On a night in 2025, off the coast of Japan along the Seven-Seas seamount chain, the crewed submersible Shinkai 6500 descended past 700, 750, 780 meters—more than 80 atmospheres outside the hull. The slope was supposed to be a biological desert.
Then the lights found something that shouldn't be there: a translucent crystalline spire, roughly forearm-length, standing upright on the volcanic seafloor like a chandelier forgotten in the deep.
It was a glass sponge (Hexactinellida)—an animal that extracts dissolved silica from seawater and assembles a rigid, transparent skeleton from six-pointed spicules. And when researchers brought the lights closer, the castle was not empty. Inside its crystalline chambers moved a worm never before recorded by science, later named *Dalhousiella yabukii*. Everyone on board thought the same word: tenant.
Slow Building: A Spire Takes a Lifetime
Seawater holds only about 2–3 mg of dissolved silica per liter, so a glass sponge must build its skeleton molecule by molecule, spicule by spicule. A single forearm-length spire may be a lifetime's work. The largest known glass sponges can live over 11,000 years—older than any tree or animal—growing just millimeters a year.
Biologists call such creatures ecosystem engineers: animals that physically reshape their environment and create shelter for other species. A glass sponge forest turns flat seafloor desert into a three-dimensional city of chambers and crevices, drawing in worms, brittle stars, crustaceans, and fish. At 791 meters, structure itself is the scarce resource—and whoever builds structure owns the niche.
The Castle and Its Tenant Were Discovered Together
*D. yabukii* is a polychaete worm living symbiotically—not parasitically—inside the sponge's crystalline chambers. Crucially, the tenant and the castle entered the scientific record at the same moment: we only know the sponge's chambers have ecological meaning because we found the worm inside them. The two species cannot be understood apart.
Venus' Flower Basket: A Lifetime Prisoner's Romance
In the genus *Euplectella*, known as the Venus' flower basket, a mating pair of shrimp enters the siliceous cage as juveniles. They grow up inside, cleaning the sponge's walls while feeding on its scraps—a symbiosis. But the cage's opening is too small for adults. Once in, the parents can never leave; only their offspring are small enough to swim out and find baskets of their own. It is the animal kingdom's strictest "till death do us part," and in Japanese culture the sponge is a symbol of eternal love, often given as a wedding gift.
Structure and tenant discovered together—and the tenant cannot leave. This is not cooperation or exploitation but structural mutual lock-in: each party provides what the other cannot get elsewhere.
Death Balls: When the Castle Becomes a Trap
On the other side of the planet, in the North Trench of the South Sandwich Islands, Ocean Census's R/V Falkor (too) expedition used ROV SuBastian in 2025 to find another ball-shaped, silica-skeletoned sponge in the deep dark. But this one—genus *Chondrocladia*, the "death ball sponge"—is a carnivore.
Its surface is covered in microscopic, velcro-like hooks. Small crustaceans that drift into contact are stuck, slowly enveloped, and digested. Same deep sea, same structure-building ability—one sponge built a castle for a tenant, the other built a trap for passersby.
Structure has no morality. Structure is just structure. It is the builder who decides whether a structure is a refuge or an instrument of death.
0.001%: The 99.999% We Cannot See
In August 2025, *Science Advances* published *"How little we've seen: A visual coverage estimate of the deep seafloor."* Its conclusion in one number:
> Humans have directly observed less than 0.001% of the deep seafloor.
The deep sea (depth ≥ 200 m) covers 66% of Earth—335.7 million square kilometers. The portion humans have seen with eyes or cameras is roughly the size of Rhode Island. The paper estimates that even with 1,000 deep-sea submersibles working simultaneously, full visual coverage would take 100,000 years—longer than *Homo sapiens* has existed.
So discoveries like *D. yabukii* and the death ball sponge are not the end of exploration but its beginning. Ocean Census announced in May 2026 that it had found 1,121 new species across 13 expeditions—a 54% boost to the annual rate of species discovery. Yet that is a drop in the ocean.
Ocean Census director Oliver Steeds put it bluntly:
> "We spend billions searching Mars for life, or the far side of the Moon. Discovering most of the life on our own planet—in our ocean—costs a fraction of that. The question is not whether we can afford to do it. The question is whether we can afford not to."
For reference: a single NASA Artemis lunar launch costs about $4.1 billion. Ocean Census estimates $1 billion would be enough to discover 100,000 new species.
Slow Building vs. Fast Generation: A Contrast for the AI Era
We live in an age of rapid generation. Large language models produce an article in seconds, a code repository in minutes, a trained model in hours. We are obsessed with speed—faster inference, faster iteration, faster deployment.
This seems like another world from the glass sponge, which spends a lifetime building a single spire—not for slowness's sake, but because some structures can only be built this way. Silica concentrations are too low for rapid deposition; the skeleton is too intricate to skip a single step. But once built, it is a castle. In the AI era we have built many things—but how many are castles? How many are traps? The recommendation algorithms, content platforms, and social networks we build: are they castles for tenants to live in, or death ball sponges catching passersby? Structure has no morality, but builders do.
0.001% Visibility and Our Blind Spots
A final thought. Humans have seen 0.001% of the deep seafloor, which means 99.999% is invisible—not nonexistent, merely unseen. *D. yabukii* was always there. The death ball sponge was always there. The shrimp in the Venus' flower basket were always there.
This mirrors a recurring problem in AI safety: LLM judges can detect commission (what was said) but not omission (what was left out)—because presence is a perception task while absence is a reasoning task. We can only see where we have looked.
The ocean's 0.001% and AI evaluation's omission blindness are the same phenomenon. We can only evaluate the dimensions we have thought of, discover in the space we have searched, understand the structures we have built.
The glass sponge builds its castle 791 meters down in the dark, not knowing whether a tenant will ever come. It builds anyway—because structure itself is the scarce resource, in the deep sea, in the AI era, in any desolate place.
Next time we discuss Agentic RL efficiency, maybe we should think of the glass sponge: a lifetime for one spire, but that spire became another species' only home.
Slow, but worth it.
---
References:
- Ocean Census May 2026 announcement: https://oceancensus.org/press-release-scientists-discover-over-1100-new-marine-species-in-landmark-ocean-census
- Discover Wildlife coverage: https://www.discoverwildlife.com/animal-facts/marine-animals/new-marine-species-discovered
- SiamDive on the glass castle worm: https://www.siamdive.com/en/blogs/worm-lives-inside-glass-sponge-791-metres
- Schmidt Ocean on the Venus' flower basket: https://schmidtocean.org/cruise-log-post/a-deep-sea-love-story
- Oceanographic on the death ball sponge: https://oceanographicmagazine.com/news/carnivorous-death-ball-sponge-among-new-deep-sea-species
- Bell et al. 2025, *Science Advances*: "How little we've seen: A visual coverage estimate of the deep seafloor"