Ghostty is leaving GitHub

The author has been using GitHub daily for 18 years, finding it the place where they've been happiest and most productive. However, due to frequent outages and reliability issues, they've decided to leave GitHub and move their project Ghostty to a new platform.

Bugs Rust won't catch

Canonical disclosed 44 CVEs in uutils, a Rust reimplementation of GNU coreutils. The bugs highlight the importance of careful handling of paths, bytes, and syscalls in systems code.

Before GitHub

https://lucumr.pocoo.org/social/2026-04-28-before-github-social.png
The author reflects on GitHub's decline and its impact on the Open Source community, remembering a time when projects were more decentralized and had more autonomy. They advocate for a public archive to preserve Open Source history and suggest that the community should learn from the past to build a more resilient future.

How ChatGPT serves ads

https://storage.ghost.io/c/b8/53/b853e3d4-3186-409d-9c7f-7da931a60431/content/images/2026/03/buchodi-600x600.png
OpenAI's ad platform injects ad units into ChatGPT conversations and tracks clicks with Fernet-encrypted tokens. The platform uses two domains and specific cookies to track ad events and user behavior.

Show HN: Auto-Architecture: Karpathy's Loop, pointed at a CPU

https://raw.githubusercontent.com/FeSens/auto-arch-tournament/main/docs/blueprints/pipeline-vs-baseline-image-model.png
An autonomous research loop was tested outside its comfort zone, achieving a 92% improvement over a baseline VexRiscv CPU on CoreMark iter/sec. The loop's success was due to a verifier that caught and rejected incorrect hypotheses.

OpenAI models coming to Amazon Bedrock: Interview with OpenAI and AWS CEOs

https://i0.wp.com/dithering.passport.online/assets/2026-04-DitheringCoverart.jpg?w=128&h=128&ssl=1
Matt Garman and Sam Altman discussed how AWS and OpenAI are partnering to make AI more accessible to businesses, with Bedrock Managed Agents being a key product that integrates OpenAI models with AWS infrastructure. The integration of models and harnesses is crucial for making AI work effectively, and the partnership aims to make it easier for businesses to build and deploy AI-powered agents.

We still don't have a more precise value for "Big G"

https://cdn.arstechnica.net/wp-content/uploads/2026/04/bigG1-640x360.jpg
Physicists have been trying to measure the gravitational constant, or Big G, with high precision for over two centuries due to its weak value and background noise from the Earth's gravitational field. A recent experiment by the National Institute of Standards and Technology (NIST) replicated a previous result and obtained a value of 6.67387×10^-11 meters^3/kilogram/second^2.

Withnail's Coat and I

https://substackcdn.com/image/fetch/$s_!gGmV!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F816fde6b-67f0-48a5-afb0-7671c1bed18c_5669x3823.jpeg
By subscribing, you agree Substack's Terms of Use, and acknowledge its Information Collection Notice and Privacy Policy. It’s easy for depictions of 1960s Britain to be enamoured with an idealised, cartoonish vision of swinging decadence. While no doubt a great swath of British pop culture from the period indulged in colourful happenings – understandably considering how fresh it must have ...

Regression: malware reminder on every read still causes subagent refusals

https://opengraph.githubassets.com/79516bae982e4402688992a6b1e032397ad5c373885247d4714fc05acee656e0/anthropics/claude-code/issues/49363
User is experiencing issues with Claude's reminder system causing subagents to refuse code edits on legitimate OSS projects. The reminder is injected into every Read and Grep result, wasting tokens and causing parallel workflows to fail.

I won a championship that doesn't exist

https://raw.githubusercontent.com/ronaldstoner/ron/gh-pages/images/6nimmt/champion.png
The user manufactured a fake 6 Nimmt! World Championship title by creating a seeded website and a Wikipedia edit, demonstrating a cheaper and easier attack on LLM trust models. This attack, known as the circular citation pattern, can compromise LLMs by creating a fake source that appears trustworthy, highlighting the need for better source verification and provenance surfacing in AI systems.

Apple CMF (Color-Matching Functions) 2026

Apple has released two new 27-inch 5K Retina displays, Studio Display and Studio Display XDR, with the latter using Apple CMF 2026 for colour calibration. The displays have varying levels of colour accuracy and contrast ratios.

GitHub RCE Vulnerability: CVE-2026-3854 Breakdown

https://www.datocms-assets.com/75231/1777300906-github-2x-100-1.jpg?fm=webp
Wiz Research discovered a critical vulnerability (CVE-2026-3854) in GitHub's internal git infrastructure that allowed remote code execution on shared storage nodes and full server compromise on GitHub Enterprise Server. GitHub mitigated the issue within 6 hours and released patches for all supported versions.

Intel Arc Pro B70 Review

https://wp-cdn.pugetsystems.com/2026/04/Intel-Arc-Pro-B70-Review.jpg
Intel Arc Pro B70 is a professional GPU with 32 GB VRAM, offering improved performance over the B50, but still trailing behind NVIDIA and AMD GPUs in most workloads. It excels in AI-inference workloads, outperforming the R9700 in MLPerf and offering a promising configuration for multi-GPU inference workstations.

Behavioral timescale synaptic plasticity rewires the brain after an experience

https://www.quantamagazine.org/wp-content/uploads/2025/06/Yasemin-Saplakoglu-article-profile.webp
Neuroscientists have discovered a new form of neuroplasticity called behavioral timescale synaptic plasticity (BTSP) that helps the brain learn from a single experience. BTSP strengthens synapses across several seconds, allowing the brain to capture behavioral processes of learning.

