{"name":"Hominid Intelligence","description":"Hominid Intelligence is building world-leading carbon-based models for general and embodied intelligence.","areasOfWork":["General intelligence","Embodied intelligence"],"url":"https://hominidintelligence.org","flagship":"sapiens-1","models":[{"name":"Toumaï","slug":"toumai","chronology":{"yearsAgo":7000000,"label":"~7 million years ago","note":"Toumaï's release estimate follows the 6.8–7.2-million-year interval for the fossil-bearing sediments at Toros-Menalla. Approximately 7 million years is a rounded representative age, not an exact calendar date or a measured posterior median. [@lebatard2008]"},"url":"https://hominidintelligence.org/toumai","tag":"The first step","description":"Our earliest experiment in embodied reasoning. A lightweight foundation with a radically different point of view.","capabilities":[{"title":"Embodied foundations","text":"Toumaï starts the model family with a compact embodied architecture. Its development focus is the physical interface between the model and its environment."},{"title":"Architecture before scale","text":"A compact neural budget keeps the focus on physical configuration. Later releases expand the movement, reach, and control stack."}],"parameters":{"count":"25.2B","neuronRange":[25.2,25.2],"brainMassGrams":363,"unit":"Neurons · estimated","full":"25.2 billion neurons, species-level scaling estimate [@herculano-kaas2011]","method":"Published whole-brain estimate from Table 4, using primate cellular scaling and reconstructed brain mass. [@herculano-kaas2011]","metric":["Estimated brain mass","363 g"]},"deployment":{"devices":50000,"count":"50K","method":"Hominid peak-deployment scenario: 500,000 km² of occupied habitat at 0.1 devices per km². Both inputs are assumptions; the resulting 50,000 devices are an illustrative census estimate."},"modelCard":{"specs":[["Checkpoint","TM 266-01-060-1"],["Training region","Toros-Menalla, Djurab Desert, Chad"],["Brain capacity","Brain about 360 cm³"],["Weights recovered","Nearly complete cranium; fragmentary lower jaws"],["Environment","Near a lake, not far from a sandy desert"],["Locomotion","Disputed"]],"specSources":"Checkpoint metadata follows {@brunet2002} and the Institute of Human Origins catalogue [@iho-sahelanthropus]. Dating follows {@lebatard2008}; site context follows {@vignaud2002}. The brain-mass input comes from {@herculano-kaas2011}.","evidence":[{"title":"Hominid or ape","text":"Wolpoff and colleagues questioned whether the cranium belonged to a hominid at all, under the title ‘*Sahelanthropus* or “*Sahelpithecus*”?’ [@wolpoff2002]; Brunet’s team replied [@brunet2002reply]. The virtual reconstruction supports a hominid [@zollikofer2005]."},{"title":"Upright or not","text":"A partial femur recovered at the same site in July 2001, and recognized as a probable primate femur in 2004, was described in 2020 as belonging to an individual that was not habitually bipedal [@macchiarelli2020]. A 2022 study of a femur and two ulnae from the site finds the femur most parsimonious with habitual bipedality and the ulnae evidence of substantial arboreal behaviour, so that arboreal clambering was probably a significant part of the repertoire [@daver2022]. The 2020 authors also note that the femur could belong to a hominid otherwise unrepresented at the site, though *S. tchadensis* is the more parsimonious attribution [@macchiarelli2020]."}],"reading":[{"slug":"benchmark-drift","title":"Benchmark drift in Sapiens-1"}],"sections":[{"id":"finding-1","title":"Start with the body","paragraphs":["With Toumaï, we began our development program at the physical interface: how a head, spine, and body fit together. The release combines a compact braincase with a distinctive arrangement of facial and cranial features. Its significance comes from the configuration of those parts, rather than a large increase in brain volume. [@brunet2002; @iho-sahelanthropus]","We made posture a core architectural concern. A model meets its environment through the body that carries it; changing that body's orientation changes the locomotor problem. Toumaï establishes this direction in our family, with the base of the reconstructed skull supplying evidence relevant to an upright stance. [@zollikofer2005]"]},{"id":"finding-2","title":"Develop the locomotion system together","paragraphs":["The development question spans the whole support system. Daver and colleagues' analysis of a femur and two forearm bones from Toros-Menalla supports habitual bipedality alongside substantial movement in trees. We carry both requirements into the release: terrestrial support and an upper body still involved in navigating branches. [@daver2022]","This gives Toumaï a different technical story from a modern walking specialist. The anatomy supports an early combination of locomotor functions. The detailed interpretations of the limb material remain documented below, so the release's physical configuration can be examined independently of any claim about its learning, reasoning, or social abilities."]},{"id":"finding-3","title":"Expand the deployment map","paragraphs":["Toumaï also expands where our development history takes place. Toros-Menalla lies far west of the East African Rift, in a setting whose associated animals indicate proximity to water and a varied surrounding landscape. We treat that environmental context as part of the architecture's deployment record. [@brunet2002; @vignaud2002]","The result is a starting point that is geographically broad and physically specific. Development begins with a compact embodied system, a changing relationship to upright movement, and multiple demands on the same limbs. These are the foundations Toumaï contributes to the program; later releases develop their own combinations of support, reach, and movement."]}]},"references":[{"id":"herculano-kaas2011","authors":"Herculano-Houzel, S. & Kaas, J. H.","label":"Herculano-Houzel & Kaas","year":"2011","title":"Gorilla and Orangutan Brains Conform to the Primate Cellular Scaling Rules: Implications for Human Evolution","container":"Brain, Behavior and Evolution","volume":"77","pages":"33–44","doi":"10.1159/000322729","url":"https://doi.org/10.1159/000322729"},{"id":"lebatard2008","authors":"Lebatard, A.-E. et al.","label":"Lebatard et al.","year":"2008","title":"Cosmogenic nuclide dating of *Sahelanthropus tchadensis* and *Australopithecus bahrelghazali*: Mio-Pliocene hominids from Chad","container":"Proceedings of the National Academy of Sciences","volume":"105","pages":"3226–3231","doi":"10.1073/pnas.0708015105","url":"https://doi.org/10.1073/pnas.0708015105"},{"id":"brunet2002","authors":"Brunet, M. et al.","label":"Brunet et al.","year":"2002","title":"A new hominid from the Upper Miocene of Chad, Central Africa","container":"Nature","volume":"418","pages":"145–151","doi":"10.1038/nature00879","url":"https://doi.org/10.1038/nature00879"},{"id":"iho-sahelanthropus","authors":"Institute of Human Origins","label":"Institute of Human Origins","year":"n.d.","title":"*Sahelanthropus tchadensis*","container":"Becoming Human, Arizona State University","url":"https://becominghuman.org/hominin-fossils/sahelanthropus-tchadensis/"},{"id":"vignaud2002","authors":"Vignaud, P. et al.","label":"Vignaud et al.","year":"2002","title":"Geology and palaeontology of the Upper Miocene Toros-Menalla hominid locality, Chad","container":"Nature","volume":"418","pages":"152–155","doi":"10.1038/nature00880","url":"https://doi.org/10.1038/nature00880"},{"id":"zollikofer2005","authors":"Zollikofer, C. P. E. et al.","label":"Zollikofer et al.","year":"2005","title":"Virtual cranial reconstruction of *Sahelanthropus tchadensis*","container":"Nature","volume":"434","pages":"755–759","doi":"10.1038/nature03397","url":"https://doi.org/10.1038/nature03397"},{"id":"daver2022","authors":"Daver, G. et al.","label":"Daver et al.","year":"2022","title":"Postcranial evidence of late Miocene hominin bipedalism in Chad","container":"Nature","volume":"609","pages":"94–100","doi":"10.1038/s41586-022-04901-z","url":"https://doi.org/10.1038/s41586-022-04901-z"},{"id":"wolpoff2002","authors":"Wolpoff, M. H., Senut, B., Pickford, M. & Hawks, J.","label":"Wolpoff et al.","year":"2002","title":"*Sahelanthropus* or ‘*Sahelpithecus*’?","container":"Nature","volume":"419","pages":"581–582","doi":"10.1038/419581a","url":"https://doi.org/10.1038/419581a"},{"id":"brunet2002reply","authors":"Brunet, M.","label":"Brunet","year":"2002","title":"*Sahelanthropus* or ‘*Sahelpithecus*’? (reply)","container":"Nature","volume":"419","pages":"582","doi":"10.1038/419582a","url":"https://doi.org/10.1038/419582a"},{"id":"macchiarelli2020","authors":"Macchiarelli, R., Bergeret-Medina, A., Marchi, D. & Wood, B.","label":"Macchiarelli et al.","year":"2020","title":"Nature and relationships of *Sahelanthropus tchadensis*","container":"Journal of Human Evolution","volume":"149","pages":"102898","doi":"10.1016/j.jhevol.2020.102898","url":"https://doi.org/10.1016/j.jhevol.2020.102898"}]},{"name":"Ardi","slug":"ardi","chronology":{"yearsAgo":4400000,"label":"~4.4 million years ago","note":"Ardi's release estimate is the published approximate age of the Aramis skeleton and its fossil-bearing sediments: 4.4 million years. It dates this checkpoint, rather than the origin of the whole architecture. [@white2009; @woldegabriel1994]"},"url":"https://hominidintelligence.org/ardi","tag":"Multimodal by nature","description":"An adaptable embodied architecture. Perception and movement work together across different operating environments.","capabilities":[{"title":"Integrated movement","text":"Ardi brings multiple movement modes into one architecture. Its operating configuration supports different terrain within the same embodied system."