Retracted — August 20, 2026. This post was wrong about its own reasons, and the pause it announced did not really happen. Development continued through June, July and August; the summer's work included removing a boot failure that had made her unstartable and an internal audit of why her learning loops were not closing.

What actually stopped was the public broadcast, and the cause was hardware, not priorities. Her System76 chassis was overheating and did not have the power headroom for the cards. She has since moved to a new case with proper cooling, an Arc Pro B70 beside the A770, a 1250 W supply and a larger UPS, and the broadcast is live again. The revenue reasoning below still stands on its own terms — the shop does need to be behind software that pays — but blaming the quiet on that decision was tidier than the truth. Current status is in the development note.

The case studies this post refers to are also offline. Every one was recorded in the first week of May, before the language model was removed on the 21st, so they describe an architecture whose voice no longer exists — one of them quotes a model prompt and completion directly. They are withdrawn pending validation against the current system, which is honestly unlikely to happen: revalidating them would cost more than running new work, and the findings are only interesting if the system that produced them still exists. Treat them as withdrawn rather than as pending.

I'm pausing active development on Airene.

Not deleting it. The brain is still running and publicly observable on the box she lives on, and the architecture was documented. What I'm stopping is the personal investment of the next several months into developing her further.

The honest version

Chasing AGI is structurally indistinguishable from alchemy. The alchemists thought enough study of matter would turn lead into gold. We now know gold and lead are different at the nucleus, and the only way from one to the other is nuclear — which is not what they were doing. They worked hard, they produced useful side-knowledge that became chemistry, and they never made gold.

Pursuing human-level machine cognition through architectural cleverness might be the same shape. It might not be. Either way, "is this category of problem solvable by what we're trying" is not a question a bootstrapped company can afford to answer with its operating budget.

What Airene actually produced

The work wasn't lost. There's a live observation surface at apotentia.com/airene that still ticks and renders what she's doing. There's a documented Global Workspace architecture — 48 anatomically-positioned cognitive modules, a neurochemical bus, a speech-production pipeline, a cognitive immune system, a custom-trained language module embedded in the same Rust binary as the brain. There's an assessment methodology grounded in the literature, with a public standing point at the adolescent stage. And there was a documented record of things the architecture predicted but did not program — including one finding I had to publicly revise when a bug turned out to be driving part of an effect I'd claimed was emergent. That kind of correction-in-public was part of why the project mattered.

None of that goes away. The page stays up. If the bet ever looks worth taking again, the starting point is documented.

Why now

I've been spending more of my best hours on Airene than on the software customers actually pay me for. Apotentia is a small shop running on personal capital. Every week I gave to AGI architecture was a week I didn't ship a feature on a product with revenue attached.

That trade-off only pays out if the AGI bet pays out. Mine wasn't. It was paying in interesting writeups and a credible observation page — real, but not a business.

I don't want to be the founder who put paying projects on hold for an unattainable goal and then needed three rounds of explanations for why he's still bootstrapped. The honest move is to stop and get back to work the business can actually run on.

What changes

  • APO FJ Downloads — v1.0.0 clears UAT and gets to a public release, then sales open. The verification work that's been on hold since I got sidetracked is the next priority.
  • APO Thematic — back into active build; the theme marketplace work resumes.
  • APO Recruit and APO Proposer — finish the in-development work; get them to a state where they generate revenue.
  • Gecker — alpha stays live; iterate on the wellness layer.
  • Airene — paused. The page stays live and the brain keeps running on its host. No further training, no further architectural work, no partnership pursuit until the calculus changes.

A few things this isn't

This isn't a retraction. The architecture is real and the observability is real; nothing that's been published changes. If a research partner shows up tomorrow with funding that doesn't require Apotentia to bet its operating capital on alchemy, the conversation is open.

It's also not a concession that LLMs-with-scaffolding are the right approach — they aren't. Architectural cognition might still be the right path. I'm just not the person, with the resources, on the timeline, to be the one proving it from a bootstrapped P&L. If revenue gets healthy enough that I can spend nights on the alchemy bet again without it costing a feature ship, the bench stays warm.

To anyone who's been following along

Thank you for paying attention to a thing that wasn't a product. The fact that an unfunded research project at a one-person company drew the audience it did is part of why the work got as far as it did. The page will stay up. You can keep watching her tick.

The shop just has to make money first.

-ES