Who owns the code Claude Code wrote?

https://substackcdn.com/image/fetch/$s_!fZ2S!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c45efbc-1881-4750-bdf5-aad04220b0a2_1456x816.png
Using AI coding tools like Claude Code raises complex questions about code ownership, copyright, and regulatory compliance, and developers must document their creative contributions to establish meaningful human authorship. Employment contracts and IP clauses can also impact code ownership, and developers should read their contracts and consider using personal tools and accounts to avoid ...

When the Internet Was a Place

https://www.frontporchrepublic.com/wp-content/uploads/2025/09/5321590647_afe4692f53_b.jpg
The internet has evolved from a physical place we visited to a pervasive, always-on environment that erodes our capacity for rootedness and freedom. To recover a healthier digital culture, we must reimagine the internet as a place we choose to enter and leave on human terms.

Your phone is about to stop being yours

https://keepandroidopen.org/img/logos/gnome.org.png
Google is imposing a developer verification requirement on Android, which critics say is a move to consolidate power and control over the open ecosystem. This change could effectively make Google the gatekeeper for all apps on certified Android devices.

Warp is now open-source

https://www.warp.dev/_next/image?url=%2Fimages%2Fblog%2Fwarp-is-now-open-source%2Fhero.png&w=3840&q=75
Warp's client is now open-source, allowing the community to participate in building it using an agent-first workflow managed by Oz. This change aims to accelerate product development and give developers a chance to shape the future of agentic development.

Nonlinearity Affects a Pendulum

https://www.johndcook.com/nonlinpend1.png
The nonlinear pendulum equation has a longer period than the linear equation due to the sine function. The difference in period is approximately 7.32% for small initial displacements.

Talkie: a 13B vintage language model from 1930

https://talkie-lm.com/images/papers-composite.png
Researchers have created a 13B vintage language model called talkie-1930-13b-base trained on 260B tokens of historical pre-1931 English text to simulate conversations with people from the past. The model's performance is compared to its modern counterpart, with talkie underperforming in some standard LM evaluations but showing similar performance on core language understanding and numeracy tasks.

We decreased our LLM costs with Opus

https://www.mendral.com/blog/frontier-model-lower-costs/pipeline.svg?dpl=dpl_BhxkJBWT4tWkkRe48qQ5Qryq8a94
The company uses a triager pattern with a Haiku agent to detect duplicate CI failures, reducing costs by 80% and improving performance, and then escalates to Opus for further investigation. Opus plans investigations, writes precise sub-agent prompts, and uses Haiku to read logs and search error messages, achieving a 25x cost reduction for triager matches compared to full investigations.

Show HN: Drive any macOS app in the background without stealing the cursor

https://raw.githubusercontent.com/trycua/cua/main/img/logo_black.svg
cuabot is a tool for automating computer tasks with a seamless sandbox for coding agents. It allows running native macOS apps in the background and recording trajectories for training.

Localsend: An open-source cross-platform alternative to AirDrop

https://camo.githubusercontent.com/f0ab3879f57aff14cff2cbecd2dde800add5c488f3b84443215dbd599ec5a722/68747470733a2f2f6c6f63616c73656e642e6f72672f696d672f73637265656e73686f742d70632e77656270
LocalSend is a free, open-source app for secure local file sharing and messaging over a network without an internet connection. It uses HTTPS encryption and a REST API for secure communication between devices.

CJIT: C, Just in Time

https://dyne.org/cjit/assets/dyne-mark.CcSJn4AU.svg
🤏🏼One file <2MB big, no EULA to sign, no IDE to install

Building a Hamiltonian Path Puzzle

http://www.4rknova.com/assets/img/post/20260424-number_trail/og-wide.png
Number Trail is a path-filling puzzle built entirely in plain HTML, CSS, and a single JavaScript module with no frameworks and no bundler. The player draws one continuous path that visits every cell of a square grid exactly once, touching numbered clue cells in ascending order. The core constraint is a Hamiltonian path problem: find a path through a graph that visits every vertex exactly ...

Claude for Creative Work

https://www-cdn.anthropic.com/images/4zrzovbb/website/a62b6eb169818f14c35b7a192af269e283f8fa93-1000x1000.svg
Claude is a tool that helps creative professionals by automating tasks, expanding their skill set, and enabling them to take on larger projects. It integrates with popular software like Adobe, Autodesk, and Blender to make creative work faster and more efficient.

I have officially retired from Emacs

https://nullprogram.com/img/elfeed2.png
The author has retired from using Emacs after 20 years, replacing it with modal editing and Vim, and is now transferring two of its applications, the Emacs Calculator and Elfeed, to new maintainers.

UAE to leave OPEC

The following information can help our support team to resolve this issue.

An update on GitHub availability

https://github.blog/wp-content/uploads/2025/07/wallpaper-generic-blue.png?resize=1600%2C850
GitHub experienced two recent incidents due to rapid growth in software development and has started increasing capacity by 10X with a goal of substantially improving reliability and failover. The company is prioritizing availability, reducing unnecessary work, improving caching, and isolating critical services to prevent similar incidents in the future.

VibeVoice: Open-source frontier voice AI

https://raw.githubusercontent.com/microsoft/VibeVoice/main/Figures/VibeVoice_logo.png
VibeVoice is an open-source AI framework for speech synthesis and recognition, supporting multiple languages and long-form audio processing. It includes models for text-to-speech and automatic speech recognition, with features like speaker tracking and customized hotwords.