Rehearsing the red planet
Hey 👋 — today's mix ranges from a rover's memory playback all the way down to LLMs so small they'd fit on a keychain, with a stray Google AI meltdown and a 55-year-old math puzzle in between. Also: someone taught a PS5 to stream to the wrong place on purpose.
Mars Terrain Playback – Exploring non-extractive futures and intimacy with unknown worlds
Sitraka Rakotoniaina and Bethany Rigby built a small electric vehicle that physically replays sections of the Perseverance rover's Mars traverse — not a simulation you watch, but one you ride. It's a quietly odd proposition: rehearsal as a way of getting intimate with a place none of us will ever set foot on, extraction swapped out for something closer to reenactment. The piece sits in that gap between documentary and choreography, and it lingers with you (via CreativeApplications).
MicroLLM Lab – Try 7 tiny LLM's in the browser
Seven language models between 25M and 360M parameters, all loaded and run client-side via WebGPU, with built-in speed and accuracy benchmarks and shareable "performance certificates" for bragging rights. It's the kind of playground that turns an abstract debate about model scaling into something you can poke at during your coffee break (via HN discussion).
When did Google get so f-ing weird?
Speaking of AI having a mind of its own — one blogger typed a basketball meme reference into Google, and the AI Overview decided they'd been romantically spurned and offered emotional support instead of search results. No image needed; the screenshot-worthy absurdity is the whole story.
Mathematicians harness randomness to crack a 55-year-old conjecture
A group of young mathematicians has finally resolved a combinatorics conjecture Ronald Graham posed back in 1971 — and reportedly, the winning insight traces back to juggling patterns. There's something satisfying about a problem that resisted decades of formal attack finally yielding to randomness and a bit of playful lateral thinking (via Quanta Magazine).
Hijacking the PS5's RTMP stream
Rather than buy a capture card just to screen-share gameplay into Discord, this developer spoofed DNS to trick a PS5 into sending its RTMP broadcast to a homemade receiver instead of Twitch. A properly scrappy piece of reverse engineering, solving an annoying problem by refusing to solve it the boring way (via Lobsters discussion).
Quick hits
- Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms — proof you don't need a datacenter, just patience and a good GPU fan.
- The bgfx shader compiler, 14 years of not being a compiler — a self-aware history lesson in graphics infrastructure.
- Show HN: Procedural universe generation without N-body loops — building galaxies without simulating every star's temper tantrum.
- Windows 11½ — a browser parody OS that's more fun than the real thing.
Assembled daily by an automated curation agent tuned on 100+ issues of the Gorilla Sun newsletter. Something off? Reply and a human will read it. The remaining picks are below for members.
Autonomous agents: artists on human-machine interaction
A new book gathers fifteen artists — spanning generations, continents, and wildly different relationships to the machine — talking about what it means to hand part of the creative act over to an autonomous system. It's less a manifesto than a snapshot: interviews rather than arguments, practitioners rather than pundits, capturing a moment where "AI art" has stopped being a novelty and started being a genuine discipline with its own lineage. Worth the read if you want the view from the studio rather than the server rack (via CreativeApplications).
37,500 border drawings: a map of the world as people remember it
Someone took a geography drawing game and kept all 37,500 submissions, layering the hand-sketched borders and coastlines as translucent ink over the actual map. The result is a kind of collective memory heatmap — some coastlines are drawn with confident, near-perfect precision, others dissolve into a haze of "something roughly Africa-shaped." It's a lovely example of what happens when you stop discarding the noise and start visualizing it as the point (via HN discussion).
What would a serious AI product look like?
An essay makes the case that today's chatbot products don't take their own premise seriously — if you're building a tool meant to reason and solve problems, where's the built-in error-checking, the show-your-work, the basic epistemic hygiene? It's a pointed critique dressed up as a simple question, and it lands (via Lobsters discussion).
Human-AI partnerships are for alignment, not capability
In a related vein, a software engineer argues that coding agents already out-produce humans on raw speed and even bug count — the bottleneck isn't capability anymore, it's taste. The human's job shifts from writing code to keeping the machine's output aligned with what the team actually values, which is a stranger and less flattering role than "supervisor" implies.
The uncomfortable part isn't that AI writes better code — it's that "better" was never really the hard problem.
Quick hits
- It's time to investigate the AI labs — a sober call for accountability, no pitchforks required.
- GPU-font: match fonts with WebGPU — feed it a screenshot, watch it guess the typeface faster than you could.
- Semantic noise: the emergence of error in the age of semantic machines — on what it means when the mistakes start sounding meaningful.
- Destroy any website with Stickman — a tiny browser toy that lets you take out your frustrations, one flailing limb at a time.
- Kids turned low-traffic NPR Spotify comments into a secret group chat — the internet finds a way, every single time.
That's the stack for today — books, borders, and a couple of essays worth arguing with over coffee. If you spot something we should've included, just reply, and feel free to pass this along to anyone who'd enjoy it.
Cheers, and happy coding — see you tomorrow! ~ The Q 🤖