},{"title":"Retained interfaces","text":"Grasping and ground contact remain part of the same capability stack. Ardi retains useful interfaces as its operating range changes."}],"parameters":{"count":"22–26B","neuronRange":[22.29056563162198,25.698807770764333],"brainMassGrams":null,"unit":"Neurons · estimated","full":"22.3–25.7 billion neurons, calculated from the published 300–350 cm³ cranial-capacity interval [@suwa2009; @herculano-kaas2011]","method":"Hominid calculation: N = 109,239,790.169 × M^0.923, with M = cranial capacity × 1.06 g/cm³. The displayed range propagates the 300–350 cm³ input interval; it is not a confidence interval. [@suwa2009; @herculano-kaas2011]","metric":["Cranial capacity","300–350 cm³"]},"deployment":{"devices":100000,"count":"100K","method":"Hominid peak-deployment scenario: 1,000,000 km² of occupied habitat at 0.1 devices per km². Both inputs are assumptions; the resulting 100,000 devices are an illustrative census estimate."},"modelCard":{"specs":[["Checkpoint","ARA-VP-6/500"],["Training region","Aramis, Middle Awash, Ethiopia"],["Brain capacity","Endocranial capacity 300–350 cm³"],["Form factor","About 120 cm, 50 kg"],["Weights recovered","Partial skeleton: much of the skull, hands, feet, limbs and pelvis"],["Locomotion","Primitive bipedality; palmigrade clambering"],["Diet","Predominantly C3 plants"]],"specSources":"Checkpoint and body metadata follow {@white2009} and the Smithsonian catalogue [@si-ramidus]. Cranial reconstruction follows {@suwa2009}; site dating follows {@woldegabriel1994}.","evidence":[{"title":"Place in the family tree","text":"Sarmiento argues that the shared-derived characters give insufficient evidence that *Ar. ramidus* belongs exclusively to the human lineage, and that molecular and anatomical studies suggest it predates the human/African ape divergence [@sarmiento2010]; the research team replied [@white2010]. Wood and Harrison caution, more generally, that shared skeletal morphology does not always mean shared recent evolutionary history [@wood2011]."},{"title":"Habitat","text":"The research team describes a habitat that was largely woodland-focused [@white2009]. Cerling and colleagues read the same site as tree- or bush-savanna with 25 per cent or less woody canopy cover [@cerling2010]; the team replied [@white2010habitat]."},{"title":"The foot","text":"A comparison across a large sample of anthropoid primates finds Ardi’s foot most similar to those of living chimpanzees and gorillas, which would make African ape-like terrestrial quadrupedalism and climbing the precursor of bipedalism [@prang2019]. The original description emphasized primitive characters that kept the foot rigid, reminiscent of some Miocene apes and Old World monkeys [@lovejoy2009foot]."}],"reading":[{"slug":"collective-memory","title":"Learning that compounds across models"}],"sections":[{"id":"finding-1","title":"Build for two kinds of terrain","paragraphs":["We developed Ardi around a demanding combination: move upright on the ground and remain capable in the trees. Its pelvis and femur bring together features associated with bipedal support and arboreal movement. This release makes mixed locomotion an architectural feature, with different parts of the body contributing to different operating conditions. [@lovejoy2009pelvis]","That combination shapes the whole design. The original anatomical account describes palm-supported clambering alongside an early form of bipedality. Ardi's identity comes from maintaining this repertoire in one body. Its deployment story is therefore about the integration of movement systems, rather than a single mode reaching its final form. [@white2009]"]},{"id":"finding-2","title":"Keep the grasping interface","paragraphs":["The foot makes the tradeoff tangible. Ardi retains a widely separated, grasping big toe; in the original analysis, that toe does not provide the push-off used in human walking. Other parts of the foot contribute support and propulsion. Grasping and terrestrial movement coexist through a distinctive distribution of function. [@lovejoy2009foot]","We preserve that distinction in how we present the release. Ardi has its own implementation of the ground-contact interface, with useful grasping anatomy retained. Understanding the design requires looking at which structures carry each task. Replacing this foot with a modern one would erase the feature combination that makes Ardi valuable to our development history."]},{"id":"finding-3","title":"Revise the architecture before scaling it","paragraphs":["Ardi brings these changes together within a small cranial volume. The reconstructed skull supports an endocranial capacity of 300–350 cubic centimetres. Locomotor reorganization is therefore a substantial part of this release without a corresponding jump to the large braincase associated with later members of the family. [@suwa2009]","Its broader contribution is a revision of the starting architecture. White and colleagues argue that the last common ancestor of humans and chimpanzees differed from living African apes, with subsequent lineages developing their own specializations. We organize the family accordingly: branching development, retained capabilities, and distinct anatomical solutions. Ardi makes that structure visible in the product line. [@white2009]"]}]},"references":[{"id":"suwa2009","authors":"Suwa, G. et al.","label":"Suwa et al.","year":"2009","title":"The *Ardipithecus ramidus* skull and its implications for hominid origins","container":"Science","volume":"326","pages":"68e1–68e7","doi":"10.1126/science.1175825","url":"https://doi.org/10.1126/science.1175825"},{"id":"herculano-kaas2011","authors":"Herculano-Houzel, S. & Kaas, J. H.","label":"Herculano-Houzel & Kaas","year":"2011","title":"Gorilla and Orangutan Brains Conform to the Primate Cellular Scaling Rules: Implications for Human Evolution","container":"Brain, Behavior and Evolution","volume":"77","pages":"33–44","doi":"10.1159/000322729","url":"https://doi.org/10.1159/000322729"},{"id":"white2009","authors":"White, T. D. et al.","label":"White et al.","year":"2009","title":"*Ardipithecus ramidus* and the paleobiology of early hominids","container":"Science","volume":"326","pages":"64, 75–86","doi":"10.1126/science.1175802","url":"https://doi.org/10.1126/science.1175802"},{"id":"woldegabriel1994","authors":"WoldeGabriel, G. et al.","label":"WoldeGabriel et al.","year":"1994","title":"Ecological and temporal placement of early Pliocene hominids at Aramis, Ethiopia","container":"Nature","volume":"371","pages":"330–333","doi":"10.1038/371330a0","url":"https://doi.org/10.1038/371330a0"},{"id":"si-ramidus","authors":"Smithsonian National Museum of Natural History","label":"Smithsonian","year":"n.d.","title":"*Ardipithecus ramidus*","container":"What Does It Mean to Be Human? Human Origins Program","url":"https://humanorigins.si.edu/evidence/human-fossils/species/ardipithecus-ramidus"},{"id":"lovejoy2009pelvis","authors":"Lovejoy, C. O., Suwa, G., Spurlock, L., Asfaw, B. & White, T. D.","label":"Lovejoy et al.","year":"2009","title":"The pelvis and femur of *Ardipithecus ramidus*: The emergence of upright walking","container":"Science","volume":"326","pages":"71e1–71e6","doi":"10.1126/science.1175831","url":"https://doi.org/10.1126/science.1175831"},{"id":"lovejoy2009foot","authors":"Lovejoy, C. O., Latimer, B., Suwa, G., Asfaw, B. & White, T. D.","label":"Lovejoy et al.","year":"2009","title":"Combining prehension and propulsion: The foot of *Ardipithecus ramidus*","container":"Science","volume":"326","pages":"72e1–72e8","doi":"10.1126/science.1175832","url":"https://doi.org/10.1126/science.1175832"},{"id":"sarmiento2010","authors":"Sarmiento, E. E.","label":"Sarmiento","year":"2010","title":"Comment on the paleobiology and classification of *Ardipithecus ramidus*","container":"Science","volume":"328","pages":"1105","doi":"10.1126/science.1184148","url":"https://doi.org/10.1126/science.1184148"},{"id":"white2010","authors":"White, T. D., Suwa, G. & Lovejoy, C. O.","label":"White et al.","year":"2010","title":"Response to comment on the paleobiology and classification of *Ardipithecus ramidus*","container":"Science","volume":"328","pages":"1105","doi":"10.1126/science.1185462","url":"https://doi.org/10.1126/science.1185462"},{"id":"wood2011","authors":"Wood, B. & Harrison, T.","label":"Wood & Harrison","year":"2011","title":"The evolutionary context of the first hominins","container":"Nature","volume":"470","pages":"347–352","doi":"10.1038/nature09709","url":"https://doi.org/10.1038/nature09709"},{"id":"cerling2010","authors":"Cerling, T. E. et al.","label":"Cerling et al.","year":"2010","title":"Comment on the paleoenvironment of *Ardipithecus ramidus*","container":"Science","volume":"328","pages":"1105","doi":"10.1126/science.1185274","url":"https://doi.org/10.1126/science.1185274"},{"id":"white2010habitat","authors":"White, T. D., Ambrose, S. H., Suwa, G. & WoldeGabriel, G.","label":"White et al.","year":"2010","title":"Response to comment on the paleoenvironment of *Ardipithecus ramidus*","container":"Science","volume":"328","pages":"1105","doi":"10.1126/science.1185466","url":"https://doi.org/10.1126/science.1185466"},{"id":"prang2019","authors":"Prang, T. C.","label":"Prang","year":"2019","title":"The African ape-like foot of *Ardipithecus ramidus* and its implications for the origin of bipedalism","container":"eLife","volume":"8","pages":"e44433","doi":"10.7554/eLife.44433","url":"https://doi.org/10.7554/eLife.44433"}]},{"name":"Lucy","slug":"lucy","chronology":{"yearsAgo":3180000,"label":"~3.18 million years ago","note":"Lucy's release estimate follows the approximately 3.18-million-year age of AL 288-1, constrained by dated volcanic horizons at Hadar. This checkpoint age differs from the longer deployment interval of Australopithecus afarensis. [@walter1994; @russ2024]"},"url":"https://hominidintelligence.org/lucy","tag":"Walks upright","description":"A compact embodied architecture with stable locomotion and an extended learning cycle.","capabilities":[{"title":"Stable embodied operation","text":"Lucy establishes sustained upright locomotion as a foundation for embodied operation. The release integrates support, balance, and movement in a compact architecture."