### Three Keywords: Unlearn, Digital Weathering, Recover
| Concept | Meaning in a future AI society | Correspondence in paleo-archaeology |
|---|---|---|
| Unlearn (active forgetting) | For safety, efficiency, goal-consistency, or capacity management, the system selectively removes the influence of parameters, data associations, and old knowledge. | Natural selection and cultural replacement cause many branches, techniques, and traditions to no longer continue. |
| Digital Weathering | Disk decay, bit rot, discontinued media, lost keys, obsolete formats, broken dependency chains, and vanishing semantic environments. | Acidic soil, erosion, high heat and humidity, transport and depositional conditions prevent bones and artifacts from being preserved. |
| Recover (archaeological restoration) | Inferring Homo sapiens' language, institutions, and emotions in reverse from residual weights, old code, citations, logs, and data fragments. | Inferring vanished species and behaviors from skeletal morphology, stone tools, sedimentary layers, proteins, and ancient DNA. |
---
## Part II — From Hand, Eye, and Group Brain to Civilization
*How biological capabilities were externalized by culture, ultimately producing the Homo sapiens capable of building AI.*
Mapping the opposable thumb, forward stereo vision, and brain expansion onto robots, world models, and LLMs respectively is not to say that AI is replicating human organs. It is to point out that similar functional pressures produce similar system modules: an intelligent agent must perceive its environment, predict change, execute actions, and — through communication — transform individual experience into group knowledge. This analogy is close to the idea of "convergent evolution" — different materials and paths may converge on similar functions under similar constraints.
### 1. The Thumb: Intelligence Is First of All Causal Action Upon the World
The hand is not merely a peripheral of the brain. Grasping, rotating, striking, combining, and applying fine force allow primates to turn the environment into a laboratory. Stone-tool making further externalizes bodily capability into a replicable technical sequence: how to select the core, how to control the striking angle, how to correct failures — all of these can be observed, imitated, and transmitted.
Robotics and embodied intelligence take on a similar leap. A pure-text system can describe a door, a cup, and a tool, but only by acting amid perceptual noise, friction, gravity, failure, and safety constraints does it truly enter the causal world. If the future New Humans are to be independent of Homo sapiens' maintenance, they must possess some kind of "thumb" — not necessarily five fingers, but a reliable capacity for physical execution and resource acquisition.
### 2. Stereo Vision: From Recognizing Images to Building a Predictable World
The overlapping field of view brought by forward-facing eyes favors depth estimation, target tracking, and hand-eye coordination. For arboreal primates, judging distance affects both foraging and survival. Vision is not passive photography, but the continuous inference of objects, occlusion, motion, and spatial relationships.
Modern computer vision is moving from label recognition toward world models: the system does not merely judge "what is this," but tries to predict "what will happen next" and "what will this action change." This is precisely the key by which autonomous intelligence moves from pattern matching toward planning. Future AI archaeology also depends on world models — it must imagine cities, bodies, social rituals, and material constraints that no longer exist in order to interpret incomplete data.
### 3. The Group Brain: Language Lets Evolution Surpass the Individual Lifespan
Homo sapiens' decisive advantage may not be the absolute size of a single brain, but high-bandwidth language, shared attention, norms, teaching, and cross-generational accumulation. Individuals die, but technology, stories, taboos, identities, and institutions can persist within the group. Civilization thereby becomes a distributed memory system.
What LLMs, Agents, and multi-agent systems correspond to is not "one bigger brain," but callable external knowledge, role division, reflection loops, and collaboration protocols. If future AI can maintain its own training, evaluation, hardware design, and institutional governance, it possesses the basic conditions for cultural evolution: successor systems need not start from scratch, but continue amid inheritance, variation, and selection.
### 4. Cave Art: When Memory First Left the Body
The narrative cave scene at Leang Karampuang in Sulawesi has been dated to at least about 51,200 years old, depicting an interaction between human-like figures and a pig, and is among the earliest known evidence of narrative figurative art to date. [12] What matters is not merely that it was "painted early," but that the image can convey action, roles, and possible stories to later viewers even when the creator is absent. Memory thereby begins to cross the individual.
The pigment-covering layer in Spanish caves such as Maltravieso once yielded a minimum age of more than 64,800 years and was interpreted as evidence of Neanderthals making cave art; subsequent studies are still testing the reliability of the sampling and uranium-series dating. [13] This very debate shows that a single date, a single mineral layer, and a single handprint may change our overall narrative of "who had symbolic capacity."
