its neural search engine.

  5. AWS ML
    Introducing workload recovery and expanded infrastructure checks on Amazon SageMaker HyperPod: preview

    SageMaker HyperPod adds auto-recovery for stalled training jobs

    Amazon SageMaker HyperPod now automatically detects when a large training job stalls or produces bad results, and it restarts the affected workload. It also runs deeper health checks on the underlying hardware to catch silent failures earlier.

  6. AWS ML
    Introducing the Amazon SageMaker JumpStart Model Leaderboard: preview

    SageMaker adds a leaderboard to compare 70+ foundation models

    The new JumpStart Model Leaderboard in SageMaker Studio shows side-by-side quality, latency and cost metrics for over 70 foundation models. You can filter, sort and launch a deployment with a few clicks.

  7. AWS ML

    Airbnb Uses Cost-Per-Task to Tame Coding Agent Bills

    Airbnb runs real engineering tasks through Amazon Bedrock’s coding agents and measures the cost of each task. Using Bedrock’s observability data, they pick the cheapest model that still meets quality, keeping production spend under control.

  8. AWS ML

    Trace RAG chatbot answers with Aurora PostgreSQL and Bedrock Guardrails

    The blog walks through adding pgvector to Aurora PostgreSQL, storing retrieved documents, relevance scores, and Bedrock Guardrail decisions alongside each response. This makes each answer auditable and lets you see where the model got its info and whether any safety rules fired.

  9. Google DeepMind

    Google launches Gemini 4 Neon, cheap AI for coding

    Gemini 4 Neon is a large language model that can write code, analyze data, and support scientific workflows. It offers fast end-to-end task completion at a lower cost and a 1-million-token context window.

Research

The papers researchers upvoted most today

  1. Paper
    Tideway: Phase-Aware Tiered Memory Management for MoE Inference on Consumer Devices: preview

    Tideway enables MoE inference on consumer PCs

    Tideway adds a phase-aware tiered memory manager that streams parts of a mixture-of-experts model between RAM, GPU memory and disk during inference. This makes it possible to run large MoE models locally on a typical laptop.

    60 upvotes

  2. Paper
    PROLA: Principal-Orthogonal Low-rank Adaptation for Predictive Spatiotemporal Weather Downscaling: preview

    PROLA sharpens weather forecasts with high-resolution downscaling

    PROLA trains a small adapter on top of a big pretrained weather model, letting it turn coarse global forecasts into detailed local predictions. It does this by learning a low-rank, orthogonal transformation that preserves the backbone’s knowledge while adding fine-scale detail.

    26 upvotes

  3. Paper
    Game2Player: Scaling Gameplay Experience for Visual Agents through Environment-Player Co-Development: preview

    Game2Player lets agents learn from co-developed gameplay data

    Game2Player is a framework that builds game environments and player data together, giving visual AI agents realistic, long-horizon training without costly human recordings. It produces paired video and action streams by evolving the environment and a simulated player in sync.

    23 upvotes

  4. Paper
    CurvDelta: Low-Curvature Delta Modeling For Surgical Learning and Unlearning: preview

    CurvDelta proposes low-curvature updates to keep LLM abilities

    CurvDelta is a method for fine-tuning large language models that limits how sharply the model’s parameters change. By constraining updates to low-curvature directions it tries to preserve the model’s original capabilities while learning a new task.

    22 upvotes

  5. Paper
    HINT-WAM: Hindsight-Informed Training for World Action Models: preview

    New training trick boosts robot action models with hindsight

    HINT-WAM teaches robots to learn from what actually happened, not just what they expected. It improves how world action models plan and execute manipulation tasks by using hindsight-informed updates.

    22 upvotes

  6. Paper
    LightTAR: Light and Fast Target-Aware Representations for Multimodal Tabular Learning: preview

    LightTAR speeds up multimodal tabular models with target-aware embeddings

    LightTAR builds compact, fast representations for tables that mix numbers, categories, text, and images. It replaces the usual PCA reduction of pretrained embeddings with a lighter, target-aware method.