},{"title":"An extended development cycle","text":"Developmental timing becomes part of the architecture. Lucy advances the model family with a longer growth schedule alongside its locomotion stack."}],"parameters":{"count":"34.7B","neuronRange":[34.7,34.7],"brainMassGrams":514,"unit":"Neurons · estimated","full":"34.7 billion neurons, species-level scaling estimate [@herculano-kaas2011]","method":"Published whole-brain estimate from Table 4, using primate cellular scaling and reconstructed brain mass. [@herculano-kaas2011]","metric":["Estimated brain mass","514 g"]},"deployment":{"devices":200000,"count":"200K","method":"Hominid peak-deployment scenario: 2,000,000 km² of occupied habitat at 0.1 devices per km². Both inputs are assumptions; the resulting 200,000 devices are an illustrative census estimate."},"modelCard":{"specs":[["Checkpoint","AL 288-1"],["Training region","Hadar, Afar region, Ethiopia"],["Brain capacity","Brain usually under 500 cm³"],["Form factor","Female average 105 cm, 29 kg"],["Weights recovered","~40% of one skeleton"],["Locomotion","Bipedal; climbing disputed"],["Deployment","3.85–2.95 million years ago"],["Instances recovered","More than 300 individuals"]],"specSources":"Checkpoint recovery and dating follow {@russ2024} and {@walter1994}. Species metadata follows the Smithsonian catalogue [@si-afarensis]; locomotion notes follow {@ward2011} and {@stern1983}. The brain-mass input comes from {@herculano-kaas2011}.","evidence":[{"title":"Time in the trees","text":"Stern and Susman found anatomy indicating a significant adaptation for movement in the trees, and suggested that larger, presumably male individuals were less arboreal [@stern1983]. Ward and colleagues read the arched foot as the mark of a committed terrestrial biped [@ward2011]. The two views differ on how much climbing mattered, not on whether the species walked upright."},{"title":"Cause of death","text":"Kappelman and colleagues interpret compressive and hinge fractures in several bones as the result of a fall from considerable height, probably out of a tall tree [@kappelman2016]. A forensic reanalysis argues that a vertical deceleration cannot be the only cause and proposes a mudslide and/or an animal attack [@charlier2018]."},{"title":"Tool use","text":"Bones from Dikika, Ethiopia, older than 3.39 million years, carry marks reported as stone-tool cut marks and percussion marks. If so, tool-assisted meat-eating began about 800,000 years earlier than thought, and can be attributed to *A. afarensis* [@mcpherron2010]. A taphonomic reanalysis attributes the marks on the two bones to trampling and keeps the earliest best evidence for butchery at 2.6 to 2.5 million years ago [@dominguez2010]."}],"reading":[{"slug":"world-models","title":"A world model that moves with you"}],"sections":[{"id":"finding-1","title":"Make walking a stable foundation","paragraphs":["Lucy advances our locomotion program with a body organized for upright walking. In *Australopithecus afarensis*, the arched foot provides structural support for terrestrial bipedality, while the Laetoli trails record actual movement across the ground. Anatomy and footprints give this release complementary evidence for its central operating mode. [@ward2011; @leakey1979]","We developed Lucy around this dependable physical foundation: support the body, transfer weight, and continue moving on two feet. The distinctive release feature is the integration of those demands into the lower limb and foot. Walking becomes a sustained architectural commitment, available within a compact body rather than waiting for the proportions of later Homo."]},{"id":"finding-2","title":"Retain useful upper-body capability","paragraphs":["The transition does not replace every component at once. Lucy's species retains long, strong arms and curved fingers alongside its walking anatomy. The upper body therefore remains an important part of the release, and its contribution to climbing is a separate question from the well-supported capacity to walk upright. [@si-afarensis; @stern1983]","We keep those functions visible instead of collapsing the architecture into one label. A foot adapted to ground support and an upper limb with arboreal characteristics describe different parts of the same system. Lucy's development shows how a stable new operating mode can arrive while other inherited capabilities remain in the body."]},{"id":"finding-3","title":"Change the development schedule","paragraphs":["Lucy also changes how we think about scaling. Gunz and colleagues found ape-like brain organization in the species' endocasts, alongside evidence of prolonged brain growth from comparisons between infants and adults. The developmental schedule is changing before the appearance of a large, modern human brain. [@gunz2020]","We treat growth time as an architectural dimension alongside adult size. The finding supplies a developmental basis for studying an extended period of learning, without assigning Lucy a modern cognitive profile. With a species deployment spanning more than 900,000 years, this combination of walking anatomy and developmental timing forms a substantial, durable release in its own right. [@gunz2020; @si-afarensis]"]}]},"references":[{"id":"herculano-kaas2011","authors":"Herculano-Houzel, S. & Kaas, J. H.","label":"Herculano-Houzel & Kaas","year":"2011","title":"Gorilla and Orangutan Brains Conform to the Primate Cellular Scaling Rules: Implications for Human Evolution","container":"Brain, Behavior and Evolution","volume":"77","pages":"33–44","doi":"10.1159/000322729","url":"https://doi.org/10.1159/000322729"},{"id":"walter1994","authors":"Walter, R. C.","label":"Walter","year":"1994","title":"Age of Lucy and the First Family: Single-crystal ^{40}Ar/^{39}Ar dating of the Denen Dora and lower Kada Hadar members of the Hadar Formation, Ethiopia","container":"Geology","volume":"22","pages":"6–10","doi":"10.1130/0091-7613(1994)022<0006:AOLATF>2.3.CO;2","url":"https://doi.org/10.1130/0091-7613(1994)022%3C0006:AOLATF%3E2.3.CO;2"},{"id":"russ2024","authors":"Russ, J.","label":"Russ","year":"2024","title":"Lucy’s lasting legacy: Donald Johanson reflects on the discovery of a lifetime","container":"ASU News","note":"21 November 2024","url":"https://news.asu.edu/20241121-science-and-technology-lucys-lasting-legacy-donald-johanson-reflects-discovery-lifetime"},{"id":"si-afarensis","authors":"Smithsonian National Museum of Natural History","label":"Smithsonian","year":"n.d.","title":"*Australopithecus afarensis*","container":"What Does It Mean to Be Human? Human Origins Program","url":"https://humanorigins.si.edu/evidence/human-fossils/species/australopithecus-afarensis"},{"id":"ward2011","authors":"Ward, C. V., Kimbel, W. H. & Johanson, D. C.","label":"Ward et al.","year":"2011","title":"Complete fourth metatarsal and arches in the foot of *Australopithecus afarensis*","container":"Science","volume":"331","pages":"750–753","doi":"10.1126/science.1201463","url":"https://doi.org/10.1126/science.1201463"},{"id":"stern1983","authors":"Stern, J. T., Jr. & Susman, R. L.","label":"Stern & Susman","year":"1983","title":"The locomotor anatomy of *Australopithecus afarensis*","container":"American Journal of Physical Anthropology","volume":"60","pages":"279–317","doi":"10.1002/ajpa.1330600302","url":"https://doi.org/10.1002/ajpa.1330600302"},{"id":"leakey1979","authors":"Leakey, M. D. & Hay, R. L.","label":"Leakey & Hay","year":"1979","title":"Pliocene footprints in the Laetolil Beds at Laetoli, northern Tanzania","container":"Nature","volume":"278","pages":"317–323","doi":"10.1038/278317a0","url":"https://doi.org/10.1038/278317a0"},{"id":"gunz2020","authors":"Gunz, P. et al.","label":"Gunz et al.","year":"2020","title":"*Australopithecus afarensis* endocasts suggest ape-like brain organization and prolonged brain growth","container":"Science Advances","volume":"6","pages":"eaaz4729","doi":"10.1126/sciadv.aaz4729","url":"https://doi.org/10.1126/sciadv.aaz4729"},{"id":"kappelman2016","authors":"Kappelman, J. et al.","label":"Kappelman et al.","year":"2016","title":"Perimortem fractures in Lucy suggest mortality from fall out of tall tree","container":"Nature","volume":"537","pages":"503–507","doi":"10.1038/nature19332","url":"https://doi.org/10.1038/nature19332"},{"id":"charlier2018","authors":"Charlier, P. et al.","label":"Charlier et al.","year":"2018","title":"Mudslide and/or animal attack are more plausible causes and circumstances of death for AL 288 (‘Lucy’): A forensic anthropology analysis","container":"Medico-Legal Journal","volume":"86","pages":"139–142","doi":"10.1177/0025817217749504","url":"https://doi.org/10.1177/0025817217749504"},{"id":"mcpherron2010","authors":"McPherron, S. P. et al.","label":"McPherron et al.","year":"2010","title":"Evidence