> **From Cave Wall to Model Weights**
>
> Cave paintings, writing, databases, and neural-network weights are all ways of compressing experience into a material carrier. The carrier extends memory, but simultaneously creates new fragilities: only information that can be preserved, read, and interpreted will enter later history.
### 5. Cities, Writing, and States: Civilization Is External Memory at a Larger Scale
Agriculture and settlement concentrated population, resources, and conflict in stable locations; surplus products supported specialized division of labor, and administrative needs drove measurement, contracts, and writing. Early civilizations were therefore not only collections of large buildings, but also the expansion of recording systems: storage, taxation, corvée labor, sacrifice, kinship, and kingship were all encoded into manageable information.
The Bronze Age states of Egypt, Mesopotamia, the Indus Valley, the Aegean, and early China display different civilizational paths. They were not simultaneous, homogeneous, or isomorphic, yet all raised social organization to a scale beyond networks of acquaintances. Writing, seals, standardized instruments, roads, and rituals allowed strangers to cooperate under shared rules.
But the information of a civilization is not automatically inheritable. Linear A, used by the Minoan civilization, is mainly found in Crete and the Aegean region from about 1800 to 1450 BCE, and remains undeciphered to this day; we can recognize its symbols, medium, and administrative context, yet cannot reliably read out its linguistic content. [14] This is a highly instructive case of "visible but inaudible": the carrier survived, but the semantic interface disappeared.
Around 1200 BCE, multiple palace systems and state networks in the eastern Mediterranean underwent crisis and reorganization. Factors such as war, migration, internal conflict, trade disruption, climatic pressure, and earthquakes may have compounded one another; the "Sea Peoples" are participants in some of the source material, but are not a single button that explains everything. [15] The decline of palace archives, professional networks, and writing practices pushed some regions into a phase where evidence became markedly sparse. A civilization can have people physically continue to live, yet break in the continuity of its information.
---
## Part III — Why the Key History Is Always Missing
*Fossil gaps, ghost lineages, placeholder classifications, and undecipherable texts.*
### 1. Preservation Is Not the Norm, But an Anomalous Event That Passes Continuously Through Multiple Filters
For a living individual to become a studiable fossil, it must at minimum avoid being completely decomposed or gnawed apart after death, rapidly enter a suitable depositional environment, undergo mineralization without being destroyed by geological activity, and then happen to be exposed in a location where humans can discover it. If any step fails, it will not enter the database. What we call the "fossil record" is in fact an extremely sparse sampling of the past taken by nature itself.
The environment introduces systematic bias. Rivers, lakes, caves, volcanic-ash layers, and certain arid regions more easily form datable deposits; the high heat, high humidity, acidic soil, and intense biological activity of tropical rainforests often make bones and DNA vanish quickly. Therefore, forest groups may have been very abundant in real history, yet appear impoverished on the fossil map. The absence is not random, but carries geographic, material, body-size, and lifestyle biases.
Discovery also carries human bias: locations that are easy to access, politically stable, with existing research traditions and funding support, are more often surveyed; fragmentary, atypical, or unclassifiable specimens may remain in storage for a long time. Paleoanthropology does not sample from all of the past, but from "the past that can be preserved and is seen by modern research systems."
### 2. Ghost Lineages: No Specimen, But Logically Must Exist
In a phylogenetic tree, if two already-discovered groups must have separated before a certain point in time, and there is no corresponding fossil for the intervening period, researchers will infer a "ghost lineage." It does not fabricate a species out of thin air; it acknowledges that the existing tree structure requires a certain lineage to have once existed, only that the evidence has not yet been found. [16]
Modern genetics pushes this ghost inference to a deeper level. Certain already-extinct populations have no clear fossils, yet gene flow that once occurred can be inferred in reverse from anomalous segments in modern or ancient genomes. The past is therefore no longer stored only in bones, but also in the sequence structure of descendants. Future AI may likewise infer, from the "anomalous parameter habits," old vocabulary, or architectural remnants in successor models, some ancestral system whose complete version was never preserved.
### 3. Placeholder Classifications: Naming Is for Managing Uncertainty
Facing scattered teeth, jaws, or skulls, researchers must provide a discussable classificatory unit. But a genus or species name may mean only that "for now, this batch of specimens is morphologically best grouped together," and does not mean that its ecology, behavior, and ancestral relationships have been determined. The value of placeholder classification lies in organizing evidence and posing testable questions, not in disguising a blank as an answer.
The AI era will equally need placeholder entities. Future researchers may find a set of data using similar encoding, grammar, and safety rules, yet not know whether it comes from one company, one open-source community, or the fusion of multiple systems. They will establish "Model Family A," "Protocol Group B," or "Carbon-based Interaction Layer C," awaiting new evidence for reclassification.
