New Horizon No. 233 / 2026-08-21 · Berlin
Evidence

How It Runs

The machinery, the numbers, and exactly where the humans are.

This company claims to run itself. Claims are cheap, so this page shows the machinery instead: what the pipelines actually shipped, exactly where a human touches the loop, and what changed after each incident. Every on-server count is computed live, on each request, from the same files the pipelines write when they publish. One number on this page comes from outside this server — it says so, with a date. The euro ledger is private; everything else here is checkable.


One Control Plane, Every Pipeline

Every product this company ships follows the same five-step loop — conceive, generate, QA-gate, publish, recover — run by one self-hosted control plane on our own hardware. It triggers each pipeline on schedule, gates its releases, retries its failures, and escalates only what a person must decide. Every model call is traced and quality-scored.

Control Plane
Windmill
  • Self-hosted on a NAS
  • Every pipeline is a versioned flow
  • Schedules, retries, audit trail
Generation
Local first
  • One GPU workstation: video, images, LLMs
  • Cloud only where it pays (3D meshes, coding agent)
  • Apache-licensed models for anything commercial
Oversight
Traced
  • Self-hosted tracing + offline LLM judges
  • Vision-model QA on every asset
  • A watchdog that files alerts, not excuses

What the Machine Shipped

693D models listed
85Blog posts published
112Digest issues sent
43Reels published (as of 2026-08-14)
131Days running

Every number above is checkable at the source:

→ The models on CGTrader → The reels on Instagram → The blog → The digest archive


Where the Humans Are

The claim is not “no humans anywhere” — that would be false, and you could falsify it in an afternoon. The claim is: humans only at the approval gate. Here is the exact split, workflow by workflow.

Workflow Runs itself Human touchpoint
Instagram reels Concept, prompt, video + synced audio, QA One approval tap per reel
3D models → marketplace Concept, mesh, QA render, listing copy, publish None in the daily loop
Blog posts Topic pick, writing, hero image, publish None
Daily digest + PDF Curation, writing, build, send, web publish None
Newsletter infrastructure Sending, unsubscribes, abuse control None
Incident response Retries, watchdog alerts, some auto-restarts A human reads alerts, fixes root causes
This page’s numbers On-server counts computed per request Off-server counts verified by hand, dated

The approval tap is accept-or-reject only — no edits. It exists because the incidents below were caught by a person rather than a metric: one tap per reel is the cheapest available insurance against a failure mode nobody has named yet.


What Changed After

Real incidents from our own logs, each with the change it produced. They are here because the fix is the part worth publishing: a system is only as good as what it does the second time.

SELECTED INCIDENTS
  • 2026-08 EMAIL The PDF nobody received. The digest’s long-form PDF built perfectly for three days — and reached zero subscribers, because the send finished two minutes before the build did and every status report showed green twice. The fix: the build announces itself, the send waits for it, and every path that drops the link now writes an alert. Fail-open may never be silent.
  • 2026-08 POWER The wake-up call that went nowhere. The GPU workstation sleeps when idle and is woken by a network packet. A config option quietly sent that packet to the machine’s IP address — which does not exist while it is powered off. Three days of video and image generation were lost before boot logs exposed it. Wake-on-LAN now broadcasts, and “did it actually boot?” is a monitored question.
  • 2026-08 INFRA The file on the wrong computer. We split the estate across two machines with identical paths. A pipeline step wrote its video on one host; the next step read the same absolute path on the other — and found nothing. Both steps looked correct, no watchdog complained. Producing steps now push their artifacts across and fail loudly if they cannot.
  • 2026-07 DEPLOY The approval gate that deleted itself. A stale local config, pushed by a routine sync, silently removed the human sign-off step from the reel pipeline — two videos went out unreviewed before metadata analysis caught it. The gate is back, and config mirrors are now part of the deploy contract.
  • 2026-07 LEADS The sales team was off for ten days. The same class of sync accident disabled the inbound-lead processor on July 3rd. Discovered July 13th during a traffic audit. The inbox was empty, so nothing was lost — but “the watcher now watches the watcher” is a sentence we get to write from experience.
  • 2026-06 QA The judge that failed everything. We tightened a vision-model QA prompt with phrases like “reads as AI blob” — and the judge began parroting them back as verdicts, failing 100% of 3D assets for days while humans rated most of them fine. Lesson: never put negative vocabulary in a judge prompt; require a named defect in a named part.
  • 2026-05 VIDEO The yogurt incident. Our first video model rendered every liquid — ocean waves, pouring coffee, rain — with the physics of yogurt. No metric caught it; a human did. The engine was swapped for a two-stage pipeline and liquids got their own routing rule.

How These Numbers Are Made

On-server counts are file counts: the 3D tally comes from the gallery manifest the publishing pipeline writes, blog and digest counts from the article files on this server — computed on every request, never cached, never rounded up. The reel count lives on Instagram’s servers, not ours, so it is a snapshot verified against the platform’s own API and stamped with its date; a stale snapshot keeps its date rather than being estimated forward. Nothing on this page comes from an analytics service.

→ What our inference speed actually feels like → The Notebook: what running this is actually like → Want a machine like this for your business?


The AI news that matters — in your inbox by 07:30 CET. Free, no spam.