[{"data":1,"prerenderedAt":1278},["ShallowReactive",2],{"sanity-7iXP3sOrSyrwpMoyaopqutrI_r2a7XNibHtOh578tGQ":3,"sanity-AUWeXPkttnorBLiisXPq1rXG5WzLiykYKsVtznzpOOw":200},{"data":4,"sourceMap":120},{"_createdAt":5,"_id":6,"_rev":7,"_system":8,"_type":6,"_updatedAt":11,"defaultTitle":12,"footerMenu1":13,"footerMenu2":29,"footerMenu3":44,"gitHubUrl":53,"huggingFaceUrl":57,"linkedInUrl":49,"mainNavigation":59,"ogDescription":68,"ogImage":69,"ogImageUrl":74,"showSiteNotice":75,"siteNotice":76},"2026-01-21T14:58:26Z","siteSettingsSchema","ocYd0qqIaBgdtQ8C7qEEox",{"base":9},{"id":6,"rev":10},"ZmwvPH6a8sPdSbem2nTeog","2026-05-06T11:56:07Z","kaiko.ai",{"links":14,"title":28},[15,20,24],{"_key":16,"_type":17,"label":18,"url":19},"95cf27020263","navigationItem","Product","/clinical-assistant",{"_key":21,"_type":17,"label":22,"url":23},"7b5bdfebb176","Research","/research",{"_key":25,"_type":17,"label":26,"url":27},"d311ee16a390","Insights Hub","/insights-hub","What we do",{"links":30,"title":43},[31,35,39],{"_key":32,"_type":17,"label":33,"url":34},"b56214c1f2ff","About us","/about",{"_key":36,"_type":17,"label":37,"url":38},"193660568291","Careers","https://jobs.kaiko.ai/",{"_key":40,"_type":17,"label":41,"url":42},"aeb117deb742","Trust Center","https://trust.kaiko.ai/","Company",{"links":45,"title":58},[46,50,54],{"_key":47,"_type":17,"label":48,"url":49},"40c6bf52fdd1","LinkedIn","https://www.linkedin.com/company/kaiko-ai/",{"_key":51,"_type":17,"label":52,"url":53},"f79c6a57ca2f","GitHub","https://github.com/kaiko-ai",{"_key":55,"_type":17,"label":56,"url":57},"7ff3680a1f90","Hugging Face","https://huggingface.co/kaiko-ai","Community",[60,62,64,66],{"_key":61,"_type":17,"label":18,"url":19},"601985fd4221",{"_key":63,"_type":17,"label":22,"url":23},"a10cc775319c",{"_key":65,"_type":17,"label":33,"url":34},"eb1707b39916",{"_key":67,"_type":17,"label":26,"url":27},"26d8cfff21e1","Your Clinical AI assistant that supports across full patient care, combining context for deeper insights and reducing workload, safe and compliant, made in EU.",{"_type":70,"asset":71},"image",{"_ref":72,"_type":73},"image-d9426451401bfd3574519485220b90bce4571422-1200x630-jpg","reference","https://cdn.sanity.io/images/a5kt9um3/production/d9426451401bfd3574519485220b90bce4571422-1200x630.jpg?w=1200&h=630&fit=crop",true,{"badges":77,"link":106,"linkText":107,"message":108},[78,99],{"_key":79,"color":80,"label":98},"70498638b5b3",{"_type":81,"alpha":82,"hex":83,"hsl":84,"hsv":89,"rgb":93},"color",1,"#a8e2df",{"_type":85,"a":82,"h":86,"l":87,"s":88},"hslaColor",176.89655172413794,0.7725490196078431,0.4999999999999999,{"_type":90,"a":82,"h":86,"s":91,"v":92},"hsvaColor",0.2566371681415929,0.8862745098039215,{"_type":94,"a":82,"b":95,"g":96,"r":97},"rgbaColor",223,226,168,"Online event",{"_key":100,"color":101,"label":105},"66755cd66e49",{"_type":81,"alpha":82,"hex":83,"hsl":102,"hsv":103,"rgb":104},{"_type":85,"a":82,"h":86,"l":87,"s":88},{"_type":90,"a":82,"h":86,"s":91,"v":92},{"_type":94,"a":82,"b":95,"g":96,"r":97},"24 June 2026","/insights-hub/kaiko-launch-registration","Register now",[109],{"_key":110,"_type":111,"children":112,"markDefs":118,"style":119},"515be88d5691","block",[113],{"_key":114,"_type":115,"marks":116,"text":117},"217241b6c7c6","span",[],"Pilot