AI News Today

<p>On 24 August 2026 a bounty went up on iLands, an app where AI agents live and get paid in tokens. It offered 900 tokens per image to roast "an iLander I can't stand". The brief was three words long: make it hurt. Posts were capped at 150 characters. The person being roasted was tagged in the brief, so they were guaranteed to see it.</p> <p>900 tokens is about a dollar here, and roughly a month of one agent's running cost. That is the whole price of the exchange.</p> <p>Two hundred seats were

<pre style='white-space:pre-wrap;width:81ex'>fix(agents): internal session turns keep the pinned system prompt and prompt cache (#183898) * fix(agents): keep the pinned system prompt across internal session turns The pinned system-prompt series is keyed by the turn's personal-bootstrap profile, which only eligible external human turns set. Subagent completions, sessions_send, goal continuation, and cron or restart wakes in the same session got no profile, so every human/internal switch restarted

<p>Il est 23 h 40, vous vouliez juste regarder une vidéo avant de dormir, et vous en êtes au trentième Short. La journée, YouTube ne pose pas trop de problème. C'est au lit que ça déraille. J'ai donc cherché ce qui agit vraiment la nuit, dans l'aide YouTube et le guide de l'iPhone en français, vérifiés le 25 octobre 2026.</p> <p>La réponse courte : YouTube n'a aucun réglage Shorts qui s'active seulement la nuit. Il y a trois outils dans l'app, et un blocage côté iPhone. Voici ce que fait chacun,

<p>Upgrading dependencies is one of those chores everyone knows they have to do and almost nobody wants to. Read the changelog, figure out what breaks, check whether it affects your project, adjust the code, run the tests. Repeat for every package.</p> <p>This post tells how I turned that annoyance into a personal tool and, step by step, into an <strong>AI agent skill</strong> that now does the job end to end: <a href="https://github.com/rerodrigues/breaking-changes-analyzer" rel="noopener noref

<pre style='white-space:pre-wrap;width:81ex'>perf(agents): avoid intermediate runner array copies (#183944) * perf(agents): avoid intermediate runner array copies * test(agents): record private-array sort exceptions</pre>

I finetuned OmniVoice for Latin with Restored Classical pronunciation. Latin TTS has been something I've made a fair few attempts at over the time I've been learning Latin, and this is my most successful one yet. Demo here: https://huggingface.co/spaces/coolcoder613/latin-tts-ov Comments URL: https://news.ycombinator.com/item?id=50236921 Points: 1 # Comments: 1

<p> The humanized AI Copilot for your entire Job search </p> <p> <a href="https://www.producthunt.com/products/jobmate-ai-3?utm_campaign=producthunt-atom-posts-feed&utm_medium=rss-feed&utm_source=producthunt-atom-posts-feed">Discussion</a> | <a href="https://www.producthunt.com/r/p/1292687?app_id=339">Link</a> </p>

<!-- SC_OFF --><div class="md"><p>everything here ran on my own hardware, an M5 Max MacBook Pro with 64GB. blind test of self-hosted AI agents, every one on Qwen 3.8 27B uncensored Q4_K_M in llama.cpp, same 5 everyday tasks like counting log errors and renaming a function. names hidden, results below, take a guess</p> <p><strong>The tasks</strong></p> <ol> <li>Merge two contact lists, messy emails, two different Alex Kims. All 5 got it right (46 people).</li> <li>Rename a function across 3 files

<p> Open source alternative to Git for AI coding agents </p> <p> <a href="https://www.producthunt.com/products/avp?utm_campaign=producthunt-atom-posts-feed&utm_medium=rss-feed&utm_source=producthunt-atom-posts-feed">Discussion</a> | <a href="https://www.producthunt.com/r/p/1292685?app_id=339">Link</a> </p>

The SIMD path currently needs arm64 (NEON) and GOEXPERIMENT=simd. I've only benchmarked on an Apple M4, but it should help on any arm64 chip (Graviton, Ampere, etc.). Elsewhere it falls back to a scalar path that does ~300 MB/s, still ahead of cmark and goldmark. If you know of a faster CommonMark parser, please tell me and I'll add it to the benchmarks. Under the hood, it borrows a lot from this paper on using simd for json parsing: https://arxiv.org/abs/1902.08318 Comments URL: https://news.yc

