The Denominator: art direction reference sheet
The question from Ben (2026-10-05 22:06 ET): our current images feel like Fast Forward Labs but too simple, and Increment images feel colorful and chaotic with no link to the article. What should our art design language be?
Ray's proposal is conceptual illustration: each image is a metaphor that carries the article's argument, made with real craft. The set below covers a range on purpose, so Ben's picks decide the direction. Sheets: sheet-1.png (01-06), sheet-2.png (07-12), sheet-3.png (13-18), sheet-4.png (19-24). Full-size captures are in shots/.
Not included: The Economist and Bloomberg Businessweek blocked headless access (403 / bot wall), so they were skipped rather than worked around. The Atlantic (Altman profile) was captured but cut because its image carried weak meaning.
| # | Source + URL | Article (1 line) | Style tag | How the image carries meaning |
|---|---|---|---|---|
| 01 | The New Yorker - https://www.newyorker.com/tech/annals-of-technology/chatgpt-is-a-blurry-jpeg-of-the-web | Ted Chiang: an LLM is a lossy compression of the web | conceptual | A face made of scattered letters dissolves at the edges, so the image shows "lossy text compression" directly. |
| 02 | Quanta - https://www.quantamagazine.org/the-unpredictable-abilities-emerging-from-large-ai-models-20230316/ | Emergent abilities appear suddenly as LLMs scale | conceptual (animated) | Random letters drift and then cluster into words, showing order emerging from scale. |
| 03 | Quanta - https://www.quantamagazine.org/does-computer-science-need-computers-20260828/ | Theoretical CS does not need machines, but machines posed its questions | collage | Torn strips of logic gates, proofs and vintage hardware are woven together, so theory and machine are shown as one fabric. |
| 04 | MIT Technology Review - https://www.technologyreview.com/2024/03/04/1089403/large-language-models-amazing-but-nobody-knows-why/ | LLMs work, but nobody fully knows why | conceptual (engraved texture) | An eye peers out through a keyhole: we only see the model through a narrow opening. |
| 05 | MIT Technology Review - https://www.technologyreview.com/2026/09/28/1145197/whos-liable-when-ai-agents-go-rogue/ | Law lags on liability when AI agents misbehave | collage | Handcuffs close around an AI logo and an eye, so the image asks who gets arrested when an agent goes rogue. |
| 06 | Noema - https://www.noemamag.com/artificial-general-intelligence-is-already-here/ | Agüera y Arcas & Norvig: frontier models are already AGI | abstract | A glitchy orb against fractured color reads as "intelligence", but the specific argument is weak; this is the abstract end. |
| 07 | Asterisk - https://asteriskmag.com/issues/15/what-to-expect-when-you-re-expecting-a-world-of-multi-agent-systems | Modeling future work with a village of AI agents | geometric | Little geometric robots build and work in a toy village, so the "AI village" metaphor is the whole picture. |
| 08 | Asterisk - https://asteriskmag.com/issues/15/what-will-happen-to-indian-it | Past and future of India as the "world's back office" | two-tone | A lone worker at two monitors is cut out of flat red, a single figure that stands for a whole outsourced industry. |
| 09 | Asterisk - https://asteriskmag.com/issues/15/all-work-and-no-play | Issue 15 opener (theme: work and automation) | riso texture | A worker in a whirlwind of robot arms and screens, with grain and limited ink colors; shows work being reshaped around him. |
| 10 | Works in Progress - https://worksinprogress.co/issue/weapons-of-mass-decentralization/ | Iron toppled Bronze Age empires; could drones do the same? | engraving | A dense symmetrical engraved structure reads as an ancient fortified order, the empire that cheap technology could break. |
| 11 | Works in Progress - https://worksinprogress.co/issue/just-bury-your-trash/ | What if what we know about recycling is wrong? | fine-art collage | Garbage bags float in a 19th-century landscape painting; the clash shows the article's contrarian idea about land and waste. |
| 12 | Distill - https://distill.pub/2020/circuits/zoom-in/ | Interpretability: zoom in on neural network circuits like a microscope | engraving / technical | Hooke's Micrographia sits next to a neuron visualization, so the article's "microscope for neural nets" analogy is the image. |
| 13 | Fast Forward Labs - https://ff13.fastforwardlabs.com/ | Report: Causality for Machine Learning | minimal tech | A node-and-edge graph on a plain cover is literal (a causal graph) but carries no tension: this is the current baseline. |
| 14 | Fast Forward Labs - https://ff06-2020.fastforwardlabs.com/ | Report: Interpretability | minimal tech | A black box with a table coming out of its side is a clean, diagrammatic metaphor at low craft. |
| 15 | Stripe Press - https://press.stripe.com/ | Stripe Press book catalog | print object | Marbled and engraved book spines get their meaning from materials and craft, not a picture of the topic. |
| 16 | Increment - https://increment.com/software-architecture/ | Issue on software architecture | flat editorial | Figures sketch inside an impossible staircase building; loosely "architecture", mostly mood. |
| 17 | Increment - https://increment.com/programming-languages/ | Issue on programming languages | chaotic | A Rube Goldberg tangle of shapes and creatures: energetic, and gives no hint of the topic (Ben's "too chaotic" pole). |
| 18 | Bartosz Ciechanowski - https://ciechanow.ski/mechanical-watch/ | Interactive explainer: how a mechanical watch works | technical 3D | A clean interactive render of the actual mechanism: the image is the explanation itself, not a metaphor. |
| 19 | The Pudding - https://pudding.cool/2017/05/song-repetition/ | Are pop lyrics getting more repetitive? | data-as-image | The hero is the finding: a histogram with "This Essay" and "Cheap Thrills" marked at opposite ends. |
| 20 | Wait But Why - https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html | The AI Revolution, part 1: exponential progress | low-fi | A stick figure stands where a progress curve starts to bend upward; crude drawing, sharp argument. |
| 21 | xkcd - https://xkcd.com/1838/ | "Machine Learning" comic | low-fi | Someone stirs a pile of linear algebra until the answers look right, so a whole critique fits in one gag. |
| 22 | Julia Evans (Wizard Zines) - https://wizardzines.com/comics/dns-records/ | Zine page: the 5 parts of a DNS record | zine | Hand-lettered annotations on a real record: explanatory, warm, and dense with meaning. |
| 23 | Maggie Appleton - https://maggieappleton.com/garden-history | History and ethos of the digital garden | watercolor | Plants grow from a browser window: "garden" as the publishing metaphor, drawn with care. |
| 24 | Google DeepMind / Unsplash - https://unsplash.com/blog/visualizing-ai-with-google-deepmind | Visualising AI: artists render AI concepts as 3D art | 3D clay | Soft clay forms and tactile renders give AI a physical feel; meaning varies by piece, and the craft is high. |
Founder votes (2026-10-06 05:07 ET, iMessage)
"I don't really care for any of those illustration styles. None of them are jumping out at me."
- 01 creepy. 02-05 okay but not clear / not a good style. 06 unclear. 07-09 okay but not a style he'd use.
- 13 + 14 (Fast Forward Labs) = good for research booklets.
- 16 (Increment, flat editorial) = "may be the closest to a good one."
- Next: round 2 sheet narrowed to the 16 family (flat editorial, hand texture, small figures in a space). Asked him A texture / B figures-at-work / C warm palette / D all.