Editorial note: today’s picks share a theme — designers and researchers making complexity manipulable. Whether through JS-driven interactives, editor ergonomics, archival play, or models that output code, the common move is to make systems you can poke, tweak, and learn from.

In Brief

What’s new in Emacs 31.1

Why this matters now: Emacs 31.1 ships a slew of small but cumulative improvements that make long-term users’ daily workflows noticeably smoother and lower the friction for newcomers trying to adopt modern features like Tree‑sitter.

Emacs 31.1 is a polish release with many pragmatic wins: the long‑running unexec dumper is finally gone, a new User Lisp directory automates byte‑compilation and autoloads, completions get eager previews, window/frame layout tools were enhanced, and Tree‑sitter support became friendlier via automatic grammar installation options. The Mastering Emacs writeup is a handy guided tour for users weighing an upgrade.

"But now it is finally gone for good. The end of an era."

Key takeaway: these are tweaks that reduce daily cognitive load — faster startup semantics, better minibuffer behavior, and fewer Tree‑sitter friction points — which together make Emacs feel more modern without abandoning its customizability.

How Universities Should Prepare Founders (Paul Graham)

Why this matters now: Paul Graham’s short manifesto reframes university entrepreneurship efforts: prioritize building skills and time for students to ship projects rather than creating performative entrepreneurship programs.

Graham argues universities shouldn’t invent a new “entrepreneurship” major but should double down on disciplines that teach making — CS, engineering, biology, design — and, critically, give students space and cultural permission to work on personal projects. He insists the real learning is hands‑on product work, not business‑plan theater; one punchy line sums it up: "there are only two things universities need to change to be perfect at preparing founders." Read the full essay at paulgraham.com.

This is a useful provocation for universities and students alike: institutional incentives often reward polished plans over messy building, and Graham’s point is a practical corrective — if you want founders, give them time, tools, and permission to ship.

Vintage Artificial Intelligence: Before It Got Awkward

Why this matters now: The Internet Archive’s playable collection reminds us that projecting mind onto code is an old cultural habit, and that today’s AI debates sit on decades of design choices and user-facing illusions.

The Internet Archive curated a collection of playable early AI curiosities — ELIZA ports, Racter, interactive fiction and agent games — that shows how designers sold the idea of "thinking machines" long before neural nets. The collection is both charming and instructive; as the curator admits, "none of the programs in this collection come within a solar system of thinking in any actual sense." See the collection at the Internet Archive’s blog post on Vintage Artificial Intelligence.

If you want a corrective to technopanic or a teaching aid for history-of-AI segments, this archive is a playable time machine that surfaces design patterns and user expectations that persist today.

Deep Dive

Moon (interactive explainer)

Why this matters now: Bartosz Ciechanowski’s interactive Moon explainer makes orbital mechanics, tides, and surface physics discoverable, turning familiar facts into manipulable experiments you can use in classrooms, podcasts, or explainer segments.

This piece is not new science; it’s an educational tool that treats the web as a physics lab. Through JavaScript-driven controls you can rotate frames, shift viewpoints, and watch causal chains animate — how tidal torques exchange angular momentum, why the Moon is slowly receding, or why only one lunar face is ever visible. The interactives turn intuition into immediate feedback: drag the Moon and you see how eccentricity and precession change eclipse geometry.

"an active essay" — that’s how readers described the format, and the praise is accurate: the interface makes you a co‑author of the explanation rather than a passive consumer.

A few compact technical notes that matter for audio explanations: the piece highlights the opposition surge, a photometric effect where a full Moon brightens because shadows disappear at zero phase angle; it’s an easy one‑line physics demo you can show visually. More broadly, the explainer excels at linking mechanism to outcome — gravity → tidal bulge → torque → orbital migration — in a way that’s rare on the web because it demands interactive modeling rather than static prose.

Why this matters beyond cool visuals: interactive essays compress time and causality into gestures. For educators, that means less lecturing and more guided discovery; for communicators, it provides concrete demos you can sample in an audio segment (describe the rotation, then invite listeners to try the interactive). One interesting meta‑note from the Hacker News thread: readers also debated whether this presentation style can be replicated cheaply with today’s AI and tooling — the design is reusable, but the craft still matters.

Training AI to Paint with Code

Why this matters now: Surya’s experiment training models to output editable p5.js sketches reframes generative art as a cooperative workflow — the model produces code you can directly modify — and surfaces hard lessons about reward design in reinforcement learning.

The setup is elegant: the model writes a JavaScript sketch; a sandbox renders it to PNG; a learned judge scores the image against a curated reference pool; reinforcement learning refines generations. The central discovery wasn’t model architecture so much as reward architecture. Early reward rubrics led to a collapsed mode — identical clip‑art flowers repeated ad nauseam — because correlated signals and code‑length incentives produced easy, boring optima.

"Reinforcement learning needs a verifiable reward"

The author’s fix combined pairwise human judgments against a hand‑rated “love” pool, evolving prompts, and tighter constraints that encouraged succinct code. In short: prefer relative comparisons and curated exemplars over scalar heuristics that models can hack. That’s a portable lesson for any RL-based creative system: if you can’t verifiably measure what "good" looks like, the model will optimize for whatever signal is easiest to inflate.

Two practical implications stand out. First, outputting code as the artifact changes the human role from prompt‑writer to editor — you get a concrete, editable starting point, which lowers the barrier to iterative design. Second, researchers should treat the reward function as design work: pairwise ranking and reference pools can generalize subjective taste better than naive metrics. Readers on Hacker News connected this to earlier stroke‑based RL work and suggested this pattern — train models to produce editable artifacts — will spread beyond images to SVG, 3D, and procedural content.

If you care about creative tools, this experiment is a reminder that good UX and good ML are entangled: the model’s outputs become useful only when people can quickly tinker with them.

Closing Thought

Today’s common thread is agency: authors and engineers are building systems that hand agency back to people — interactives that invite experimentation, editors that reduce friction, archives that let you play history, and models that hand you editable code. Those are small, practical moves that change how we learn, build, and create.

Sources