This layer is for high-novelty bets that are worth writing down before they look normal.
These pages are not claims of established originality in the global literature. They are a stricter local standard:
- not just a restatement of one paper
- not just a shallow interpolation of two adjacent papers
- shaped around the actual Parameter Golf objective: quality per stored byte under a hard artifact cap
- paired with a cheap falsifier so they can die quickly if they are fake-deep
If frontiers names the best seams suggested by recent papers, moonshots ask:
what would we pursue if we stopped assuming the winning artifact still has to look like a normal compressed checkpoint?
Epistemic status
These notes should be read as:
- original in this garden’s current graph
- outside the strongest visible priors we have already collected
- not guaranteed globally novel across all unpublished, obscure, or adjacent literature
For that reason, each page should be judged by:
- whether it opens a genuinely different mechanism space
- whether it is challenge-native rather than trend-chasing
- whether it gives a falsifiable experiment plan
For the novelty discipline itself, see Originality and novelty discipline.
Current moonshots
Artifact-native objectives
Alternative model representations
Alternative architectures
How to use this layer
Use moonshots when:
- the main graph is converging too hard on familiar priors
- you want mechanisms that might look weird but are tightly matched to the challenge objective
- you want ideas that can fail usefully and teach us something structural
Do not use moonshots as a dumping ground for vague sci-fi. Every page here should still answer:
- what bytes does this save?
- what capability does it try to reconstruct?
- what is the cheapest proxy test?
- what result would kill it?