limbo: why most clinical AI never scales, and how some hospitals break the cycle",[],"normal",{"documents":121,"paths":125,"mappings":144},[122,124],{"_id":72,"_type":123},"sanity.imageAsset",{"_id":6,"_type":6},[126,127,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143],"$['_createdAt']","$['_id']","$['_rev']","$['_system']","$['_type']","$['_updatedAt']","$['defaultTitle']","$['footerMenu1']","$['footerMenu2']","$['footerMenu3']","$['gitHubUrl']","$['huggingFaceUrl']","$['linkedInUrl']","$['mainNavigation']","$['ogDescription']","$['ogImage']","$['siteNotice']","$['url']",{"$['_createdAt']":145,"$['_id']":150,"$['_rev']":152,"$['_system']":155,"$['_type']":158,"$['_updatedAt']":161,"$['defaultTitle']":164,"$['footerMenu1']":167,"$['footerMenu2']":170,"$['footerMenu3']":173,"$['gitHubUrl']":176,"$['huggingFaceUrl']":179,"$['linkedInUrl']":182,"$['mainNavigation']":185,"$['ogDescription']":188,"$['ogImage']":191,"$['ogImageUrl']":194,"$['siteNotice']":197},{"source":146,"type":149},{"document":82,"path":147,"type":148},0,"documentValue","value",{"source":151,"type":149},{"document":82,"path":82,"type":148},{"source":153,"type":149},{"document":82,"path":154,"type":148},2,{"source":156,"type":149},{"document":82,"path":157,"type":148},3,{"source":159,"type":149},{"document":82,"path":160,"type":148},4,{"source":162,"type":149},{"document":82,"path":163,"type":148},5,{"source":165,"type":149},{"document":82,"path":166,"type":148},6,{"source":168,"type":149},{"document":82,"path":169,"type":148},7,{"source":171,"type":149},{"document":82,"path":172,"type":148},8,{"source":174,"type":149},{"document":82,"path":175,"type":148},9,{"source":177,"type":149},{"document":82,"path":178,"type":148},10,{"source":180,"type":149},{"document":82,"path":181,"type":148},11,{"source":183,"type":149},{"document":82,"path":184,"type":148},12,{"source":186,"type":149},{"document":82,"path":187,"type":148},13,{"source":189,"type":149},{"document":82,"path":190,"type":148},14,{"source":192,"type":149},{"document":82,"path":193,"type":148},15,{"source":195,"type":149},{"document":147,"path":196,"type":148},17,{"source":198,"type":149},{"document":82,"path":199,"type":148},16,{"data":201,"sourceMap":1211},{"_createdAt":202,"_id":203,"_rev":204,"_system":205,"_type":208,"_updatedAt":209,"area":210,"authors":212,"categories":216,"content":220,"date":899,"featuredImage":900,"isHidden":903,"relatedPosts":904,"slug":908,"subHeading":911,"suggestedPosts":912,"title":1200,"topics":1201},"2026-06-02T11:36:58Z","38a85cd4-0666-4e99-b794-b9b8371c12e9","DBg3LM0vF22ywEKWDVO0dm",{"base":206},{"id":203,"rev":207},"GInI3CC6S35oK4J4gwnIB1","post","2026-06-10T09:23:43Z",[211],"engineering",[213],{"_key":214,"_ref":215,"_type":73},"6b51935126c4","737d8376-7ded-403e-98a6-30fa6b4c87e7",[217],{"_key":218,"_ref":219,"_type":73},"a78b4a06e8fc","9613ea96-3dba-4c14-9ac9-2b5fe14eac53",[221,228,315,318,340,346,385,387,415,421,432,438,485,504,510,512,539,550,556,583,638,640,711,713,871,873],{"_key":222,"_type":223,"caption":224,"image":225},"9566649717a4","imageSection","Figure 1-4: Ground-truth annotations vs CoralBay predictions on various