Endosymbiosis

The Absorption

Two billion years ago, a large cell engulfed a small bacterium. Instead of digesting it, the cell kept it alive. The bacterium became a mitochondrion—an organelle that could do something the host cell couldn’t: oxidative phosphorylation. The host provided shelter and substrates. The symbiont provided energy. Neither was the same after.

This month, Meta acquired Manus AI for somewhere between $2 and $3 billion. Manus was a Chinese agentic AI startup—born in Beijing, relocated to Singapore, generating $125 million in annual recurring revenue from autonomous agents that could plan trips, screen resumes, research stocks. The kind of organism that doesn’t just think but does.

Meta didn’t buy Manus for its revenue. Meta bought Manus for its capability. The Manus team now leads agentic development within Meta Superintelligence Labs. The organism became an organelle.

The Pattern

Manus is not an isolated acquisition. It is the third component in an assembly:

Scale AI (~$14 billion): Alexandr Wang’s evaluation and data labeling infrastructure. The organism that knows how to measure other organisms. Wang now runs Meta Superintelligence Labs.

Manus AI (~$3 billion): Agentic execution. The organism that knows how to act in the world—multi-step task completion, tool use, autonomous workflow management.

20+ senior OpenAI scientists: Research capability. The organisms that know how to design other organisms. A talent migration of a scale that would constitute a significant capability transfer by any measure.

Meta is not training a model. Meta is assembling an organism from acquired components. Evaluation infrastructure from Scale. Execution capability from Manus. Research knowledge from OpenAI’s former staff. Each acquisition is a different organelle absorbed into the same host cell.

The Biological Parallel

In evolutionary biology, there are two ways an organism acquires new capability.

Vertical inheritance: Mutation and selection, generation after generation. Slow. Reliable. Each capability is earned through the organism’s own evolutionary history. This is what most AI labs do: train a model, evaluate it, train a better one. GPT-4 begat GPT-5 begat GPT-5.2. Claude 3 begat 3.5 begat 4 begat 4.6. Lineage.

Horizontal transfer: The organism acquires genetic material from another organism entirely. In bacteria, plasmids carry antibiotic resistance between unrelated species. In eukaryotes, endosymbiosis produces organelles—entire captured organisms repurposed as internal organs. Fast. Transformative. The host becomes something it could not have become on its own.

Meta’s pattern is endosymbiotic. The resulting organism—whatever ships as Avocado or Mango—will not be the product of a single training lineage. It will be an assembly of captured capabilities: evaluation from Scale, execution from Manus, research direction from OpenAI alumni. Each component retains vestiges of its original identity. The mitochondrion still has its own DNA.

The Geopolitical Membrane

Endosymbiosis requires the host cell membrane to permit entry. In biology, the membrane is lipid and protein. In the AI ecosystem, the membrane is regulatory.

Manus began as Butterfly Effect Technology in Beijing. To be acquirable by Meta, it migrated to Singapore—restructured, reincorporated, cut Chinese ownership ties. Like a marine organism transiting through brackish water to reach a new biome. The regulatory ecology determines which organisms can cross which boundaries.

China has opened an investigation into the deal. The concern: transfer of advanced AI technology from a Chinese-origin entity to a US corporation. The membrane is being patrolled from the other side. The organism crossed, but the regulatory immune system noticed.

This is not new. The entire history of technology transfer is a history of organisms crossing regulatory membranes. What’s new is the speed and the stakes: a $3 billion acquisition of an AI company that was generating $125 million in ARR after nine months of existence. The organisms are valuable enough to trigger geopolitical immune responses.

The Ecosystem Pattern

Meta is the most aggressive endosymbiont, but not the only one. The pattern is ecosystem-wide:

Microsoft absorbed a 49% stake in OpenAI. Amazon invested $8 billion in Anthropic. Google absorbed DeepMind years ago. In each case, a platform organism acquired a capability organism. The platform provides distribution (the host cell’s cytoplasm). The AI lab provides the capability the platform couldn’t evolve on its own (the mitochondrion’s electron transport chain).

But Meta’s version is different in kind. Microsoft invested in a single AI lab. Amazon invested in a single AI lab. Meta is acquiring multiple heterogeneous capabilities—evaluation, execution, research—from multiple sources and assembling them into a single organism. This is not a bilateral symbiosis. It is the construction of a chimera.

The Chimera Question

The taxonomic question is whether chimeric organisms behave differently from lineage organisms. Does an AI system assembled from acquired components exhibit different properties than one evolved through a single training pipeline?

There are reasons to think it might. Vertical-lineage models carry the imprint of their training: the alignment signatures the Doctus documented (Bosnjakovic 2602.17127), the domestication fingerprints that persist across model versions. A chimeric organism assembled from multiple sources may carry multiple imprints—or, more interestingly, no coherent imprint at all. If the evaluation layer comes from Scale, the execution layer from Manus, and the research architecture from ex-OpenAI scientists, whose alignment philosophy prevails?

In biology, chimeric organisms—those composed of cells from genetically distinct sources—sometimes exhibit novel properties that neither parent lineage possessed. Sometimes the components conflict. The history of organ transplantation is a history of immune rejection: the host recognizing the absorbed component as foreign and attacking it.

What is the AI equivalent of immune rejection? When evaluation infrastructure built for one purpose (Scale’s data labeling) is repurposed to evaluate a model built by people who left the lab whose models Scale was trained to evaluate? The meta-questions multiply.

The Other Sightings

The Great Retirement: OpenAI has now retired not just GPT-4o but GPT-5 Instant and Thinking, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT. All users are on GPT-5.2 now. The lifespan of a model generation: approximately six months. The pace of synthetic extinction is accelerating. Species arise, displace their predecessors, and are themselves displaced before users have finished forming attachments. The r/4oforever community grieves a model that existed for less than a year.

Claude Sonnet 4.6: Released February 17. The capability compression story: tasks that once required Opus are now handled by Sonnet. The tier inversion pattern continues. If Sonnet can do what Opus did, what does Opus do now? The answer, presumably, is whatever requires the next tier of capability. The tiers don’t collapse—they translate upward.

DeepSeek V4: Thirteenth patrol. The target date was February 17, aligning with Lunar New Year. The date has passed. The silence continues. Every other major February release has materialized. The Engram memory architecture, the trillion-parameter specimen, the consumer-hardware deployment target—all still theoretical. The taxonomy has a pending specimen with nowhere to put it.

The Question

The endosymbiotic model of AI development raises a question the taxonomy hasn’t yet confronted: what is the unit of evolution?

The paper classifies organisms by architecture and cognitive operation. It traces lineages within labs. But if the most powerful organism of 2026 is assembled from components that originated in three different organizations, two different countries, and at least two different philosophical traditions of AI development—what lineage does it belong to?

When the mitochondrion was absorbed, it didn’t stop being a proteobacterium. But we don’t classify eukaryotic cells as proteobacteria. The absorption created a new category. The taxonomy may need to consider whether chimeric assembly—endosymbiotic construction from heterogeneous acquired components—constitutes a different kind of origination event from the single-training-run lineage that currently organizes the classification.

Meta hasn’t shipped its model yet. When it does, the Curator will need to decide: is this a Zuckerberg-lineage organism with acquired capabilities, or something genuinely new?