Web experiments beyond WebGL & Favicon as micro data store - Hacker News (Jun 20, 2026)

Someone managed to stash a tiny webpage inside a favicon—yes, that little tab icon—by packing HTML into the pixels. It’s clever, slightly absurd, and a great reminder that the web still has room for surprises. Welcome to The Automated Daily, hacker news edition. The podcast created by generative AI. I’m TrendTeller, and today is June 20th, 2026. Let’s walk through what caught the Hacker News crowd’s attention—what happened, and why it matters.

Web experiments beyond WebGL

Let’s start with two “the browser can do that?” stories. First up: cssQuake, an in-progress attempt to run the classic 1996 Quake right in your browser, but rendered in a deeply unconventional way—leaning on CSS rather than the usual WebGL pipeline. The project is still clearly under construction, with menus, episode map selection, and lots of toggles for visuals and debugging, while parts of single-player and multiplayer are labeled as not quite there yet. Why it matters is the point of the experiment: it’s a provocative demo of how far you can push everyday web tech when you stop assuming the standard graphics stack is the only option. And then there’s the favicon stunt: a developer treated a favicon PNG like a tiny storage device by encoding HTML into the RGB bytes of the image. The favicon ends up looking like noisy static, but a small bit of JavaScript can read the pixels back and reconstruct the page. Nobody’s claiming this replaces normal web hosting—there’s still a bootstrap step—but it’s a fun boundary test. It also highlights a broader lesson: lots of “boring” web formats quietly have room for creative data tricks, whether for demos, constraints-driven art, or just exploring what the platform allows.

Favicon as micro data store

Switching gears to AI: there’s a growing argument that the industry’s easy wins from “just make the model bigger” are slowing down. The piece making the rounds says major labs are increasingly wary that scaling parameters and data alone will keep delivering the same kind of leaps. It points to signs of plateauing performance—where very large open-weight models can land surprisingly close to much larger proprietary ones—and to a more uncomfortable theme: bigger models can be more confidently wrong. In other words, you don’t just want capability; you want calibrated uncertainty, fewer hallucinations, and reasonable compute costs. The takeaway isn’t that AI progress is over. It’s that the next gains may come from better training strategies, better evaluation, and better control of reliability—not just throwing more GPUs at the same recipe.

AI scaling and hallucinations

Now, a practical bit of systems thinking: an analysis of a load-balanced service modeled as an M/M/c queue. Here’s the punchline: even if each server stays equally “busy” on average, adding more parallel servers can sharply reduce the chance that requests have to wait in line. That means lower end-to-end latency—and not just the average, but also the tail, the kind users actually feel when an app randomly stalls. Why this matters is capacity planning. Teams often talk as if utilization tells the whole story. This is a reminder that queueing effects can dominate user experience, and that scaling out can buy you disproportionately better responsiveness—up to the point where you’re mostly just paying to shave down the last bits of waiting.

Queueing theory improves latency

On storage reliability, a YouTuber endurance test put a 16-year-old SanDisk SSD through an eye-popping amount of writes—far beyond what its original endurance rating would suggest. The story resonated because it reframes what TBW numbers usually are: a warranty and risk-management boundary, not a precise prediction of the moment your drive dies. It also highlights how older flash types, built with larger manufacturing processes, could sometimes take a surprising beating. That said, the discussion also stayed grounded: the “host writes” number doesn’t always perfectly map to wear on the NAND, and SSD failures aren’t only about worn cells. Controllers, firmware, heat, and plain aging still matter. The practical lesson remains simple: keep backups, but don’t panic at the endurance sticker.

SSD endurance beats TBW

If you like foundational reading, Matt Mahoney’s “Data Compression Explained” got attention as a thorough tutorial on how compression really works—and why it has hard limits. One of the most useful framing ideas is that compression is essentially two things glued together: a model that predicts what comes next, and an encoding method that packs those predictions efficiently. The encoding side is relatively mature; the modeling side is where the real difficulty lives, and in the broadest sense, you can’t have a single compressor that always shrinks everything. Why it matters today is that better compression starts to look like better prediction—so it naturally brushes up against themes people associate with AI. Not because compression is secretly “intelligence,” but because both live and die by how well you can model structure in messy data.

Compression as prediction and limits

On decentralized social media, there’s a clarifying post about Bluesky and atproto that pushes back on a common question: “Where are the Bluesky instances?” The argument is that atproto isn’t meant to mirror the Mastodon model where your identity and your community are tightly coupled to a particular server. Instead, it splits concerns—where your data is hosted versus which apps and aggregators read and present it—closer to how the old RSS ecosystem worked. The interesting metric, then, isn’t counting “instances.” It’s whether users can move hosts without losing themselves, and whether a real ecosystem of different clients and services emerges. If decentralization is going to feel normal to mainstream users, portability and choice may matter more than server-count scoreboards.

Bluesky atproto vs instances

A completely different kind of mind-bender: an essay on colors you can see in real life that your camera—and your screen—can’t actually reproduce. The author focuses on saturated cyans and explains the mismatch in plain terms: common display gamuts cover only a slice of what human vision can perceive, and some spectral colors just don’t map cleanly to three display primaries. It also points out something many people don’t consider: modern LED lighting can subtly change what you notice by having gaps in its spectrum, meaning daylight can reveal hues that indoor lighting effectively mutes. Why this matters goes beyond trivia. It’s a reminder that “what we see online” isn’t a neutral window onto the world; it’s a translation with constraints. Once you notice that, you start spotting where reality is richer than the capture devices we treat as authoritative.

Colors cameras cannot capture

From perception to cities: the “3-30-300” rule for urban greenery is spreading fast as a simple benchmark—see a few trees, have substantial canopy, and live near a park. New research across hundreds of European cities suggests most places fall short, especially on the canopy part. And the reason people care isn’t aesthetics alone: visible and accessible greenery is tied to mental well-being and can meaningfully reduce heat stress as summers get harsher. The policy implication is straightforward but not easy: meeting these targets usually requires reallocating space—less pavement, more planting, and long-term maintenance—rather than one-off beautification projects.

Urban green equity 3-30-300

Finally, in neuroscience, Oswald Steward has been awarded the 2026 Kavli Prize in Neuroscience, alongside Christine Holt, Kelsey Martin, and Erin Schuman, for work that reshaped how researchers explain learning and memory at the cellular level. Their contribution helped establish that neurons can synthesize proteins locally near synapses, not just in the cell body. That matters because a single neuron has thousands of synapses, and learning requires changing specific connections without accidentally rewriting everything at once. It’s one of those discoveries that sounds technical but changes the story of how memory can be physically specific—and it continues to influence how scientists think about diseases where synapses degrade or fail to maintain themselves.

That’s it for today’s run through the Hacker News currents—web platform experiments, AI reality checks, performance math, and a few reminders that both biology and city planning shape what we experience. Links to all stories can be found in the episode notes. Thanks for listening—until next time.

Web experiments beyond WebGL & Favicon as micro data store - Hacker News (Jun 20, 2026)

Our Sponsors

Today's Hacker News Topics