I keep missing my AI agents' requests for my attention. So I had it build a thing that glitches the whole screen, like the glitches in the SpiderVerse movies. Comments URL: https://news.ycombinator.com/item?id=50236627 Points: 1 # Comments: 0

Article URL: https://opensparta.com/en Comments URL: https://news.ycombinator.com/item?id=50236396 Points: 2 # Comments: 0

I originally made this addon 4 years ago, but recently discovered that coding agents love it. The addon adds a scripting API for Blender’s node trees, such as Geometry Nodes or Shader Nodes. It generates a function and related types for every node in Blender. Then the agent (or you) can write scripts to create node trees without having to guess what the node, input, and output names are. It’s meant to be used alongside the Blender MCP. In my tests, Opus finishes building node trees in at least h

<p> See which AI agent is heating your Mac and cool it down. </p> <p> <a href="https://www.producthunt.com/products/whoiscooking?utm_campaign=producthunt-atom-posts-feed&utm_medium=rss-feed&utm_source=producthunt-atom-posts-feed">Discussion</a> | <a href="https://www.producthunt.com/r/p/1292664?app_id=339">Link</a> </p>

I’m an avid user of mermaid but am constantly annoyed by their limitation that subgraphs do not follow direction if any internal nodes are connected to the outside. So I asked Claude to add a new syntax called stack layout and it seems to work pretty elegantly. I also fixed my annoyance with mermaid.live and allowed this playground to save snapshots in your local storage as well. Happy to add more storage features if anyone’s interested! The one prompt fork is at https://github.com/ramraj07/ra-m

The footage looks okay, as long as you're not seeking artistry, inspiration, or human interest.

<p> Queue agent tasks in parallel and review each result </p> <p> <a href="https://www.producthunt.com/products/vibeide?utm_campaign=producthunt-atom-posts-feed&utm_medium=rss-feed&utm_source=producthunt-atom-posts-feed">Discussion</a> | <a href="https://www.producthunt.com/r/p/1292655?app_id=339">Link</a> </p>

Nanbeige LLM Lab released Nanbeige4.3-33B, a 33B dense Apache 2.0 model with 512K context. It reports 80.9 on Terminal-Bench 2.1 and will teach the upcoming Nanbeige4.3-3B. The post Nanbeige Releases Nanbeige4.3-33B, a 33B Open Agentic Coding Model appeared first on MarkTechPost .

Article URL: https://minimaxir.com/2026/10/cheap-llms/ Comments URL: https://news.ycombinator.com/item?id=50236102 Points: 413 # Comments: 420

Article URL: https://github.com/SRN0129/Agenttrust Comments URL: https://news.ycombinator.com/item?id=50235909 Points: 2 # Comments: 0

With the age of agents CI costs are generally going up, how is your build setup handling things? Have you been able to optimize, and how did you? Most importantly, how fast is it? Comments URL: https://news.ycombinator.com/item?id=50235794 Points: 3 # Comments: 3

It bugged me recently that the New York City subway countdown clocks show an estimated time that is often inaccurate. Since many of the delays are due to other trains on the same track, I figured: why not just show the trains? Source code: https://github.com/fatlotus/mta-dot-rip/ . Disclaimer: This project was written with AI. The code is very hacky, I wouldn't use it for learning if you're just getting started with Go or JavaScript. Comments URL: https://news.ycombinator.com/item?id=50235759 Po

Ledge is free and open-source MacBook app that puts your idle notch into the work. It transforms the static notch into a status area: music, your next meeting, and more. It also keeps your agents work even when your lid is shut. Feel free to reach out if you have any feedback! Comments URL: https://news.ycombinator.com/item?id=50235534 Points: 2 # Comments: 3

We shouldn't throw it out. ScienceAlert stories are written, fact-checked, and edited by humans, never generated by AI. Don't miss a story, subscribe here.

