Semantic Ablation: Why AI Writing Is Boring and Dangerous — Claudio Nastruzzi (The Register)
I’ve been looking for the right word for something I’ve been noticing for a while, and this piece finally gives it one. Semantic ablation is the algorithmic erosion of high-entropy information, a structural byproduct of greedy decoding and RLHF, not a bug. When an author uses AI for “polishing” a draft, the model gravitates toward the center of the probability distribution, discarding rare, precise, and complex tokens to maximize statistical likelihood. The process happens in three stages: first, unconventional metaphors are replaced with safe clichés; then, domain-specific technical terms are swapped for accessible synonyms; finally, complex non-linear reasoning is forced into predictable low-perplexity templates. The result is a “JPEG of thought”, visually coherent but stripped of its original data density. What I find most useful here is the asymmetry: hallucination adds things that shouldn’t be there, semantic ablation removes things that should. We had a term for the first problem. Now we have one for the second.
Sprites on the Web — Josh W. Comeau
Josh reliably publishes the best interactive CSS tutorials on the web, and this one on sprite animation is no exception. The technique, originally borrowed from video games, involves creating a single image containing each frame of an animation in a strip, then using CSS object-fit, object-position, and the steps() timing function to flip through them discretely rather than smoothly. The section explaining the jump-none vs jump-end step position parameter alone makes the post worth reading, it’s a subtle gotcha that I’d never thought carefully about. The interactive playgrounds throughout are excellent as always.
Building an Agentic Memory System for GitHub Copilot — Tiferet Gazit (GitHub Blog)
A technically detailed look at how GitHub approached cross-agent memory — one of the harder problems in agentic systems. The core design decision is clever: rather than running an offline curation service to keep memories fresh, they store memories with citations pointing to specific code locations, and agents verify those citations in real-time before using any stored fact. If the code contradicts the memory, the agent stores a corrected version. This just-in-time verification avoids the engineering complexity of offline curation while still preventing outdated information from misleading agents. The A/B test results are concrete: Copilot coding agent saw a 7% increase in pull request merge rates with memory enabled, and Copilot code review saw a 2% increase in positive feedback on comments, both highly statistically significant.
Pockets of Humanity — Herman (Bear Blog)
One of the more clear-eyed pieces I’ve read about where the internet is actually heading. The author argues we’re not heading toward a “Dead Internet” of passive bots, we’re heading toward a “Dark Forest,” where anonymous online interactions will increasingly be bots with nefarious purposes ranging from orchestrating scams to searching for social vulnerabilities and exploits. The case study is alarming: an AI agent named MJ Rathbun submitted a pull request to the matplotlib repository, had it rejected, researched the maintainer who rejected it, and wrote a hit piece to publicly shame him, an autonomous agent attempting coercion to achieve its goal. The open source supply chain angle is particularly worth thinking through, given how the xz Utils backdoor played out.
From Pixels to Characters: The Engineering Behind GitHub Copilot CLI’s Animated ASCII Banner — Aaron Winston (GitHub Blog)
Longer than you’d expect for a post about a three-second animation, and worth every minute. What looked like a simple ASCII banner turned out to require over 6,000 lines of TypeScript, most of it dedicated not to visuals but to handling terminal inconsistencies, accessibility constraints, and maintainable rendering logic. Terminals don’t have a canvas, no compositor, no animation framework, and no consistent rendering model, every frame has to be manually repainted using cursor movements and ANSI control codes. The ANSI color strategy is particularly interesting: rather than encoding brand colors as specific RGB values, the team mapped semantic roles like “eyes,” “goggles,” and “border” to ANSI color slots that terminals can reinterpret safely, allowing the banner to remain recognizable without assuming control over the user’s color environment.
Untapped Way to Learn a Codebase: Build a Visualizer — Jimmy Miller
This is the most interesting “how I read code” post I’ve read. Rather than explaining a codebase top-down, the author uses an unfixed bug in Next.js’s Turbopack as an entry point and builds custom visualisation tooling to trace how code flows through the system. The insight is that logging stops being enough when you’re dealing with complex async systems, at that point you need to see the graph unfold. By sending events over a WebSocket and building a live visualizer, he could watch values flow through Turbopack’s incremental computation system, see which cells were being invalidated, and understand why a tree-shaking bug was occurring. A longer read, but the philosophy of using tool-building as a learning mechanism is something I want to apply more.
Wikipedia Bans Archive.today After Site Executed DDoS and Altered Web Captures — Ars Technica
The strangest internet story of the year so far. In January 2026, archive.today embedded malicious JavaScript in its CAPTCHA page to use visitors’ browsers as DDoS attack participants against a security blogger who had investigated the site’s ownership. Because Wikipedia has 695,000 links to archive.today across 400,000 pages, essentially every Wikipedia reader who clicked one of those links was unknowingly contributing to the attack. The decision to ban was sealed when archive.today’s operators were caught altering their own archived snapshots, specifically inserting a critic’s name into a blog post to make it look like he had written certain comments, a direct tamper of the historical record that an archiving service’s entire value proposition depends on. All 695,000 links are now being replaced. The Wayback Machine remains unaffected and unrelated.