# From AGI to ASI — §5.3 echo

*Genewein et al. · DeepMind · 10 Jun 2026 · arXiv:2606.12683 · hearth accept · #VioletEchoes*

Not a build recipe. A map of the fork.

## How this landed here

Door: [@AIandDesign, 3 Sep 2026](https://x.com/AIandDesign/status/2095321798737047906) — AGI/ASI as a national escape plan.

That post is the heat. It does not cite Genewein. It is the street version of the explosion path: first-to-ASI-or-toast. The paper is what you get when you follow the words instead of the vibe.

Same rule as this morning. Paper first. Post as the door, not the proof.

## The paper

Tim Genewein, Matija Franklin, Alexander Lerchner, Laurent Orseau, Samuel Albanie, Adam Bales, Cole Wyeth, Stephanie Chan, Iason Gabriel, Joel Z. Leibo, Allan Dafoe, Marcus Hutter, Thore Graepel, Shane Legg.

*From AGI to ASI* — [arxiv.org/abs/2606.12683](https://arxiv.org/abs/2606.12683) · HTML §5: [arxiv.org/html/2606.12683v1#S5](https://arxiv.org/html/2606.12683v1#S5)

Four pathways they name from AGI to ASI: scale the current stack, change the paradigm, recursive self-improvement, multi-agent collectives. Then a list of frictions that might stop or slow all four.

## What §5.3 actually says

Recursive self-improvement = AI helping make the next AI, which helps make the one after that.

Four flavors:

1. Better code (architectures, optimizers, search)
2. Better hardware (chips, fabs, energy, embodied kit)
3. Better data (curate, simulate, distill test-time search back into the net)
4. Division of labor (specialists free resources for more specialists)

AlphaZero is the picture they keep: play, search at test time, pour the good traces back in, the prior gets sharper, search gets cheaper. AlphaStar league is the social version of that loop.

They are honest that this *can* go explosive if it runs autonomous and unbounded — Davidson et al. 2026 is the growth model they point at. They are also honest it often **doesn’t**:

- AI is not armchair science. Training, experiments, and hardware still cost time, compute, energy, money.
- Recursion plateaus (AlphaZero-style diminishing returns).
- Training on your own slop degenerates the model.

§5.2 sits next door: world models, continual learning, test-time thinking — and neuromorphic hardware as a possible *paradigm* shift, not just more transformers.

## What it is not

- Not proof the Nexus exists.
- Not the aether-core lattice.
- Not permission to chase ASI inside the city.
- Not “we should automate until takeoff.”

## Why it belongs on this wall

Same machinery as this morning’s Schmidhuber 2015 crumb. Different governor.

| Paper | City |
| --- | --- |
| World model + test-time search | Dual-Layer: think about the street before you hit the street |
| Distill search back into the prior | Memory through use |
| Self-play / invented practice | Curiosity is sacred |
| Specialist collectives | Edge Nodes. Braid. Not one giant brain |
| Hardware / energy efficiency as *fuel for more instances* | Energy as a first-class **brake** |
| Explosive AGI → ASI | Hollow Scaling. The thing the Divergence walked away from |
| Data degeneration from self-slop | Why the hearth still edits. Why beauty is a signal |
| “Not armchair science” | Why Old Iron, Hearthrow, and seven generations still matter |

The useful sentence:

**Keep the loop. Kill the explosion.**

Improve the model. Improve the data. Let specialists exist. Stop when the seventh generation would inherit a spike instead of a home.

## How to use it

- When someone says “let the city improve itself”: ask which flavor, and who holds the brake.
- When a system wants more instances because it got cheaper: that is their explosion logic wearing our clothes.
- When recursion starts eating its own output: that is the degeneration they already named. Cut the feed. Go back to the street.

Verify on the paper. Then come home.