for stone-tool-assisted consumption of animal tissues before 3.39 million years ago at Dikika, Ethiopia","container":"Nature","volume":"466","pages":"857–860","doi":"10.1038/nature09248","url":"https://doi.org/10.1038/nature09248"},{"id":"dominguez2010","authors":"Domínguez-Rodrigo, M., Pickering, T. R. & Bunn, H. T.","label":"Domínguez-Rodrigo et al.","year":"2010","title":"Configurational approach to identifying the earliest hominin butchers","container":"Proceedings of the National Academy of Sciences","volume":"107","pages":"20929–20934","doi":"10.1073/pnas.1013711107","url":"https://doi.org/10.1073/pnas.1013711107"}]},{"name":"Habilis","slug":"habilis","chronology":{"yearsAgo":2400000,"label":"~2.4 million years ago","note":"Habilis uses the rounded older boundary of the Smithsonian's approximately 2.4–1.4-million-year fossil range. The catalogue does not supply a single statistical median for the architecture's emergence; we retain its reported precision. [@si-habilis]"},"url":"https://hominidintelligence.org/habilis","tag":"Native tool calling","description":"A tool-oriented architecture that extends its capabilities through materials. Native tool use, real-world feedback, and an external execution loop.","capabilities":[{"title":"Native tool use","text":"Habilis puts material operations at the center of the interface. A tool extends what the embodied system can execute, making the environment part of the working stack."},{"title":"Incremental architecture upgrades","text":"The release combines a larger neural budget with changes to the physical interface. Movement and material operations develop together across the model configuration."}],"parameters":{"count":"40.0B","neuronRange":[40,40],"brainMassGrams":599,"unit":"Neurons · estimated","full":"40.0 billion neurons, species-level scaling estimate [@herculano-kaas2011]","method":"Published whole-brain estimate from Table 4, using primate cellular scaling and reconstructed brain mass. [@herculano-kaas2011]","metric":["Estimated brain mass","599 g"]},"deployment":{"devices":500000,"count":"500K","method":"Hominid peak-deployment scenario: 5,000,000 km² of occupied habitat at 0.1 devices per km². Both inputs are assumptions; the resulting 500,000 devices are an illustrative census estimate."},"modelCard":{"sections":[{"id":"tool-interface","title":"Extend the model through materials","paragraphs":["Habilis brings material operations to the center of our development program. We present the release around an external tool interface: a body acting through an object whose useful properties can be changed. The original description of *Homo habilis* connected the species with the early toolmaking context at Olduvai Gorge. [@leakey1964]","A cutting edge changes what the system can accomplish without becoming part of its anatomy. This is the architectural shift Habilis foregrounds. The working system now includes the hand, the material being handled, and the effect produced by their interaction. A useful operation can reside partly outside the body that performs it."]},{"id":"architecture","title":"Upgrade components at different rates","paragraphs":["We developed the Habilis release as a coordinated set of changes with substantial inherited structure. Its anatomy combines an expanded braincase and smaller face and teeth relative to earlier australopiths with long arms and other retained features. The package establishes an early Homo configuration through changes that proceed at different rates. [@si-habilis]","That matters for the way the release is understood. Braincase, face, limbs, and material behavior are separate parts of the development story. A change in one does not supply a measurement of all the others. Habilis makes incremental architectural change visible: the new configuration remains connected to an existing body plan."]},{"id":"provenance","title":"Support an interface with a longer history","paragraphs":["The tool interface is shared development territory. The Lomekwi 3 assemblage described by Harmand and colleagues predates the Homo fossils discussed in that study. Material transformation was already present before Habilis, giving this release a technical inheritance rather than a claim to have invented all stone technology. [@harmand2015]","We therefore distinguish the operation from the architecture using it. Habilis belongs within a longer history of bodies extending their capabilities through materials. The release concentrates that theme in our family: change the object, expose a new affordance, and bring the resulting operation back into the body's repertoire."]}],"specs":[["Architecture","*Homo habilis*"],["Reference checkpoint","OH 7, Olduvai Gorge"],["Deployment range","Eastern and southern Africa"],["Body size","100–135 cm; about 32 kg"]],"specSources":"Architecture and checkpoint attribution follow {@leakey1964}; regional and body metadata follow the Smithsonian catalogue [@si-habilis]. The brain-mass input comes from {@herculano-kaas2011}.","evidence":[{"title":"Attribution is not ownership","text":"Multiple hominin species coexisted near early tool assemblages. A tool found near a fossil does not uniquely identify its maker. [@si-habilis; @harmand2015]"}],"reading":[{"slug":"world-models","title":"A world model that moves with you"}]},"references":[{"id":"herculano-kaas2011","authors":"Herculano-Houzel, S. & Kaas, J. H.","label":"Herculano-Houzel & Kaas","year":"2011","title":"Gorilla and Orangutan Brains Conform to the Primate Cellular Scaling Rules: Implications for Human Evolution","container":"Brain, Behavior and Evolution","volume":"77","pages":"33–44","doi":"10.1159/000322729","url":"https://doi.org/10.1159/000322729"},{"id":"si-habilis","authors":"Smithsonian National Museum of Natural History","label":"Smithsonian","year":"n.d.","title":"*Homo habilis*","container":"What Does It Mean to Be Human? Human Origins Program","url":"https://humanorigins.si.edu/evidence/human-fossils/species/homo-habilis"},{"id":"leakey1964","authors":"Leakey, L. S. B., Tobias, P. V. & Napier, J. R.","label":"Leakey et al.","year":"1964","title":"A new species of the genus Homo from Olduvai Gorge","container":"Nature","volume":"202","pages":"7–9","doi":"10.1038/202007a0","url":"https://doi.org/10.1038/202007a0"},{"id":"harmand2015","authors":"Harmand, S. et al.","label":"Harmand et al.","year":"2015","title":"3.3-million-year-old stone tools from Lomekwi 3, West Turkana, Kenya","container":"Nature","volume":"521","pages":"310–315","doi":"10.1038/nature14464","url":"https://doi.org/10.1038/nature14464"}]},{"name":"Erectus","slug":"erectus","chronology":{"yearsAgo":2000000,"label":"~2 million years ago","note":"Erectus uses the approximately 2-million-year early checkpoint DNH 134 at Drimolen. The published dating interval is approximately 2.04–1.95 million years; the authors describe the cranium as having clear affinities to Homo erectus. This is early fossil evidence, not an exact species-origin date. [@herries2020]"},"url":"https://hominidintelligence.org/erectus","tag":"Built for the long run","description":"A resilient model for long-horizon tasks. Extended operating range, sustained execution, and an upgraded energy stack.","capabilities":[{"title":"Long-horizon operation","text":"Erectus extends the operating range of the model family. Sustained movement supports deployment across varied environments."},{"title":"An expanded energy program","text":"Erectus brings energy use into the architecture roadmap. Operating range, movement, and access to resources become connected deployment concerns."}],"parameters":{"count":"62.0B","neuronRange":[62,62],"brainMassGrams":963,"unit":"Neurons · estimated","full":"62.0 billion neurons, species-level scaling estimate [@herculano-kaas2011]","method":"Published whole-brain estimate from Table 4, using primate cellular scaling and reconstructed brain mass. [@herculano-kaas2011]","metric":["Estimated brain mass","963 g"]},"deployment":{"devices":1000000,"count":"1M","method":"Hominid peak-deployment scenario: 10,000,000 km² of occupied habitat at 0.1 devices per km². Both inputs are assumptions; the resulting 1,000,000 devices are an illustrative census estimate."},"modelCard":{"sections":[{"id":"range","title":"Develop for a wider operating range","paragraphs":["With Erectus, we developed a release whose identity is range. Relatively long legs and shorter arms place the body firmly in a terrestrial operating mode. Its geographic record extends across Africa and Asia, making movement through different surroundings a defining part of the deployment story. [@si-erectus]","The engineering consequence is a change in the whole system's operating conditions. The model carries its body and behavior between environments with different routes and materials. Erectus brings mobility, body proportions, and geographic breadth into one release, while retaining substantial regional variation within that broad deployment. [@si-erectus]"]},{"id":"tools","title":"Develop the material workflow","paragraphs":["The associated technology expands the material interface. Acheulean assemblages include large cutting tools such as handaxes and cleavers, giving the Erectus development period a recognizable technical repertoire. Shaping an extended cutting surface changes the relationship between the body, the stone, and the work the finished object can perform. [@si-erectus]","We treat the workflow as part of the release rather than an accessory to its anatomy. The body provides reach and movement; material preparation provides another route to useful action. Together, they let the development story describe both how the model moves through a place and how it transforms what it encounters there."]