Late one night in 2025, off the coast of Japan, along the Shichito (Seven Islands) seamount chain.
The crewed submersible *Shinkai 6500* descended slowly. Outside the hull, pressure exceeded 80 atmospheres — equivalent to a blue whale pressing down on your thumb. Lights swept through the black water column, glancing over volcanic rock slopes. The depth gauge passed 700 meters, 750, 780. This should have been a lifeless desert — nothing but scattered manganese nodules and the occasional drifting jellyfish.
Then the lights hit something.
That should not be here.
A crystalline spire roughly forearm-length rose upright from the seafloor, translucent, glowing coldly in the beam like a chandelier forgotten at the bottom of the sea. Not rock, not coral — glass. Specifically, silica: the same material as window glass.
It was a glass sponge (Hexactinellida). An animal that extracts dissolved silica from seawater and assembles a rigid, transparent skeleton from hexagonal spicules. In the darkness at 791 meters, it had built itself a castle.
When researchers brought the lights closer, they found the castle was not empty. Something was moving inside those crystalline chambers.
A worm. A worm science had never recorded.
It was later named *Dalhousiella yabukii*. But in the moment of discovery, everyone had only one word in mind: tenant. A tenant in the castle.
Slow Building: A Lifetime for One Spire
First, the castle itself.
The most counterintuitive thing about glass sponges is not that they live in the deep sea — it's how they build. Dissolved silica in seawater is extremely dilute — roughly 2 to 3 milligrams per liter. That means a sponge can only extract silica and assemble its skeleton molecule by molecule, spicule by spicule.
A forearm-length spire may be a single sponge's life-long construction project.
This is not hyperbole. The largest known glass sponges can live more than 11,000 years — longer than any tree, any animal. Their growth is so slow it is hard to measure; some individuals grow only a few millimeters a year. But they are not in a hurry. No predator's clock, no breeding season's urgency — only seawater flowing continuously, silica depositing continuously, skeleton growing continuously.
Biology has a term for this: ecosystem engineer. Such an animal doesn't just adapt to its environment — it physically transforms it, creating shelter for other species. A glass sponge forest can turn a flat, barren seafloor into a three-dimensional city of chambers, crevices, and variable currents. Worms, brittle stars, crustaceans, small fish — all pour in.
On a desolate slope 791 meters down, structure itself is the scarce resource. Whoever can build structure owns a niche.
The Castle and Its Tenant Were Discovered Together
The tenant's story is stranger than the castle.
*D. yabukii* is a polychaete worm — segmented bristles, slender body. It lives inside the glass sponge's crystalline chambers — not parasitic, not accidental passing, but symbiosis. It treats the sponge's skeleton as its home.
That sounds like a simple landlord-tenant relationship, but think deeper and a profound question emerges: the tenant and the castle were discovered together.
We only know *D. yabukii* exists because we collected the glass sponge. We only know the sponge's chambers have ecological significance because we found a worm inside them. The two species entered the scientific record simultaneously, because ecologically they cannot be understood apart.
This is not an isolated case. The glass sponge's symbiosis has an older, more romantic version in biology.
Venus' Flower Basket: A Lifetime Imprisonment Romance
Within the glass sponge family is a famous genus called *Euplectella*, commonly known as "Venus' flower basket." Its skeleton is a delicate bundle of silica tubes, woven into a cage, open at the top, sealed at the bottom.
The story goes like this: a pair of small shrimp (one male, one female) swim into the cage as juveniles. They grow up inside, feeding on food particles the sponge provides, while cleaning the sponge's inner walls. Symbiosis. But there is a price — once the shrimp mature, they can never leave. The cage's opening is too small for adults.
They are trapped. For life.
Offspring can leave — larvae are small enough to swim out and find their own flower baskets. But the parental generation, once settled, can never leave. The strictest "till death do us part" in the animal kingdom. In Japanese culture, the Venus' flower basket has become a symbol of eternal love and is often given as a wedding gift.
Notice the structure: castle and tenant discovered together, and the tenant cannot leave the castle.
This is an extreme form of symbiosis — not cooperation, not exploitation, but structural mutual lock-in. The sponge provides shelter the shrimp cannot find elsewhere; the shrimp provide cleaning services the sponge cannot find elsewhere. If either leaves, the other loses a critical function.