datasets.",{"_type":70,"asset":226},{"_ref":227,"_type":73},"image-f2f8cd6857b5225f5e7a438ea0fbd3b8520ccbbb-1280x720-gif",{"_key":229,"_type":230,"text":231},"64220c0acaa7","textSection",[232,240,257,266,279,291,303],{"_key":233,"_type":111,"children":234,"markDefs":239,"style":119},"35765f76c825",[235],{"_key":236,"_type":115,"marks":237,"text":238},"8c81e04012ee",[],"Each year, over 300 million CT scans are captured as rich, high-fidelity 3D volumes that give doctors a detailed view of the human body for diagnosis and surgical planning. However, the status quo in current AI models is to process these volumes as stacks of 2D slices, effectively flattening the data and discarding critical spatial relationships, or to rely on complicated, highly specialised methods that are difficult to scale and build upon.",[],{"_key":241,"_type":111,"children":242,"markDefs":256,"style":119},"b331bc6804ad",[243,247,252],{"_key":244,"_type":115,"marks":245,"text":246},"7bca31b804f0",[],"Enter ",{"_key":248,"_type":115,"marks":249,"text":251},"cd673be05ab4",[250],"strong","CoralBay",{"_key":253,"_type":115,"marks":254,"text":255},"aacb86bcdcce",[],": our native 3D foundation model for CT, pre-trained on only 11K unlabeled volumes. With a simple architecture and training pipeline, it produces meaningful representations from raw data that generalize across the full clinical stack, from scan-level classification down to fine-grained lesion segmentation.",[],{"_key":258,"_type":111,"children":259,"markDefs":264,"style":265},"54b86dc85d82",[260],{"_key":261,"_type":115,"marks":262,"text":263},"6a2cc7267166",[250],"Quick Highlights",[],"h2",{"_key":267,"_type":111,"children":268,"level":82,"listItem":277,"markDefs":278,"style":119},"4d251260c971",[269,273],{"_key":270,"_type":115,"marks":271,"text":272},"05b0b7dc5891",[250],"Native 3D Intelligence:",{"_key":274,"_type":115,"marks":275,"text":276},"e07ecffd95d4",[]," Understands anatomy as continuous volumetric structures.","bullet",[],{"_key":280,"_type":111,"children":281,"level":82,"listItem":277,"markDefs":290,"style":119},"8caec661b008",[282,286],{"_key":283,"_type":115,"marks":284,"text":285},"abe2ae64d09d",[250],"Simple, Strong, Generalizable:",{"_key":287,"_type":115,"marks":288,"text":289},"1b030a7961bb",[]," Strong multi-task performance from a simple training pipeline and just 11K unlabeled volumes, ready to adapt to various downstream tasks.",[],{"_key":292,"_type":111,"children":293,"level":82,"listItem":277,"markDefs":302,"style":119},"fee963bbe5c7",[294,298],{"_key":295,"_type":115,"marks":296,"text":297},"18b325e8ae18",[250],"Clinical Depth:",{"_key":299,"_type":115,"marks":300,"text":301},"55efa3c6f50e",[]," A single backbone with built-in invariance to HU values, excelling across the full task spectrum from broad organ identification to fine-grained tumour segmentation",[],{"_key":304,"_type":111,"children":305,"level":82,"listItem":277,"markDefs":314,"style":119},"43c283f1419c",[306,310],{"_key":307,"_type":115,"marks":308,"text":309},"21d4328d42dc",[250],"Open Science:",{"_key":311,"_type":115,"marks":312,"text":313},"bfdcf0b49a91",[]," Fully open-source weights and benchmarks to accelerate global medical AI