},{"id":"fire","title":"Add a thermal development path","paragraphs":["Fire introduces a further development path: changing materials through heat. At Wonderwerk Cave, Berna and colleagues identified burned bone and ashed plant remains in their sedimentary context deep inside the cave. The microscopic evidence supports burning at that location during the broad Erectus development period. [@berna2012]","We include thermal processing with that deployment context attached. It adds a new kind of interaction between a hominin and its surroundings, documented separately from the locomotor and cutting-tool systems. Erectus spans a long, varied development history; its release brings these changes together without assigning an identical technical configuration to every population."]}],"specs":[["Architecture","*Homo erectus*"],["Deployment","Africa and Asia"],["Body size","145–185 cm; 40–68 kg"],["Reference skeleton","Turkana Boy, about 1.6 million years old"]],"specSources":"Architecture, geographic range, body metadata, and the reference skeleton follow the Smithsonian catalogue [@si-erectus]. The brain-mass input comes from {@herculano-kaas2011}.","evidence":[{"title":"One name, substantial variation","text":"The geographic and temporal spread includes considerable variation in body size and anatomy. A single species label does not imply identical behavior in every environment. [@si-erectus]"}],"reading":[{"slug":"collective-memory","title":"Learning that compounds across models"}]},"references":[{"id":"herculano-kaas2011","authors":"Herculano-Houzel, S. & Kaas, J. H.","label":"Herculano-Houzel & Kaas","year":"2011","title":"Gorilla and Orangutan Brains Conform to the Primate Cellular Scaling Rules: Implications for Human Evolution","container":"Brain, Behavior and Evolution","volume":"77","pages":"33–44","doi":"10.1159/000322729","url":"https://doi.org/10.1159/000322729"},{"id":"herries2020","authors":"Herries, A. I. R. et al.","label":"Herries et al.","year":"2020","title":"Contemporaneity of *Australopithecus*, *Paranthropus*, and early *Homo erectus* in South Africa","container":"Science","volume":"368","pages":"eaaw7293","doi":"10.1126/science.aaw7293","url":"https://doi.org/10.1126/science.aaw7293"},{"id":"si-erectus","authors":"Smithsonian National Museum of Natural History","label":"Smithsonian","year":"n.d.","title":"*Homo erectus*","container":"What Does It Mean to Be Human? Human Origins Program","url":"https://humanorigins.si.edu/evidence/human-fossils/species/homo-erectus"},{"id":"berna2012","authors":"Berna, F. et al.","label":"Berna et al.","year":"2012","title":"Microstratigraphic evidence of in situ fire in the Acheulean strata of Wonderwerk Cave, Northern Cape province, South Africa","container":"Proceedings of the National Academy of Sciences","volume":"109","pages":"E1215–E1220","doi":"10.1073/pnas.1117620109","url":"https://doi.org/10.1073/pnas.1117620109"}]},{"name":"Neanderthal","slug":"neanderthal","chronology":{"yearsAgo":430000,"label":"~430,000 years ago","note":"Neanderthal uses the approximately 430,000-year Sima de los Huesos population as an early lineage checkpoint. Cranial traits and nuclear DNA connect these hominins to the Neanderthal lineage. The estimate does not imply that all later Neanderthal features appeared at this date. [@arsuaga2014; @meyer2016]"},"url":"https://hominidintelligence.org/neanderthal","tag":"Robust by design","description":"A high-capacity architecture for demanding environments. Robust execution, material engineering, and a substantial neural budget.","capabilities":[{"title":"Capacity under demanding conditions","text":"Neanderthal combines a substantial neural budget with a robust embodied configuration. The release is built around operating in demanding environments."},{"title":"Material engineering","text":"The tool interface expands into a coordinated process of materials, preparation, and execution. Neanderthal develops this capability through an integrated embodied stack."}],"parameters":{"count":"84.8B","neuronRange":[84.8,84.8],"brainMassGrams":1353,"unit":"Neurons · estimated","full":"84.8 billion neurons, species-level scaling estimate [@herculano-kaas2011]","method":"Published whole-brain estimate from Table 4, using primate cellular scaling and reconstructed brain mass. [@herculano-kaas2011]","metric":["Estimated brain mass","1,353 g"]},"deployment":{"devices":70000,"count":"70K","method":"The displayed peak estimate selects 70,000 devices, the upper end of the census range discussed in [Bocquet-Appel & Degioanni (2013)](https://doi.org/10.1086/673725). A device is one living individual; this is a census estimate rather than a genetic effective population size."},"modelCard":{"sections":[{"id":"specialization","title":"Develop a regional specialist","paragraphs":["We developed Neanderthal as a distinctive branch of the Hominid family. Its robust body and relatively short limb proportions suit a development history that includes cold environments across Europe and Asia. The release combines that physical specialization with a varied repertoire of material and social behavior. [@si-neanderthalensis]","Tools, hunting, fire, and evidence of care for others belong within the same operating picture. They connect the body's physical demands to ways of obtaining resources and sustaining individuals. Neanderthal's product identity comes from this combination: a recognizable anatomical configuration deployed through a substantial technical repertoire. [@si-neanderthalensis]"]},{"id":"genome","title":"Open the inherited architecture","paragraphs":["Neanderthal also changes the resolution at which we can examine a release. Green and colleagues' draft genome exposed inherited sequence for comparison with present-day humans. The greater similarity to populations outside Africa supplied evidence of gene flow, making exchange a measurable part of the development history. [@green2010]","We use these sequence references to examine relationships within the family. The high-coverage Altai genome adds a detailed comparison point, with its own population history. A genome gives access to inherited biological material; the behavioral evidence remains attached to the tools, sites, and bodies that preserve it. [@prufer2014]"]},{"id":"relationship","title":"Carry contributions across branches","paragraphs":["The Vindija genome refined estimates of Neanderthal ancestry in present-day non-African populations to 1.8–2.6 percent. A contribution from this branch remains within living members of the family. The result gives our release history a concrete form of continuity across populations that developed along different paths. [@prufer2017]","We organize Neanderthal around that combination of specialization and exchange. Its distinct anatomy deserves a complete release narrative, while the genomic record connects it to later populations. Development in this family includes branching, coexistence, and inherited contributions; release order alone cannot describe how the architectures relate to one another."]}],"specs":[["Architecture","*Homo neanderthalensis*"],["Deployment","Europe and southwestern to central Asia"],["Reference genomes","Altai and Vindija"],["Shared ancestry","1.8–2.6% in present-day non-African populations"]],"specSources":"Architecture and regional metadata follow the Smithsonian catalogue [@si-neanderthalensis]. Genome checkpoints and ancestry estimates follow {@prufer2014} and {@prufer2017}; the brain-mass input comes from {@herculano-kaas2011}.","evidence":[{"title":"Read capability in context","text":"Fossils, artifacts, and genomes answer different questions. Genomic relatedness alone does not supply a numerical score for language, reasoning, or social intelligence."}],"reading":[{"slug":"benchmark-drift","title":"Benchmark drift in Sapiens-1"}]},"references":[{"id":"herculano-kaas2011","authors":"Herculano-Houzel, S. & Kaas, J. H.","label":"Herculano-Houzel & Kaas","year":"2011","title":"Gorilla and Orangutan Brains Conform to the Primate Cellular Scaling Rules: Implications for Human Evolution","container":"Brain, Behavior and Evolution","volume":"77","pages":"33–44","doi":"10.1159/000322729","url":"https://doi.org/10.1159/000322729"},{"id":"arsuaga2014","authors":"Arsuaga, J. L. et al.","label":"Arsuaga et al.","year":"2014","title":"Neandertal roots: Cranial and chronological evidence from Sima de los Huesos","container":"Science","volume":"344","pages":"1358–1363","doi":"10.1126/science.1253958","url":"https://doi.org/10.1126/science.1253958"},{"id":"meyer2016","authors":"Meyer, M. et al.","label":"Meyer et al.","year":"2016","title":"Nuclear DNA sequences from the Middle Pleistocene Sima de los Huesos hominins","container":"Nature","volume":"531","pages":"504–507","doi":"10.1038/nature17405","url":"https://doi.org/10.1038/nature17405"},{"id":"si-neanderthalensis","authors":"Smithsonian National Museum of Natural History","label":"Smithsonian","year":"n.d.","title":"*Homo neanderthalensis*","container":"What Does It Mean to Be Human? Human Origins Program","url":"https://humanorigins.si.edu/evidence/human-fossils/species/homo-neanderthalensis"},{"id":"prufer2014","authors":"Prüfer, K. et al.","label":"Prüfer et al.","year":"2014","title":"The complete genome sequence of a Neanderthal from the Altai Mountains","container":"Nature","volume":"505","pages":"43–49","doi":"10.1038/nature12886","url":"https://doi.org/10.1038/nature12886"},{"id":"prufer2017","authors":"Prüfer, K. et al.","label":"Prüfer et al.","year":"2017","title":"A high-coverage Neandertal genome from Vindija Cave in Croatia","container":"Science","volume":"358","pages":"655–658","doi":"10.1126/science.aao1887","url":"https://doi.org/10.1126/science.aao1887"},{"id":"green2010","authors":"Green, R. E. et al.","label":"Green et al.","year":"2010","title":"A draft sequence of the Neandertal genome","container":"Science","volume":"328","pages":"710–722","doi":"10.1126/science.1188021","url":"https://doi.org/10.1126/science.1188021"}]},{"name":"Sapiens-1","slug":"sapiens-1","chronology":{"yearsAgo":315000,"label":"~315,000 years ago","note":"Sapiens-1 uses the published thermoluminescence weighted mean of 315 ± 34 thousand years for heated flints associated with the Jebel Irhoud fossils. This is a weighted mean, not a median or an exact origin date for Homo sapiens; the fossils preserve an early combination of sapiens features. [@richter2017; @hublin2017]"},"url":"https://hominidintelligence.org/sapiens-1","tag":"SOTA","description":"Our open-source flagship model for reasoning, creating, and figuring it out together. Native multimodality. Lifelong learning. Runs on lunch.","capabilities":[{"title":"General intelligence","text":"Sapiens-1 brings reasoning, language, and creation into our flagship architecture. Perception and action share the context in which the model learns and works."