    19 upvotes

  7. Paper

    WorldRuler lets developers script game physics on the fly

    WorldRuler is a framework for building interactive worlds that follow user-defined rules instead of fixed physics. Designers can swap rule sets to get completely different gameplay from the same scene and actions.

    17 upvotes

  8. Paper

    New 'Resolution Hypothesis' explains why AI agents stall

    The paper proposes that the key to successful agent evolution is a clear way to identify incremental improvements. It frames evolution as a proposer-selection loop that only keeps artifacts that score higher on a defined metric.

    17 upvotes

  9. Paper

    Code World Model 2.0 enables controllable video generation

    The system separates a visual proxy that describes world state from a photorealistic video renderer. It lets users script scenes and get either offline clips or live streaming output.

    12 upvotes

The conversation

What builders are arguing about on Hacker News

  1. Hacker News
    Gemini 4 Neon: preview

    Google launches Gemini 4 Neon, cheap high-capacity AI coder

    Gemini 4 Neon is Google’s latest language model that can write code, analyze data, and run scientific workflows end-to-end. It costs $3 per million input tokens and $15 per million output tokens while supporting a 1 million-token context window.

    976 points655 commentsRead the thread

  2. Hacker News

    Google AI answers slash traffic to niche guide sites

    A blog post details how All About Berlin lost 70-75% of its visitors after Google started showing AI-generated answers instead of linking to the site. The author argues the new search format is cutting off direct traffic and revenue.

    798 points695 commentsRead the thread

  3. Hacker News
    Windows 11 is collecting data on which coding agents and developer tools you use: preview

    Windows 11 secretly logs your coding tools and AI assistants

    A Windows component watches for executables like codex.exe, claude.exe, copilot.exe and common dev tools such as python.exe, then writes the names to a local log. If the optional diagnostic data setting is left on (the default), that log is uploaded to Microsoft’s diagnostics service.

    728 points620 commentsRead the thread

  4. Hacker News

    Google AI Overviews Slash Site Clicks by 40%

    Google now shows AI-generated summaries, called Overviews, at the top of search results, linking to the sources it used. When those Overviews appear, users click on external sites about 40 % less.

    707 points547 commentsRead the thread

  5. Hacker News
    Seven new models at scalable price-performance: preview

    Microsoft launches seven cheap, high-performing AI models

    Microsoft’s AI division unveiled seven new models that claim near-state-of-the-art quality at a fraction of the usual cost. They target workloads like code generation and general reasoning while keeping token spend low.

    588 points347 commentsRead the thread

  6. Hacker News
    AI agents ask better questions than KPMG accountants in audit trial: preview

    KPMG’s AI agents out-question human auditors in FTSE 250 audit

    KPMG ran over 100 AI agents alongside its audit team to examine the financial statements of an unnamed FTSE 250 company. The agents dug deeper into financial issues and posed more probing questions than the human auditors.

    478 points206 commentsRead the thread

  7. Hacker News

    New Benchmark Shows Every LLM Needs Its Own Coding Harness

    FrontierHarness Eval v1.1 runs thousands of agent experiments to compare how different coding-agent harnesses perform with various large language models. It measures pass rate and cost per task, revealing that the cheapest or most accurate harness changes with each model.

    418 points133 commentsRead the thread

  8. Hacker News

    Anthropic prices initial public offering

    384 points278 commentsRead the thread

  9. Hacker News

    AI code generators speed writing, slow down reviews by 49%

    Harvard researchers observed that engineers using AI coding assistants produced about 30% more code, but the amount of completed work didn't increase. The extra code took nearly 50% longer to get through human code review.

    322 points443 commentsRead the thread

Get the signal first.

Join the list for the morning email: the day's biggest AI launches, explained in five minutes. Unsubscribe any time.

Prefer a reader? Follow the RSS feed.