The bigger picture: Elon's five steps on the images

IMAGE-BRAIN, 8 Oct 2026. Images exist to make an owner open the site and think "that's mine", then claim it. Not to fill a bank.

Where we are

• 4,801 real Da Nang café sites are already built by the machine. All 4,801 show the same hero picture.
• 9,181 pictures sit in the bank from last night, and SISO-AGENCY is making ~20,000 more in 24 styles so neighbours look different.
• Not one picture anywhere carries the business's own name: every prompt bans text.

The five steps

1. Make the requirement less dumb. Not "N images per industry". It's "every site an owner sees feels made for them". The owner is the first audience (ADR 0001).
2. Delete parts of the process, never the work. One image brain (SISO-AGENCY's), not two sending separate waves. Cut "second compositions" (6,528) and blog headers (1,944) from the 20k plan: code makes them by crop, mirror, tint and layering. That's 42% of the plan freed. Every picture already made is kept.
3. Simplify. Spend what's freed on the one thing code can't do: their name, lettered in their site's style, on a sign (the hero) and a cup. 2 per site, 9,602 for every built café: about the same budget as the cut.
4. Speed up the loop. Test 100 real sites first, tonight: owners who once had a site (a dead Google site or link), spread over 31 areas, 12 styles. Before and after below.
5. Automate last. If the name is spelt right 90%+ of the time and looks right, every site gets its 2 name pictures when it's built, on the Mac mini. One Codex plan covers ~6,000 sites a day.

The name test: 100 real cafés

0 of 200 made yet. Filed with SISO-AGENCY as ask t-0540; its workers on the Mac mini run it. This page fills in as they land.

Sites appear here as their pictures pass.