embedding
A concept becomes one normalized 384-dimensional local vector. It is evidence about learned language relationships, not a definition of the word.
signal / live study
LATENTGARDEN
What if meaning could reproduce after language disappeared?
Give two seeds a concept. A small local embedding model compresses each meaning into a bounded procedural genome. The browser grows the organisms, crosses their genes, introduces mutation, and lets descendants continue after the source words are gone.
germinate two meanings · cross them · mutate the descendant · plant what survives
two procedural seed forms are ready / local semantics has not been activated
generation01
child seedlg1_…
originprocedural cross
growthforming
Until local semantics is activated, these two parents use deterministic procedural preview genomes. Activating the model replaces those previews with fixed projections of the two local embedding vectors.
descendant genome
12 bounded traits / 0–100
- branching
- 50
- symmetry
- 50
- curl
- 50
- reach
- 50
- density
- 50
- taper
- 50
- phyllotaxis
- 50
- gravity
- 50
- drift
- 50
- porosity
- 50
- persistence
- 50
- spectrum
- 50
Concepts are sent only to the browser-local MiniLM worker after you choose to germinate them. Warp the Surface does not upload the text, embedding, genome, or seed. Shared seed links deliberately omit the source language, so a descendant can survive independently of the concept that produced its ancestor.
under the surface
The model does not grow the organism. It supplies ancestry.
compression
Twelve fixed deterministic projections collapse that vector into a tiny bounded genome. Similar meanings can inherit related tendencies without preserving the full embedding.
inheritance
Cross-pollination blends corresponding genes from two parents and adds small deterministic variation. Descendants can become new parents without another model call.
seed
The share token stores only quantized procedural traits and generation. Opening it reconstructs the organism without the original language or model.
system
Meaning enters once. The lineage keeps moving without it.
- input
- two short concepts / optional shared seed
- semantics
- all-MiniLM-L6-v2 / 384-dimensional normalized embedding
- projection
- 12 fixed bounded semantic-genome axes
- growth
- deterministic Canvas branching field
- inheritance
- gene-wise crossover + bounded mutation
- seed persistence
- URL fragment only when explicitly shared / no source text
- server inference
- none
- status
- experimental / mobile-first v0
This study intentionally reuses Warp's existing MiniLM runtime rather than introducing a larger generative model. The technical question is whether a learned latent representation can behave like hereditary material: compressed, recombined, mutated, and eventually detached from the language that first created it.