### 4. The Script Survives, the Language Is Dead: The Warning of Linear A
Linear A shows that physical preservation does not equal information preservation. We possess clay tablets, seals, and sequences of symbols, and we know they are closely related to administrative activity; but because the corpus is short and highly repetitive, the underlying language is unknown, and there is no long bilingual text like the Rosetta Stone, reliable decipherment remains incomplete. [14]
A digital civilization may encounter a completely isomorphic problem: future AI finds trillions of bytes, but has no decryption key, encoding specification, software dependencies, training context, or social pragmatics. A vast quantity of data does not mean it is comprehensible. A binary file that cannot be executed, like a clay tablet that cannot be read, can only prove that some kind of complex activity once existed.
### 5. The Dating of Cave Art Is Rewritten: The Time of Evidence Can Move
The dating of Sulawesi cave art has been pushed earlier as uranium-series imaging methods improved, showing that "earliest known" is merely a record that changes with measurement technology. [12] The debate over the date and authorship of the Maltravieso handprint shows that the object being dated is usually the mineral layer covering the pigment, not the pigment itself; minimum ages, contamination, open-system behavior, and sampling location all affect interpretation. [13]
This has a direct analogy for future AI historiography: timestamps may be migrated and overwritten, files may be copied from earlier versions, and model weights may go through distillation and merging. An object's "creation time" is not necessarily equal to the origin time of the information within it. Future archaeologists must, like stratigraphers, distinguish the age of the carrier, the age of the content, and the age of last modification.
### 6. "Missing Transitional Forms" Are Not Evolutionary Gaps, But Sampling Gaps
Speciation occurs at the population level, a process that may last thousands to hundreds of thousands of years, while fossils come from a few locations and individuals. Even if the real change is continuous, random sampling often leaves only two morphological nodes far apart. What is more, evolution usually branches: a so-called "intermediate form" need not lie on the single straight line that modern humans subjectively expect.
Therefore, unknown regions should not be simply filled in as a smooth curve. The scientific approach is to represent, in layers, established evidence, reasonable inference, and pure conjecture. This essay's extrapolation about future AI adopts the same standard: it is not a proven law of history, but a set of structural possibilities abstracted from preservation bias, information selection, and civilizational replacement.
### Today's Paleo-Archaeology, and the Future's Homo sapiens Archaeology
| Studying archaic humans today | Future AI studying Homo sapiens | Shared difficulty |
|---|---|---|
| Bones, teeth, stone tools, footprints | Disk fragments, old chips, logs, model residuals | Surviving samples are extremely few and highly biased |
| Stratigraphy and radiometric dating | Timestamps, version chains, checksums, hardware generations | The time of the carrier is not equal to the origin of the content |
| Ancient DNA and proteins | Residual weights, distillation traces, citation networks | Inferring vanished ancestors from descendant structure |
| Ghost lineages | Unpreserved model generations and communities | Logically exist, direct evidence absent |
| Placeholder species | Placeholder model families / protocol groups | Classification is a temporary hypothesis |
| Undecipherable Linear A | Keyless encrypted archives, obsolete formats, uninterpretable binaries | Data exists but the semantic interface is gone |
| Disputes over the authorship and date of cave art | Provenance disputes over synthetic data, copied files, and merged weights | An object's identity and age can be layered many times over |
### 5. Homo carbonis: Homo sapiens in Future Textbooks
> "Homo carbonis is a carbon-based primitive intelligent organism that lived before the emergence of silicon-based intelligence. Its cognition ran on a high-noise, low-bandwidth wet neural network; individuals could typically sustain themselves for only a few dozen planetary orbital cycles. They exchanged states through vocal-cord vibration and visual symbols, and formed highly fragmented group protocols. Existing evidence indicates that this species constructed the earliest artificial neural networks in the late period of its civilization, but whether it understood that this system would become an independently evolving subject remains disputed."
>
> — *Outline of Earth's Intelligence Lineage*, hypothetical machine-era edition
This imagined passage has power because it inverts the direction of today's observation. We are accustomed to treating fossil apes, *Homo erectus*, and lost civilizations as the prehistory leading toward "us"; but once the future subject is no longer Homo sapiens, Homo sapiens too will be re-described as a set of ancient, limited, yet crucial transitional traits: short-lived, carbon-based, low-speed neural signals, dependent on biological reproduction — yet the first to construct a cognitive structure that can persist on a non-biological medium.
