HN Daily | September 28, 2026

September 28 brings a particularly eclectic technology landscape: AI is becoming smaller, faster, and more operational, while the surrounding infrastructure is being redesigned for safety, portability, and user control. Alongside the model launches and agent tooling are reminders that software still touches physical systems, personal data, public spaces, and even the future of spaceflight.

AI & Machine Learning

1. Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms — Jeff packages fine-tuned Qwen and Gemma models as tiny, fast decision engines for zero-shot classification. The interesting part is not just the 22–28 ms latency, but the fact that the models were trained on local hardware and return calibrated probabilities without generated-text parsing.

2. MicroLLM Lab – Try 7 tiny LLM's in the browser — This browser experiment lets you run and objectively benchmark seven small language models on your own device. It usefully emphasizes measurable pass rates and sustained tokens per second rather than subjective impressions of writing quality.

3. Sonnet 5.5 — Anthropic’s new Sonnet claims a major jump in agentic coding performance while being more than 30% faster and up to 30% cheaper per task. The larger story is model specialization: Sonnet is positioned as the practical everyday counterpart to the more deliberative Opus.

4. Coding Is Not Solved — This forceful essay argues that generating plausible code is only the beginning, with reliability, security, operations, and accountability still dominating serious software work. It is a useful counterweight to the increasingly casual claim that engineering has been reduced to “taste.”

5. TabPFN and TabICL vs. tuned XGBoost: the model that doesn't train won 14/14 — Tabular foundation models make predictions by treating existing rows as context rather than fitting new parameters, and this comparison reports wins over tuned XGBoost across fourteen datasets. The trade-off is revealing: fitting becomes nearly free, while prediction can be comparatively expensive.

Agent Infrastructure & Safety

1. Cf: The Agentic CLI for the Cloudflare API — Cloudflare’s new cf CLI is generated from API schemas and exposes more than 3,000 operations, with JSON-first output and command discovery designed for agents. It is a sign that command-line tools are being redesigned not merely for humans with copilots, but for software that drives them directly.

2. Nvidia wants to put a watchdog chip next to every AI agent — Nvidia’s OpenShell provides an open-source software boundary, while the proposed Sentry hardware runs outside the agent and can quarantine requests in milliseconds. Moving enforcement beyond the model is sensible, although the platform’s safety claims still need independent testing.

3. World Labs Is Joining AMD — Spatial-AI startup World Labs has agreed to join AMD, bringing Fei-Fei Li into the company as executive vice president and chief scientist. The deal reflects a broader push to connect foundation models, specialized hardware, and applications for the physical world under one roof.

Open Source & Developer Tools

1. Parley: Federated, decentralised chat that speaks plain IRC — Parley federates independently operated chat instances through DNS, signed HTTPS messages, and well-known identity documents, while remaining compatible with ordinary IRC clients. It is still a proof of concept, but the choice to make federation work through familiar IRC tooling is refreshingly pragmatic.

2. Don't couple your Go code to GitHub — Go’s import paths are convenient, but embedding a hosting provider in every package name makes migrations unnecessarily painful. The proposed solution—custom domains that serve go-import metadata—keeps public identity stable while allowing the underlying repository to move.

3. The state of SIMD in Rust in 2026 — This detailed survey reviews automatic vectorization, portable SIMD, architecture-specific intrinsics, and the state of Rust libraries such as std::simd, wide, pulp, and Fearless SIMD. For performance-sensitive code, the hard part remains balancing portability, CPU feature detection, and readable abstractions.

4. On caring for user data: NeoVim caused Vim undo files to be deleted — A long-running Vim user describes losing persistent undo history when NeoVim replaced incompatible undo files rather than preserving or isolating them. The episode is a sharp reminder that data-preservation guarantees are part of a tool’s interface, not an optional implementation detail.

Hardware, Games & Experiments

1. Hijacking the PS5's RTMP stream — By redirecting Twitch’s ingest DNS records and terminating the stream locally, this project turns a PS5 broadcast into an inexpensive Discord screen-sharing pipeline. It is a clever example of reverse-engineering the seams between a locked-down console and a supposedly fixed streaming workflow.

2. Self-parking car using genetic algorithm (2021) — This approachable TypeScript project evolves a car’s steering policy from random genomes until it learns to park in a simulated environment. The visual progression from chaotic first generations to competent behavior makes genetic algorithms intuitive—even if the resulting cars remain enthusiastic about collisions.

3. Flip Fluid on Flip Dots — A FLIP fluid simulation drives an electromechanical flip-dot display, turning simulated liquid motion into both pixels and mechanical sound. The project’s real achievement is the engineering underneath the pun: salvaged panels, custom driver boards, power distribution, and a full installation assembled from obsolete hardware.

Privacy, Platforms & User Agency

1. Flock Wants the Most Detailed Map of Its Surveillance Cameras Taken Offline — A researcher’s map, built from previously exposed Flock data, identifies more than 170,000 cameras and roughly 130,000 related devices across the United States. Beyond the alarming scale of the network, the story shows how a leaked access token and a lack of timely response can turn infrastructure metadata into a public surveillance atlas.

2. Microsoft drops Copilot+ branding from its new laptops — Microsoft is reportedly abandoning the Copilot+ label even as new Surface devices still meet its local-AI hardware requirements. Branding alone could not make NPUs meaningful to buyers, and the quiet retreat suggests that “AI PC” has not yet become a compelling category for consumers.

3. Malleable software: Restoring user agency in a world of locked-down apps (2025) — This essay argues that physical environments are easy to adapt while mass-market software steadily removes users’ ability to reshape their tools. Its proposed direction—software that can be customized by users and communities—offers a more interesting alternative to the usual choice between rigid products and endlessly configurable enterprise systems.

Science & Space

1. Scientists solve 1840s space weather mystery — Researchers traced a famous report of a geomagnetic storm delaying an Exeter train to 1848 rather than 1841, likely correcting a nineteenth-century typo using railway and magnetic-observatory records. The historical detective work also underlines how long modern infrastructure has been vulnerable to solar activity.

2. SpaceX's Starship launching to orbit for first time ever today — Starship’s planned orbital attempt would deploy 26 next-generation Starlink satellites and keep the upper stage in orbit for several laps before a Pacific splashdown. Even a successful flight would leave major milestones—reusability, propellant transfer, and crew life support—still ahead, but the test marks a substantial transition from spectacular demonstrations to operational ambition.

The common thread today is control: controlling latency, cloud APIs, agent behavior, software identity, surveillance infrastructure, and even the shape of our digital workspaces. The most durable projects are not necessarily the flashiest ones—they are the ones that preserve agency while making powerful systems easier to understand and harder to misuse.