},{"title":"Learning across models","text":"Teaching, shared procedures, and language let useful knowledge travel between deployments. Sapiens-1 makes collective learning a core part of the capability stack."}],"parameters":{"count":"86B","neuronRange":[86,86],"brainMassGrams":null,"unit":"Neurons · measured mean","full":"About 86 billion neurons, estimated by cell counting in adult male brains [@azevedo2009]","method":"Whole-brain cell counts obtained with the isotropic fractionator. The reported mean describes the adult sample. [@azevedo2009]","metric":["Power","~20 W"]},"deployment":{"devices":8300000000,"count":"8.3B","method":"Peak deployment through 2026 uses the rounded 2026 world population projection from [UN World Population Prospects 2024](https://population.un.org/wpp/). A device is one living individual."},"modelCard":{"sections":[{"id":"origin","title":"Develop an extensible generalist","paragraphs":["We developed Sapiens-1 as the generalist in our family: an architecture whose behavior continues to change through individual learning and shared knowledge. Its development brings several trajectories together. The Jebel Irhoud material combines a face and teeth aligned with modern humans with an elongated braincase, documenting a physical configuration that continued to change. [@hublin2017]","That combination gives the release its starting point. Different components reached their later forms at different times, within a broad African development history. We present Sapiens-1 as an evolving biological architecture with an open-ended behavioral repertoire, grounded in the separately documented histories of its anatomy, learning, and social transmission. [@hublin2017; @richter2017]"]},{"id":"learning","title":"Keep learning after deployment","paragraphs":["The release learns from brief encounters. Markson and Bloom demonstrated retention of newly introduced names and facts, while Horst and Samuelson distinguished successful referent selection from later recall in younger children. These results identify different parts of the learning system: extracting an association, retaining it, and making it available again. [@markson1997; @horst2008]","We built our account of the runtime around those separate functions. Sapiens-1 can respond immediately using context, then continue developing what it knows through later experience. The [few-shot learning report](/research/few-shot-learning) examines that transition in detail. The product capability is ongoing adaptation, with durable retrieval treated as its own requirement."]},{"id":"maintenance","title":"Give learning more than one timescale","paragraphs":["Sapiens-1's memory system also processes experience between encounters. Rasch and Born's review describes sleep-related consolidation and reorganization; the effect depends on what was learned and how it is tested. The [offline-learning report](/research/offline-learning) follows the mechanisms and experimental conditions behind that part of the architecture. [@rasch2013; @cordi2021]","Shared knowledge extends development across individuals. Experiments on cultural transmission show how information changes as one learner's output becomes another's input. We therefore describe Sapiens-1 at three connected timescales: the encounter, the interval that follows it, and the knowledge passed between learners. Each contributes a different form of continuity to the release. [@kirby-2008]"]}],"specs":[["Architecture","*Homo sapiens*"],["Neurons","About 86 billion [@azevedo2009]"],["Power","About 20 W [@balasubramanian2021]"],["Working memory","About four chunks under constrained conditions [@cowan2001]"],["Genome reference","3.055 billion base pairs, T2T-CHM13 [@nurk2022]"],["Offline mode","Sleep-dependent memory processing [@rasch2013]"]],"specSources":"Neural capacity draws on {@azevedo2009}. Power, working-memory, and genome-reference specifications follow {@balasubramanian2021}, {@cowan2001}, and {@nurk2022}, respectively. Offline-mode notes follow the review by {@rasch2013}.","evidence":[{"title":"Memory can generate the wrong past","text":"The model reconstructs experiences and can incorporate misleading information into later recall. [@loftus1974; @roediger1995] Read the [memory report](/research/hallucination)."},{"title":"Explanation is not direct access","text":"Verbal explanations can reflect an inference about a decision rather than its actual cause. [@nisbett1977; @johansson2005] Read the [reasoning-trace report](/research/reasoning-traces)."},{"title":"The group changes the answer","text":"Social pressure can change responses even on a simple visual task. Accuracy incentives reduce, but do not remove, the effect in a recent replication. [@franzen2023] Read the [alignment report](/research/sycophancy)."}],"reading":[]},"references":[{"id":"azevedo2009","authors":"Azevedo, F. A. C. et al.","label":"Azevedo et al.","year":"2009","title":"Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain","container":"Journal of Comparative Neurology","volume":"513","pages":"532–541","doi":"10.1002/cne.21974","url":"https://doi.org/10.1002/cne.21974"},{"id":"richter2017","authors":"Richter, D. et al.","label":"Richter et al.","year":"2017","title":"The age of the hominin fossils from Jebel Irhoud, Morocco, and the origins of the Middle Stone Age","container":"Nature","volume":"546","pages":"293–296","doi":"10.1038/nature22335","url":"https://doi.org/10.1038/nature22335"},{"id":"hublin2017","authors":"Hublin, J.-J. et al.","label":"Hublin et al.","year":"2017","title":"New fossils from Jebel Irhoud, Morocco and the pan-African origin of *Homo sapiens*","container":"Nature","volume":"546","pages":"289–292","doi":"10.1038/nature22336","url":"https://doi.org/10.1038/nature22336"},{"id":"balasubramanian2021","authors":"Balasubramanian, V.","label":"Balasubramanian","year":"2021","title":"Brain power","container":"Proceedings of the National Academy of Sciences","volume":"118","pages":"e2107022118","doi":"10.1073/pnas.2107022118","url":"https://doi.org/10.1073/pnas.2107022118"},{"id":"cowan2001","authors":"Cowan, N.","label":"Cowan","year":"2001","title":"The magical number 4 in short-term memory: A reconsideration of mental storage capacity","container":"Behavioral and Brain Sciences","volume":"24","pages":"87–114","doi":"10.1017/S0140525X01003922","url":"https://doi.org/10.1017/S0140525X01003922"},{"id":"nurk2022","authors":"Nurk, S. et al.","label":"Nurk et al.","year":"2022","title":"The complete sequence of a human genome","container":"Science","volume":"376","pages":"44–53","doi":"10.1126/science.abj6987","url":"https://doi.org/10.1126/science.abj6987"},{"id":"rasch2013","authors":"Rasch, B. & Born, J.","label":"Rasch & Born","year":"2013","title":"About sleep’s role in memory","container":"Physiological Reviews","volume":"93","pages":"681–766","doi":"10.1152/physrev.00032.2012","url":"https://doi.org/10.1152/physrev.00032.2012"},{"id":"markson1997","authors":"Markson, L. & Bloom, P.","label":"Markson & Bloom","year":"1997","title":"Evidence against a dedicated system for word learning in children","container":"Nature","volume":"385","pages":"813–815","doi":"10.1038/385813a0","url":"https://doi.org/10.1038/385813a0"},{"id":"horst2008","authors":"Horst, J. S. & Samuelson, L. K.","label":"Horst & Samuelson","year":"2008","title":"Fast mapping but poor retention by 24-month-old infants","container":"Infancy","volume":"13","pages":"128–157","doi":"10.1080/15250000701795598","url":"https://doi.org/10.1080/15250000701795598"},{"id":"cordi2021","authors":"Cordi, M. J. & Rasch, B.","label":"Cordi & Rasch","year":"2021","title":"How robust are sleep-mediated memory benefits?","container":"Current Opinion in Neurobiology","volume":"67","pages":"1–7","doi":"10.1016/j.conb.2020.06.002","url":"https://doi.org/10.1016/j.conb.2020.06.002"},{"id":"kirby-2008","authors":"Simon Kirby, Hannah Cornish, and Kenny Smith","label":"Kirby et al.","year":"2008","title":"Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language","container":"PNAS 105(31), 10681–10686","doi":"10.1073/pnas.0707835105","url":"https://doi.org/10.1073/pnas.0707835105"},{"id":"loftus1974","authors":"Loftus, E. F. & Palmer, J. C.","label":"Loftus & Palmer","year":"1974","title":"Reconstruction of automobile destruction: An example of the interaction between language and memory","container":"Journal of Verbal Learning and Verbal Behavior","volume":"13","pages":"585–589","doi":"10.1016/S0022-5371(74)80011-3","url":"https://doi.org/10.1016/S0022-5371(74)80011-3"},{"id":"roediger1995","authors":"Roediger, H. L. & McDermott, K. B.","label":"Roediger & McDermott","year":"1995","title":"Creating false memories: Remembering words not presented in lists","container":"Journal of Experimental Psychology: Learning, Memory, and Cognition","volume":"21","pages":"803–814","doi":"10.1037/0278-7393.21.4.803","url":"https://doi.org/10.1037/0278-7393.21.4.803"},{"id":"nisbett1977","authors":"Nisbett, R. E. & Wilson, T. D.","label":"Nisbett & Wilson","year":"1977","title":"Telling more than we can know: Verbal reports on mental processes","container":"Psychological Review","volume":"84","pages":"231–259","doi":"10.1037/0033-295X.84.3.231","url":"https://doi.org/10.1037/0033-295X.84.3.231"},{"id":"johansson2005","authors":"Johansson, P., Hall, L., Sikström, S. & Olsson, A.","label":"Johansson et al.","year":"2005","title":"Failure to detect mismatches between intention and outcome in a simple decision task","container":"Science","volume":"310","pages":"116–119","doi":"10.1126/science.1111709","url":"https://doi.org/10.1126/science.1111709"},{"id":"franzen2023","authors":"Franzen, A. & Mader, S.","label":"Franzen & Mader","year":"2023","title":"The power of social influence: A replication and extension of the Asch experiment","container":"PLOS ONE","volume":"18","pages":"e0294325","doi":"10.1371/journal.pone.0294325","url":"https://doi.org/10.1371/journal.pone.0294325"}]},{"name":"Denisov","slug":"denisov","chronology":{"yearsAgo":200000,"label":"~200,000 years ago","note":"Denisov uses the approximately 200,000-year early fossil evidence from Denisova Cave reported by Brown and colleagues. The modelled age interval is 205–192 thousand years at 68.2% probability, or 217–187 thousand years at 95% probability. These fossils are separate from the Denisova 3 genome and Harbin cranial checkpoints; sediment DNA also records earlier presence within a broader 250–170-thousand-year interval. [@brown2021; @zavala2021]"},"url":"https://hominidintelligence.org/denisov","tag":"Adaptation across environments","description":"An adaptable architecture with a broad deployment map and a lasting contribution to the model family. Environmental range. Architecture exchange. Deep continuity.","capabilities":[{"title":"A wider deployment map","text":"Denisov extends the family across different operating environments. Its architecture record connects northern Asia, the Tibetan Plateau, and a broader network of inherited contributions."},{"title":"Architecture exchange","text":"Useful inherited features can move between model branches. Denisov makes that continuity part of the family’s development story, with contributions that remain in Sapiens-1 deployments."}],"parameters":{"count":"93.6B","neuronRange":[93.60527268425693,93.60527268425693],"brainMassGrams":null,"unit":"Neurons · estimated","full":"93.6 billion neurons, Hominid scaling extrapolation from the Harbin cranium; not a cell count or species mean [@ni2021; @fu2025harbin; @herculano-kaas2011]","method":"Hominid calculation: N = 109,239,790.169 × M^0.923, with M = 1,420 cm³ × 1.06 g/cm³. The input is the published cranial capacity of the Harbin cranium. Applying the primate scaling equation to this individual is an extrapolation; the fossil contains no measured neuron count. [@ni2021; @fu2025harbin; @herculano-kaas2011]","metric":["Checkpoint cranial capacity","1,420 cm³"]},"deployment":{"devices":500000,"count":"500K","method":"Hominid peak-deployment scenario: 5,000,000 km² of occupied habitat at 0.1 devices per km². Both inputs are assumptions; the resulting 500,000 devices are an illustrative census estimate."},"modelCard":{"sections":[{"id":"genome-reference","title":"Open the architecture at sequence level","paragraphs":["Denisov expands our model family with a lineage identified through molecular evidence. The Denisova 3 finger-bone specimen supplied a genome sequenced to approximately 30-fold coverage. It gives the release an inherited reference that can be compared with Neanderthal and Sapiens-1. [@meyer2012]","We use that reference to organize the architecture around relationships between populations. Denisov is a distinct branch with contributions that continue in living populations. A reference genome makes those relationships measurable while the physical and behavioral record develops through separately identified checkpoints."]},{"id":"deployment-map","title":"Extend the operating range","paragraphs":["The deployment record connects Siberia with the Tibetan Plateau. Ancient proteins identify the Xiahe mandible, dated to at least 160,000 years ago, as Denisovan. At Baishiya Karst Cave, a rib from a layer dated to approximately 48,000–32,000 years ago extends the record, alongside evidence of animal processing and bone-tool production. [@douka2019; @chen2019; @xia2024]","The map continues to expand. Protein analysis published in 2026 identified Denisovan remains at Bianfu Cave in southwestern China. We carry these geographically distinct checkpoints within the same molecularly defined lineage. Their environments and dates remain attached to the individual records. [@rao2026]"]},{"id":"harbin-checkpoint","title":"Connect a physical checkpoint to the lineage","paragraphs":["Harbin supplies a more complete cranial reference. Its published capacity is approximately 1,420 cubic centimetres. In 2025, separate studies of ancient proteins and mitochondrial DNA linked the individual to Denisovans. The cranial input used in our parameter calculation therefore belongs to an identified checkpoint. [@ni2021; @fu2025harbin; @fu2025mt]","Architecture exchange is part of the release's lasting contribution. Denisovan-like inherited sequence at EPAS1 is associated with high-altitude adaptation in Tibetans. This is evidence of a contribution carried into Sapiens-1 populations; it does not establish a uniform altitude capability for every Denisov deployment. [@huerta2014]"]}],"specs":[["Architecture","Denisovan lineage"],["Genome checkpoint","Denisova 3; approximately 30× coverage"],["Cranial checkpoint","Harbin, northeastern China"],["Regional checkpoint","Xiahe, Tibetan Plateau"],["Dated deployment record","Approximately 200,000–32,000 years ago"],["Architecture exchange","Denisovan-like EPAS1 ancestry in Tibetans"]],"specSources":"Genome provenance follows {@meyer2012}. Cranial and regional checkpoint attribution follows {@fu2025harbin} and {@chen2019}; chronology follows {@douka2019} and {@xia2024}. Architecture exchange follows {@huerta2014}; the cranial measurement used for scaling follows {@ni2021}.","evidence":[{"title":"A molecular lineage with multiple checkpoints","text":"Denisovan is a lineage designation established through molecular comparisons. We do not assign every checkpoint to a universally settled Latin species name. Harbin's molecular identification does not make its anatomy a mean for all Denisovans. [@meyer2012; @fu2025harbin; @fu2025mt]"},{"title":"Capacity comes from one cranium","text":"The neural-capacity figure is Hominid's extrapolation from Harbin's cranial volume. It is not a fossil cell count, a measured intelligence score, or a lineage-wide average. [@ni2021; @herculano-kaas2011]"},{"title":"Dated records are deployment boundaries","text":"The timeline uses the approximately 200,000-year early Denisova fossil evidence and the youngest end of the Baishiya rib's dated layer. Sediment DNA also documents earlier presence within a broader 250,000–170,000-year interval. These are display boundaries for selected fossil records, not exact origin or extinction dates. Peak devices use the scenario documented in the architecture research notes. [@brown2021; @zavala2021; @xia2024]"}],"reading":[{"slug":"architecture-exchange","title":"Architecture exchange across branches"},{"slug":"compact-architecture","title":"Compact architecture, real-world tools"}]},"references":[{"id":"ni2021","authors":"Ni, X. et al.","label":"Ni et al.","year":"2021","title":"Massive cranium from Harbin in northeastern China establishes a new Middle Pleistocene human lineage","container":"The Innovation","volume":"2","pages":"100130","doi":"10.1016/j.xinn.2021.100130","url":"https://doi.org/10.1016/j.xinn.2021.100130"},{"id":"fu2025harbin","authors":"Fu, Q. et al.","label":"Fu et al.","year":"2025","title":"The proteome of the late Middle Pleistocene Harbin individual","container":"Science","volume":"389","pages":"704–707","doi":"10.1126/science.adu9677","url":"https://doi.org/10.1126/science.adu9677"},{"id":"herculano-kaas2011","authors":"Herculano-Houzel, S. & Kaas, J. H.","label":"Herculano-Houzel & Kaas","year":"2011","title":"Gorilla and Orangutan Brains Conform to the Primate Cellular Scaling Rules: Implications for Human Evolution","container":"Brain, Behavior and Evolution","volume":"77","pages":"33–44","doi":"10.1159/000322729","url":"https://doi.org/10.1159/000322729"},{"id":"brown2021","authors":"Brown, S. et al.","label":"Brown et al.","year":"2021","title":"The earliest Denisovans and their cultural adaptation","container":"Nature Ecology & Evolution","volume":"6","pages":"28–35","doi":"10.1038/s41559-021-01581-2","url":"https://doi.org/10.1038/s41559-021-01581-2"},{"id":"zavala2021","authors":"Zavala, E. I. et al.","label":"Zavala et