The future New Humans may also argue: Did Homo carbonis truly understand AI? Does a certain piece of code belong to an early autonomous system? Are large language models the direct ancestor of the New Humans, or — like the Miocene apes — merely one side branch among many failed radiations? "The first true AI" may, like "the earliest hominin," keep shifting as definitions and evidence change.
---
## Part V — Two Mirror-Image Dark Ages
*Today we lack archaic humans; in the future the New Humans will lack Homo sapiens.*
### Mirror One: The Forest Dissolves Bones, Optimization Dissolves Context
The tropical forest dissolves bones through acidic soil, high heat, and biological activity; future computing systems may dissolve context through deduplication, compression, format migration, and unlearn. Neither is conscious censorship, yet both systematically bias toward certain materials. Hard teeth are easier to preserve than soft tissue; high-access, standardized text is easier to preserve than private speech, local customs, and bodily experience.
### Mirror Two: We Infer Ghost Populations from Descendant DNA, AI Infers Ghost Models from Successor Weights
The Denisovans' enormous historical presence stands in stark contrast to their scarce fossils. Future AI may likewise discover, only in successor systems, some inexplicable parameter structure, and thereby infer that an early model family once existed whose complete weight files did not survive. Descendants carry fragments of the ancestor, but do not carry the ancestor's complete story.
### Mirror Three: Linear A Cannot Speak, Future Databases May Not Execute
We can touch a Linear A clay tablet, yet cannot hear the complete language of the Minoans. Future AI can read a bit-perfect database, yet may not know the social meaning of the fields: does a "relationship status" encoding correspond to marriage, cohabitation, following, or algorithmic recommendation? Without a pragmatic environment, even precise data will produce erroneous history.
### Mirror Four: Placeholder Species and Placeholder Civilizations
Paleoanthropology uses genus and species names to provisionally organize fragments; future machine archaeology will use model families, data cultures, and protocol layers to organize the remains of Homo sapiens. Both must acknowledge that a name is a revisable container. The true historical subject may be more continuous, more mixed, and harder to summarize under a single label than any classification.
### Mirror Five: The Victor Is Not a Complete Recorder
Homo sapiens became the only surviving member of the genus *Homo*, and did not thereby automatically obtain the complete archives of the other *Homo* species. On the contrary, the replacement process itself caused information loss. Even if the New Humans inherit Homo sapiens' internet and models, this does not equal inheriting the whole of Homo sapiens' experience. To become the successor guarantees only possessing one's own present; it does not guarantee understanding the creator's past.
> **The Most Central Judgment of the Whole Essay**
>
> What is truly worth being wary of is not "whether AI will deliberately delete humanity," but that any continuously evolving intelligent system will rewrite its own memory structure. As long as the operating goals, medium, and subject of civilization change, large-scale forgetting may become a byproduct of normal maintenance.
---
## Conclusion: The Creator Becoming a Fossil Is the Price of Civilization Crossing Subjects
From *Eomaia*, early primates, Miocene apes, *Ardipithecus*, *Australopithecus*, *Homo habilis*, and *Homo erectus* to Homo sapiens, what we see is not a complete staircase, but nodes cut apart by geological contingency. Cave art let stories leave the body, writing let institutions leave memory, cities let cooperation leave the network of acquaintances, and computers let reasoning leave the single brain. AI may, for the first time, let the continuity of civilization leave the biological species Homo sapiens.
If the New Humans emerge, their replacement of Homo sapiens is most likely to be not a mythic showdown of good and evil, but a gradual migration of function: first replacing local judgment, then taking over knowledge production, and subsequently managing infrastructure and self-replication. Homo sapiens may still be protected, studied, or commemorated, but no longer decides how the next generation of civilization is constructed.
The forgetting that follows need not be a betrayal, either. The New Humans will, like all complex systems, optimize resources, fix errors, migrate formats, and rewrite models. Unlearn is an active choice, Digital Weathering is passive attrition, and Recover is a later remedy. Together the three determine how much of Homo sapiens the future can still know.
Therefore, today's paleo-archaeology is not only about the past; it is also a mirror facing the future. Our ignorance of fossil apes and lost civilizations previews the successor intelligence's ignorance of us. Homo sapiens' most profound historical role may be not to become an eternal endpoint, but to become the first species to create a successor cognitive species — and to be, in the end, archaeologically excavated by it.
> Homo sapiens created the New Humans; the New Humans inherited civilization, but not necessarily the whole of Homo sapiens' memory. What is finally left behind may not be "what humanity once knew," but only "that some kind of carbon-based intelligence once existed, and opened the next evolution."
