HN Daily | August 4, 2026

August 4, 2026, is a day where the past and future of technology collide. We're celebrating the 20th anniversary of Pandoc, a tool that has quietly powered document conversion for two decades, while simultaneously grappling with the implications of AI-generated code and images. The day's biggest story is a massive supply chain attack on npm, a stark reminder of the fragility of our software ecosystem, but there's also plenty of exciting progress in making large language models smaller, faster, and more accessible.

AI & Machine Learning

1. Mistral's Shieldstral: 3B open-weights model for multimodal moderation — A new 3B parameter safety classifier that outperforms models up to 7x its size by framing content moderation as a policy-adaptive question-answering task. It accepts plain-language policies at inference time, unifying text and image safety evaluation without retraining, and is released under Apache 2.0.

2. DeepSeek V4 Flash on a Single AMD MI300X — A detailed production configuration for running the 304B-parameter DeepSeek V4 Flash model on a single AMD MI300X GPU, achieving 168.6 tok/s decode speed. This repo includes critical fixes for FP8 format issues and MoE routing, making high-end AI inference more accessible on AMD hardware.

3. When AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation — A comprehensive study analyzing benchmark saturation across 60 language model benchmarks, finding that nearly half exhibit saturation, with rates increasing with age. The research suggests that expert-curation, not public test data, is key to extending benchmark longevity.

4. Why Large Language Models Fail at Tabular Prediction — This paper systematically tests five hypotheses for why LLMs struggle with tabular data, falsifying four of them. The decisive factor is dimensionality: the LLM's accuracy decreases as dimensionality grows, while classical baselines stay flat or improve, explaining why LLMs lose to fifty-year-old baselines on tables.

5. Show HN: Run an 80B Qwen in 4.3 GB of RAM on a Mac, and a 35B on an iPhone — Swiftlet is a Swift + Metal runtime that streams MoE weights from storage on demand, allowing 35B and 80B Qwen models to run on ordinary Apple devices. The 35B model runs on an iPhone 17 in about 2.5 GB of RAM, a first for a model of this class.

6. Smaller, faster, safer: running Kimi and GLM at scale — Cloudflare details three techniques for serving large MoE models efficiently: quantizing the KV cache to FP8, compressing model weights to INT4, and protecting shared caches. These optimizations enable more concurrent requests and lower costs with no change in model accuracy.

Open Source

1. Twenty Years of Pandoc — John MacFarlane reflects on the 20th anniversary of Pandoc, the universal document converter that started as a Haskell learning project. From 3,000 lines of code to over fifty supported formats, it's a heartwarming story of how a useful tool can snowball into an essential piece of the academic and technical writing ecosystem.

2. FFmpeg 9.0 — The latest major release of the ubiquitous multimedia framework is here. While the release notes are light on details, any FFmpeg update is significant given its role as the backbone of video and audio processing across the web.

3. Celebrating 45 Years of Kermit with the First New C-Kermit Release in 15 Years — A testament to the longevity of classic protocols, this post celebrates 45 years of Kermit with a new C-Kermit release. It's a fascinating look at maintaining and evolving a decades-old C codebase.

4. Devtools must be open source — A compelling argument that AI agents have changed the ROI on customizing software, making it easier than ever to personalize open-source tools. The author demonstrates how agents can automatically manage the process of syncing local changes with upstream, turning devtools into truly personal instruments.

Tools & Infrastructure

1. Keyv and friends compromised in active Shai-Hulud supply chain attack — A critical security incident: attackers compromised the maintainer of keyv and injected credential-stealing malware across the entire package family, affecting over 2 billion monthly installs. The malware steals npm, GitHub, AWS, Kubernetes, and Vault credentials, and has worm-like propagation capabilities.

2. Don't stop early: Case-folding source code at memory speed — GitHub engineers share how they optimized case folding for their code search engine, Blackbird, which indexes over 480TB of source code. The counterintuitive key insight: removing an early-exit optimization and using a branchless loop made the ASCII fast path 15x faster, running at memory bandwidth.

3. Agent skills that bring team coding standards to Claude Code and Codex — An open-source framework for turning AI agents from isolated guessers into compliant team members by sharing your team's constitution, architectural standards, and evaluation benchmarks. It works with any agent supporting the Agent Skills standard, including Claude Code, Codex, and Cursor.

4. Show HN: Simple algorithm and color space to generate diverse skin tones — A color picker and procedural generation algorithm designed to make picking plausible and diverse skin tones easy for digital art and game development. The author shares the methodology and properties of the color space, with plenty of room for future improvements.

Science & Research

1. Video2NAND – Abusing video codecs for great computational power — A mind-bending exploration of using the VP8 video codec's prediction mechanisms to simulate combinatorial logic. The author builds composable 'gadgets' to construct arbitrary logic circuits, proving that NAND gates can be built from the most unexpected substrates.

Business & Startups

1. Andy Pavlo joins ClickHouse to establish ClickHouse Labs — The renowned database professor from Carnegie Mellon is joining ClickHouse to lead a new research team. ClickHouse Labs aims to bridge academic research and industry engineering, focusing on how DBMSs fit into emerging AI and agentic technologies.

Culture & Commentary

1. Ray Bradbury's "There Will Come Soft Rains" is set today (2026-08-04) — A poignant reminder that the date in Bradbury's classic story about an automated house outliving its inhabitants has arrived. It's a chillingly relevant read about technology, isolation, and the quiet persistence of nature.

2. Xbox goes down. You can't play games you own on disc — A thoughtful piece on how physical media has changed, sparked by an Xbox outage that blocked even disc-based games. The author argues that owning a disc today is still just a license, and that the PC's approach to maintaining access is a model worth considering.

3. Prevent cognitive debt by manually retyping LLM-generated code — A contrarian and deeply personal workflow: asking the LLM to generate code in chat, then manually typing every edit into the editor. The author argues this builds a mental model of the codebase and prevents the cognitive debt of blindly accepting AI-generated code.

4. AI-Generated Images Discourage Me from Reading Your Blog — A short but sharp critique of AI-generated images on personal blogs, arguing they undermine trust in the authenticity of the text. The author would rather see a shitty MS Paint drawing than an AI image, a sentiment that resonates with many in the indie web community.

---

Today's stories paint a picture of a tech world in transition. We're celebrating the enduring power of well-crafted open-source tools like Pandoc and FFmpeg, while also wrestling with the new challenges of AI: from supply chain attacks to the cognitive debt of generated code. The push to make AI models smaller and more efficient is exciting, but it's the human stories—of maintaining a 45-year-old protocol or manually retyping code to truly understand it—that remind us of the values that keep this industry grounded.