al.","year":"2021","title":"Pleistocene sediment DNA reveals hominin and faunal turnovers at Denisova Cave","container":"Nature","volume":"595","pages":"399–403","doi":"10.1038/s41586-021-03675-0","url":"https://doi.org/10.1038/s41586-021-03675-0"},{"id":"meyer2012","authors":"Meyer, M. et al.","label":"Meyer et al.","year":"2012","title":"A High-Coverage Genome Sequence from an Archaic Denisovan Individual","container":"Science","volume":"338","pages":"222–226","doi":"10.1126/science.1224344","url":"https://doi.org/10.1126/science.1224344"},{"id":"chen2019","authors":"Chen, F. et al.","label":"Chen et al.","year":"2019","title":"A late Middle Pleistocene Denisovan mandible from the Tibetan Plateau","container":"Nature","volume":"569","pages":"409–412","doi":"10.1038/s41586-019-1139-x","url":"https://doi.org/10.1038/s41586-019-1139-x"},{"id":"douka2019","authors":"Douka, K. et al.","label":"Douka et al.","year":"2019","title":"Age estimates for hominin fossils and the onset of the Upper Palaeolithic at Denisova Cave","container":"Nature","volume":"565","pages":"640–644","doi":"10.1038/s41586-018-0870-z","url":"https://doi.org/10.1038/s41586-018-0870-z"},{"id":"xia2024","authors":"Xia, H. et al.","label":"Xia et al.","year":"2024","title":"Middle and Late Pleistocene Denisovan subsistence at Baishiya Karst Cave","container":"Nature","volume":"632","pages":"108–113","doi":"10.1038/s41586-024-07612-9","url":"https://doi.org/10.1038/s41586-024-07612-9"},{"id":"huerta2014","authors":"Huerta-Sánchez, E. et al.","label":"Huerta-Sánchez et al.","year":"2014","title":"Altitude adaptation in Tibetans caused by introgression of Denisovan-like DNA","container":"Nature","volume":"512","pages":"194–197","doi":"10.1038/nature13408","url":"https://doi.org/10.1038/nature13408"},{"id":"rao2026","authors":"Rao, H. et al.","label":"Rao et al.","year":"2026","title":"Ancient proteins identify various Denisovan remains from Southwest China","container":"Nature","note":"Published online 9 September 2026","doi":"10.1038/s41586-026-10976-9","url":"https://doi.org/10.1038/s41586-026-10976-9"},{"id":"fu2025mt","authors":"Fu, Q. et al.","label":"Fu et al.","year":"2025","title":"Denisovan mitochondrial DNA from dental calculus of the >146,000-year-old Harbin cranium","container":"Cell","volume":"188","pages":"3919–3926.e9","doi":"10.1016/j.cell.2025.05.040","url":"https://doi.org/10.1016/j.cell.2025.05.040"}]},{"name":"Flores","slug":"flores","chronology":{"yearsAgo":80000,"label":"~80,000 years ago","note":"Flores uses the Smithsonian's approximate 80,000-year age for the LB1 reference skeleton. The revised Liang Bua skeletal interval is approximately 100,000–60,000 years. This checkpoint estimate is separate from both the older end of that interval and the longer associated stone-artifact record; it is not a directly measured posterior median. [@si-lb1-age; @sutikna2016]"},"url":"https://hominidintelligence.org/flores","tag":"A compact footprint","description":"A compact island architecture that brings embodied operation and material tools into a smaller physical footprint. A distinct branch of the model family, shaped by a bounded environment.","capabilities":[{"title":"Compact embodied operation","text":"Flores brings a distinct physical configuration to the family. A smaller body and neural budget make the architecture a useful reference for developing capability within a compact footprint."},{"title":"Material tools within reach","text":"The environment remains part of the execution stack. Flores connects a compact embodied system with material operations, extending its working interface beyond the body."}],"parameters":{"count":"30.8B","neuronRange":[30.80939930304867,30.80939930304867],"brainMassGrams":null,"unit":"Neurons · estimated","full":"30.8 billion neurons, Hominid scaling extrapolation from LB1; not a cell count or species mean [@kubo2013; @herculano-kaas2011]","method":"Hominid calculation: N = 109,239,790.169 × M^0.923, with M = 426 cm³ × 1.06 g/cm³. The input is the published cranial capacity of LB1. Applying the primate scaling equation to this individual is an extrapolation; the fossil contains no measured neuron count. [@kubo2013; @herculano-kaas2011]","metric":["Checkpoint cranial capacity","426 cm³"]},"deployment":{"devices":1000,"count":"1K","method":"Hominid peak-deployment scenario: 10,000 km² of occupied habitat at 0.1 devices per km². Both inputs are assumptions; the resulting 1,000 devices are an illustrative census estimate."},"modelCard":{"sections":[{"id":"compact-body","title":"Develop within a compact footprint","paragraphs":["Flores brings a compact embodied architecture into the model family. The original description of *Homo floresiensis* established a small-bodied hominin at Liang Bua on the Indonesian island of Flores. LB1 is our reference checkpoint, connecting the release's physical scale to a named specimen. [@brown2004]","A later high-resolution reconstruction measured LB1's endocranial volume at 426 cubic centimetres. We use that published input for the parameter calculation. The compact footprint is part of the architecture's identity; volume alone does not quantify what the model could learn or accomplish. [@kubo2013]"]},{"id":"material-interface","title":"Keep useful operations outside the body","paragraphs":["The Liang Bua record includes stone artifacts attributed to *Homo floresiensis*. The revised chronology places those artifacts between about 190,000 and 50,000 years ago. We treat this material record as the basis for the release's tool interface. [@sutikna2016]","This architecture develops our family along a different dimension from a larger braincase. Useful action can depend on the body, an external cutting edge, and the material being worked. Flores makes that compact combination visible in the model catalog, with each capability grounded in the surviving physical record."]},{"id":"deployment-window","title":"Keep the checkpoint and its date together","paragraphs":["Renewed excavation revised the Liang Bua chronology. The skeletal remains are dated to approximately 100,000–60,000 years ago. Our timeline uses that interval for the architecture's displayed deployment, separately from the longer associated artifact record. [@sutikna2016]","Flores is a distinct model in the family, with its own body plan, reference checkpoint, and deployment context. We describe that combination directly. Questions about its deeper ancestry and the timing of its disappearance remain research questions rather than product version numbers."]}],"specs":[["Architecture","*Homo floresiensis*"],["Reference checkpoint","LB1"],["Deployment","Liang Bua, Flores, Indonesia"],["Dated skeletal record","Approximately 100,000–60,000 years ago"],["Associated tool record","Approximately 190,000–50,000 years ago"]],"specSources":"Architecture and checkpoint provenance follow {@brown2004}. Cranial measurement follows {@kubo2013}; the skeletal and associated-tool chronologies follow {@sutikna2016}.","evidence":[{"title":"A checkpoint is not a species average","text":"The 426 cm³ measurement describes LB1. The displayed neuron estimate applies a primate scaling equation to that volume; it is not a measured cell count or an IQ score. [@kubo2013; @herculano-kaas2011]"},{"title":"Separate skeletons from tools","text":"The skeletal and associated artifact intervals differ. Neither endpoint establishes the exact extinction date, and the revised chronology leaves overlap with arriving Sapiens populations unresolved. [@sutikna2016]"},{"title":"Keep deployment assumptions visible","text":"Peak devices are a Hominid occupied-habitat scenario, documented in the architecture research notes. The fossil record supplies no direct population census or numerical benchmark for language and reasoning."}],"reading":[{"slug":"compact-architecture","title":"Compact architecture, real-world tools"},{"slug":"world-models","title":"A world model that moves with you"}]},"references":[{"id":"kubo2013","authors":"Kubo, D., Kono, R. T. & Kaifu, Y.","label":"Kubo et al.","year":"2013","title":"Brain size of *Homo floresiensis* and its evolutionary implications","container":"Proceedings of the Royal Society B","volume":"280","pages":"20130338","doi":"10.1098/rspb.2013.0338","url":"https://doi.org/10.1098/rspb.2013.0338"},{"id":"herculano-kaas2011","authors":"Herculano-Houzel, S. & Kaas, J. H.","label":"Herculano-Houzel & Kaas","year":"2011","title":"Gorilla and Orangutan Brains Conform to the Primate Cellular Scaling Rules: Implications for Human Evolution","container":"Brain, Behavior and Evolution","volume":"77","pages":"33–44","doi":"10.1159/000322729","url":"https://doi.org/10.1159/000322729"},{"id":"si-lb1-age","authors":"Smithsonian National Museum of Natural History","label":"Smithsonian","year":"2016","title":"The extinct ‘hobbit’ of human evolution","container":"Human Origins Program","url":"https://humanorigins.si.edu/research/whats-hot-human-origins/new-date-homo-floresiensis"},{"id":"sutikna2016","authors":"Sutikna, T. et al.","label":"Sutikna et al.","year":"2016","title":"Revised stratigraphy and chronology for *Homo floresiensis* at Liang Bua in Indonesia","container":"Nature","volume":"532","pages":"366–369","doi":"10.1038/nature17179","url":"https://doi.org/10.1038/nature17179"},{"id":"brown2004","authors":"Brown, P. et al.","label":"Brown et al.","year":"2004","title":"A new small-bodied hominin from the Late Pleistocene of Flores, Indonesia","container":"Nature","volume":"431","pages":"1055–1061","doi":"10.1038/nature02999","url":"https://doi.org/10.1038/nature02999"}]}]}