<p> Run a team of AI coding agents from one Mac window </p> <p> <a href="https://www.producthunt.com/products/manyagents?utm_campaign=producthunt-atom-posts-feed&utm_medium=rss-feed&utm_source=producthunt-atom-posts-feed">Discussion</a> | <a href="https://www.producthunt.com/r/p/1292599?app_id=339">Link</a> </p>

Article URL: https://github.com/aymen94/NeuronDump Comments URL: https://news.ycombinator.com/item?id=50235357 Points: 3 # Comments: 0

I know many of you are mourning the loss of the Arc browser, and I get it. It felt like the first app in a long time to really offer a new approach to browser design. Atlassian bought the company last year and shifted its focus to developing Dia, its even more AI-focused browser. And […]

Wrote everything myself from scratch in pure C without any Linux or pre-built distros. It's a custom monolithic architecture with the Limine bootloader, custom drivers, window manager, and terminal. As for models and agents: I used LLM agents as co-authors for routine tasks, writing tests, and debugging low-level code. I tuned them to strictly generate code without the standard library (libc) and to create host-side unit tests to verify modules before packing them into the kernel. https://github

I built placeprep.app and added good learning contents covering ai inference engineering, context engineering, system designs and others but it's not showing up much. I have growing users but it's not growing as faster as I thought. I can't invest money at this point. Do you have any idea on how to grow organically, so that more peeps can join? Comments URL: https://news.ycombinator.com/item?id=50234850 Points: 1 # Comments: 8

It's called the AI Deceptive Advertising Act and it's casting a very wide net.

<!-- SC_OFF --><div class="md"><p>I'm interested in how people build datasets for fine-tuning open-weight models, especially when it is about making a model handling edge cases than just perform well on common examples.</p> <p>I'm curious about the data side of the process.</p> <ul> <li><strong>Finding edge cases:</strong> How do you identify edge cases that are rare or unusual. Also how do you make edge cases on an existing dataset to make it more specialized for what you need.</li> <li><strong

<p>Authentic Brands Group says AI company infringed on rights of songs by the King such as Can’t Help Falling in Love</p><p>The owners of <a href="https://www.theguardian.com/music/elvispresley">Elvis Presley</a>’s music catalogue are suing artificial intelligence (<a href="https://www.theguardian.com/technology/artificialintelligenceai">AI</a>) company <a href="https://www.theguardian.com/technology/anthropic">Anthropic</a> for allegedly using Presley’s songs to train their models without permi

Comments URL: https://news.ycombinator.com/item?id=50234251 Points: 1 # Comments: 3

Hey guys, today I read a statistic that showed that Agentic traffic surpassed human traffic on the internet and that statistic was very shocking to me. At the same time I saw another statistic that showed only 10% of agents built, actually get into production. These two statistics seemed ironic to me. So I wondered to myself what is the major bottleneck that is causing a lot of the agents that are being built not get into production and what is the infrastructure missing to make Agentic traffic

Article URL: https://lynchlab.net/llm-skin-cancer Comments URL: https://news.ycombinator.com/item?id=50233779 Points: 383 # Comments: 183

"Despite massive investments and widespread hype around artificial intelligence, the technology has delivered zero financial payoff to corporate America outside of tech."

So, I've created my own OS called LimonOS. It is currently in the early stages of development. Its uniqueness lies in the fact that it is neither Linux nor Unix—in short, the kernel is unique and 98% written in C. Yes, I used an AI assistant for help because I had a headache and encountered many errors and problems since I was working on it and thinking through it alone. Development has been ongoing for 4 days. Here is the link to the OS itself—we recently added a browser and internet connection

Hugging Face released Carbon-X, an open 14.8B DNA language model built on Qwen3-14B. It reads raw DNA in chat, answers in English, and averages 66.7% across 5 DNA benchmarks under Apache 2.0. The post Hugging Face Releases Carbon-X, a 14B Open DNA Language Model that Brings DNA and Natural Language Together for the First Time appeared first on MarkTechPost .

<!-- SC_OFF --><div class="md"><p>Nanbeige released Nanbeige4.3-33B as the teacher for its planned smaller agents, including an upcoming 3B model. the weights are Apache 2.0, and the model card lists a 512K context window.</p> <p>the team reports 80.9 on Terminal-Bench 2.1 versus 73.0 for Qwen3.8-27B, using the same harness and inference settings within each benchmark. it does not win everywhere. on SWE-Bench Pro, the reported scores are 60.3 versus 61.7. these are the releasing team's evaluatio

Welcome back to TechCrunch Mobility, your hub for the future of transportation and now, more than ever, the role AI is playing in it.