The things we
couldn't afford
to make.

For C. Girinandh's music video GAIA, Manish and With Media built two fictional civilizations. One draws on Chola architecture and sits underwater.

Their production breakdown ↗ is unusually specific. They modeled a city in Blender, placed cameras inside it, and used those views to guide the generated shots. Character reference sheets kept faces and costumes consistent. The composition was settled before the generation began.

The studio reports spending less than $1,000 on the video, mostly on generation credits and an upscaling license. That figure excludes recording and mixing the music. Preparation took weeks; the post doesn't put a price on that labor.

An underwater palace is a good use for this process. A small production can spend its time designing the place without having to finance its construction. Watch GAIA ↗.

Also on the watchlist: Robert Gaudette's 7:50 short A Face Only A Mother Could Love, Grand Prix winner at Runway's 2026 AI Festival. Festival lineup ↗.

Backplane is an open-source hardware workspace built around a KiCad fork, with KiCad and FreeCAD viewers and CAD-specific agent instructions. You can inspect an agent's changes and point to individual components while discussing a design. Code and installation ↗.

That could make electronics easier to learn: ask about a component, see where it sits, and follow the proposed change in the schematic. It gives the explanation something concrete to refer to.

The repository establishes what the tool offers. It doesn't establish that a novice can use it to ship reliable hardware. We'd like to see build reports next: design files, a fabricated board, measurements, and the mistakes along the way.

Simon Willison's browser-tool collection ↗ is full of small utilities built mostly with LLM assistance. He publishes the development transcripts and source code ↗, under an Apache-2.0 license.

One example: Photo Scrubber ↗ offers local face redaction and metadata removal. Its interface lets you cover faces the detector missed and explains why blurring alone doesn't guarantee anonymity.

This is a useful scale for software. Someone has an annoying problem, builds a tool, and leaves the code for the next person. It can be worth making even if nobody would pay a monthly subscription for it.

Willison is an experienced programmer. His collection is a good place to study how he works, including the checks he asks for, rather than assume the model did everything.

On October 28, Terence Tao described a new paper with Daniel Carter and Kevin O'Bryant on Sidon sets. In these sets of integers, every unordered pair has a different sum. How large can such a set be if its numbers must lie within a given range?

The paper grew out of a discussion on Tao's blog about AlphaEvolve improving bounds on that question. The researchers refined the approach, stalled for several months, and published what they had. Another method has since produced a slightly sharper bound; Tao says the method is the focus of their paper.

It's a modest, specific result with an account of how the work proceeded. The preprint gives other mathematicians something to check and extend. Tao's account ↗ · The paper ↗.

The AlphaFold Database now includes predictions of protein complexes: proteins interacting with each other.

EMBL-EBI reports 1.7 million high-confidence predictions for pairs of identical proteins, called homodimers. A May update added almost 80,000 high-confidence predictions for pairs of different proteins, or heterodimers. Lower-confidence results are available separately in bulk. Release details ↗.

Researchers still have to test these predictions. But they can begin with calculations that would be expensive to repeat in every lab. Making the results openly available is a consequential choice about who gets to use the work. Search the database ↗.