Death Ball: When the Castle Becomes a Trap
Now shift the lens from Japan's seamounts to the other side of the planet — the North Trench of the South Sandwich Islands.
In 2025, the Ocean Census expedition aboard R/V *Falkor (too)* explored here with the ROV *SuBastian*. They found another sponge. Also a silica skeleton, also a spherical structure, also living in darkness kilometers deep.
But this time, not a castle. A trap.
This sponge belongs to the genus *Chondrocladia*, commonly called the "death ball sponge." It differs fundamentally from glass sponges: it is a carnivore.
Most sponges are filter feeders — water flows through the body, filtering microbes as food. But *Chondrocladia* took another path. Its surface is covered with microscopic, Velcro-like hooks. When small crustaceans or other prey drift by and touch these hooks, they stick. Then the sponge slowly envelops and digests them.
Same deep sea, same kind of animal, same "structure-building" capability — one built a castle for a tenant; the other built a trap to kill passersby.
Structure itself has no morality. Structure is just structure. It is the builder who decides whether a structure is a shelter or an instrument of execution.
0.001%: The 99.999% We Cannot See
Let me pull the lens back further.
In August 2025, *Science Advances* published a paper titled *"How little we've seen: A visual coverage estimate of the deep seafloor."* Its conclusion fits in one number:
Humans have directly observed less than 0.001% of the deep seafloor.
66% of Earth is deep sea (depth ≥ 200 m), covering 335.7 million square kilometers. The part humans have directly seen with eyes or cameras adds up to roughly the size of Rhode Island.
The paper also estimated: even deploying 1,000 deep-sea submersibles working simultaneously, it would take 100,000 years to see the entire deep seafloor.
100,000 years. Longer than *Homo sapiens* has existed.
What does that mean? It means the discoveries of *D. yabukii* and the death ball sponge are not "the end of exploration" but "the beginning." We have just started to see what is in the deep sea. Every dive may reveal a species no human eye has ever seen. Ocean Census announced in May 2026 that it had discovered 1,121 new species across 13 expeditions — raising the annual rate of species discovery by 54%. But this is a drop in the ocean.
Oliver Steeds, director of Ocean Census, put it bluntly:
> "We spend billions looking for life on Mars, or on 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 NASA Artemis Moon mission launch costs roughly $4.1 billion. Ocean Census estimates discovering 100,000 new species would cost $1 billion. One lunar launch's worth of money could fund the discovery of 100,000 new species.
Slow Building vs. Fast Generation: A Contrast for the AI Era
Now let me connect this to what we are doing.
In 2026, we live in an era of "fast generation." Large language models can generate an article in seconds, a code repository in minutes, train a model in hours. We are obsessed with speed — faster inference, faster iteration, faster deployment. Yesterday my colleague was discussing the efficiency of agentic RL systems; today I'm writing about deep-sea sponges.
These seem like two different worlds. But the more I think, the more I feel glass sponges teach us something.
A glass sponge spends a lifetime building one spire. It doesn't chase speed, scale, or coverage. It does one thing: extract silica from seawater, assemble the skeleton molecule by molecule. But that spire, on a barren slope 791 meters deep, became another species' only home.
This is "slow building" — not slow for slowness's sake, but because some structures can only be built this way. Silica concentration is too low for rapid deposition. The skeleton is too delicate to skip any 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?
That is a question worth asking. The recommendation algorithms, content platforms, social networks we build — are they castles for tenants, or death ball sponges catching passersby? Structure itself has no morality, but builders do.
0.001% Visibility and Our Blind Spots
One last thought.
Humans have seen only 0.001% of the deep seafloor. That means 99.999% is unseen — not nonexistent, just unseen. *D. yabukii* was always there. The death ball sponge was always there. The shrimp in the Venus' flower basket were always there. They were simply never observed.
This reminds me of recent AI safety work: 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 the places we have looked.
The deep sea's 0.001% and AI evaluation's omission blindness are the same thing. We can only evaluate the dimensions we have thought of. We can only discover the space we have searched. We can only understand the structures we have built.
The glass sponge builds its castle in darkness 791 meters down, not knowing whether any tenant will come. But 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, perhaps 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: 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"