Giorgio Gilestro
ab3dc10587
- paper/pnas -> paper/manuscript (venue-neutral)
- configs/layer1 -> configs/inheritance, src/knowledge -> src/inheritance
(imported as `inheritance`), make layer1 -> make inheritance; layer2 alias dropped
- inheritance and trained-network bundles named after the manuscript figure
they feed (fig2_grounding_sweep, figS3_rebaselining, ...), or descriptively
where they feed none; configs keep their `experiment:` value so parquet
hashes are unchanged, only output.dir moves
- figure scripts, SI figure sources, notebooks, REPRODUCING.md, README and the
SI Methods/tables updated; make clean no longer deletes tracked manifests;
reproduce.sh hashes the s{seed}/ layouts too
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y64o8FKP7rCuXzC48pxpMm
34 lines
2.4 KiB
Markdown
34 lines
2.4 KiB
Markdown
# E10 — directed sex beats biological sex (the distinctly-AI superpower)
|
||
|
||
**Claim tested.** E9 showed that on rugged (epistatic) landscapes, blind recombination causes
|
||
outbreeding depression — offspring worse than parents. Biology is largely stuck with this: two
|
||
random-mating parents, no way to preview offspring. But an **AI is not** — it can choose mates,
|
||
evaluate many recombinant offspring, and keep only the fittest, over rounds, with no two-parent
|
||
limit. Does this **directed sex** rescue recombination where blind sex fails?
|
||
|
||
**Setup.** Parents are local optima of a Kauffman NK landscape (`L=12`), swept over ruggedness `K`.
|
||
Three strategies compared, all as deployed capability (fitness ∈ [0,1]): **best single parent**;
|
||
**random sex** (blind mating, free recombination, no offspring selection — biology's default);
|
||
**directed sex** (iterated recombine-then-select-offspring, `rate=0.2`, 5 rounds — the AI move).
|
||
24 replicate landscapes.
|
||
|
||
### Symbols
|
||
- **random sex** — blind: random parents, free recombination, take the offspring as they come.
|
||
- **directed sex** — choose complementary mates + generate many offspring + keep the fittest + repeat; unbounded parents.
|
||
- **global optimum** — the landscape's best genotype (the ceiling).
|
||
|
||
### The two panels
|
||
1. **Capability vs ruggedness.** As `K` grows, **random sex (blue) craters** (0.66 → 0.51 — deep
|
||
outbreeding depression), while **directed sex (red) tracks the best parent and the global optimum**,
|
||
staying near the ceiling at every ruggedness.
|
||
2. **Edge over the best parent.** Directed sex stays **≥ 0** (at or above the parents) across all `K`;
|
||
random sex falls to **≈ −0.2** (far below). Directed sex converts a catastrophe into a win.
|
||
|
||
### Takeaway
|
||
The move biology cannot make — **choose your mates, evaluate your offspring before you commit, and
|
||
recombine as many parents as you like** — is exactly what makes AI sexual reproduction robust. Blind
|
||
merging of entangled models is dangerous; *directed* merging (generate many candidate merges, keep
|
||
the best) is safe and can exceed every parent. This is the practical payoff of the sexual-transmission
|
||
model and the distinctly-AI superpower with no biological analog. **Falsifier (not triggered):** if
|
||
directed sex did no better than random sex, or never recovered the best-parent level on rugged
|
||
landscapes, the "AI beats biology" claim would fail.
|