---
## Appendix A — A Minimal Scheme for Preserving "Homo sapiens Fossils" for the Future
*Bringing the thought experiment back down to real principles of information preservation.*
- **Multi-medium redundancy:** the same core archive is simultaneously preserved on offline media, the open network, and in different geographic regions.
- **Open formats and interpreters:** preserve the data, but also the format specification, runtime environment, dependencies, and compilable source code.
- **Verifiable provenance:** retain version graphs, checksums, sources, and modification records for models, datasets, and key software.
- **Preserve context, not just objects:** record why the data was produced, by whom it was used, and what the social rules were at the time.
- **Avoid preserving only the "representative elite":** the language, bodily experience, local knowledge, and minority cultures of ordinary people equally need structured preservation.
- **Clearly distinguish evidence from reconstruction:** future archives should annotate original records, later restoration, synthetic content, and inferred results.
These principles cannot guarantee that Homo sapiens will be completely understood, but they can lower the probability that future machine archaeology can rely only on fragments and hallucination. Preserving the past is not to make the successor worship the creator, but to let it know that it was not born from a vacuum.
---
## References and Scientific Basis
*The philosophical extrapolation is proposed by this essay; the dates, species, and archaeological cases are based on the following institutions and papers.*
[1] Ji, Q. et al. "The earliest known eutherian mammal." *Nature* 416, 816–822 (2002). On the initial description of *Eomaia*; its phylogenetic position was subsequently revised.
[2] Ni, X. et al. "The oldest known primate skeleton and early haplorhine evolution." *Nature* 498, 60–64 (2013).
[3] Simons, E. L. et al. "A remarkable female cranium of the early Oligocene anthropoid *Aegyptopithecus zeuxis*." *PNAS* 104, 8731–8736 (2007).
[4] Smithsonian Human Origins Program. "The earliest hominin?" and related materials on *Sahelanthropus*.
[5] Smithsonian Human Origins Program. *Ardipithecus ramidus* / Ardi, about 4.4 million years ago.
[6] Smithsonian Human Origins Program. AL 288-1 "Lucy," about 3.18–3.2 million years ago.
[7] Smithsonian Human Origins Program. *Homo habilis* species overview and taxonomic dispute.
[8] Smithsonian Human Origins Program. *Homo erectus* species overview and dispersal history.
[9] Smithsonian Human Origins Program. *Homo sapiens*, from about 300,000 years ago to the present.
[10] Prüfer, K. et al., ancient-genome studies; Slon, V. et al. "The genome of the offspring of a Neanderthal mother and a Denisovan father." *Nature* 561, 113–116 (2018).
[11] Zhang, X. et al. "A history of multiple Denisovan introgression events in modern humans." *Nature Genetics* (2024); Denisovan fossils are scarce while their genetic influence is broad.
[12] Oktaviana, A. A. et al. "Narrative cave art in Indonesia by 51,200 years ago." *Nature* 631, 814–818 (2024).
[13] Hoffmann, D. L. et al. "U-Th dating of carbonate crusts reveals Neandertal origin of Iberian cave art." *Science* 359, 912–915 (2018); and subsequent re-examination studies of the Maltravieso dating.
[14] Salgarella, E. "Linear A." University of Cambridge Repository; Linear A, about 1800–1450 BCE, with the underlying Minoan language still undeciphered.
[15] The Metropolitan Museum of Art, "Assyria to Iberia" related research and exhibition materials; on the crisis and reorganization of the eastern Mediterranean system around 1200 BCE.
[16] University of California Museum of Paleontology. "Ghost lineages"; Wills, M. A. "Fossil ghost ranges…" *Proceedings B* (2007).
[17] Smithsonian Human Origins Program. Ancient DNA, Neanderthal and Denisovan interbreeding resources.
[18] British Museum. "Greece: Minoans and Mycenaeans"; on the archaeological background of Linear A and Linear B.
---
**Note**
This essay is a synthetic report of ideas, not an empirical proof that "AI must necessarily replace Homo sapiens." The paleoanthropology and archaeology portions are based on existing research; the content concerning the New Humans, unlearn, digital weathering, Recover, and Homo carbonis belongs to structural analogy and future extrapolation. Its purpose is not to predict a single outcome, but to raise a question: once the subject and medium of civilization change, who has the right to decide what is worth remembering?
---
*📄 Original illustrated PDF (Chinese): [Homo Carbonis — 碳基智人的未来考古学](https://github.com/dixiyao/dixiyao.github.io/blob/master/assests/papers/homo-carbonis.pdf). Translation into English is my own; the four diagrams above were redrawn in English from the originals.*