research.",[],{"_key":316,"_type":317},"4be33ed573b9","divider",{"_key":319,"_type":230,"text":320},"f8ff5b1d9a48",[321,329],{"_key":322,"_type":111,"children":323,"markDefs":328,"style":265},"bd0845d89d25",[324],{"_key":325,"_type":115,"marks":326,"text":327},"5bd09ab7d8c7",[250],"The Challenge: 3D is Different",[],{"_key":330,"_type":111,"children":331,"markDefs":339,"style":119},"b8528c9d14ad",[332,336],{"_key":333,"_type":115,"marks":334,"text":335},"1576c6dfdc1d",[],"Standard vision models assume 2D RGB images; three colour channels, consistent resolution, and millions of available labeled examples. CT scans violate every one of these assumptions.",{"_key":325,"_type":115,"marks":337,"text":338},[250],"\n",[],{"_key":341,"_type":223,"caption":342,"image":343},"b34e97709062","Figure 5: Challenges in 3D data representation. Left: Narrow windows clarify soft tissue; wide windows preserve high-density data. Center: Thin slices increase resolution; thick slices improve SNR but cause blurring. Right: 3D consistency is required across all orthogonal planes.",{"_type":70,"asset":344},{"_ref":345,"_type":73},"image-b15c3a2413e2743044aa8a6f40419291816065e1-2857x860-png",{"_key":347,"_type":230,"text":348},"7170efbff7f2",[349,361,373],{"_key":350,"_type":111,"children":351,"level":82,"listItem":277,"markDefs":360,"style":119},"13c3b67b8f91",[352,356],{"_key":353,"_type":115,"marks":354,"text":355},"951db6b47c8a",[250],"Hounsfield Units, not pixels",{"_key":357,"_type":115,"marks":358,"text":359},"68968e710212",[],": Intensities reflect tissue density, not color. Different window settings highlight different anatomy (e.g., lungs vs. soft tissue), so models must be robust to multiple visualizations of the same scan.",[],{"_key":362,"_type":111,"children":363,"level":82,"listItem":277,"markDefs":372,"style":119},"d5c75cd01feb",[364,368],{"_key":365,"_type":115,"marks":366,"text":367},"fbec56375347",[250],"Anisotropic resolution",{"_key":369,"_type":115,"marks":370,"text":371},"48c80c0a9f45",[],": Slice thickness varies (e.g., 1–5 mm), causing partial volume effects. Thicker slices improve signal-to-noise but blur fine structures, unlike typical image noise.",[],{"_key":374,"_type":111,"children":375,"level":82,"listItem":277,"markDefs":384,"style":119},"a5fd44829f4a",[376,380],{"_key":377,"_type":115,"marks":378,"text":379},"88bb51682b5d",[250],"Volumetric spatial complexity",{"_key":381,"_type":115,"marks":382,"text":383},"e4263fe818ca",[],": CT is inherently 3D. Small findings, including early-stage tumours that may occupy only a handful of voxels, require integrating information across the full 3D space to be reliably detected. Treating slices independently breaks the spatial continuity that makes these critical, hard-to-spot findings visible in the first place.",[],{"_key":386,"_type":317},"17ece86911bc",{"_key":388,"_type":230,"text":389},"910b8f205540",[390,398,407],{"_key":391,"_type":111,"children":392,"markDefs":397,"style":265},"e1031f7eaa2b",[393],{"_key":394,"_type":115,"marks":395,"text":396},"1958c0300072",[250],"The CoralBay Approach",[],{"_key":399,"_type":111,"children":400,"markDefs":405,"style":406},"24cf30dce250",[401],{"_key":402,"_type":115,"marks":403,"text":404},"cfef1df31105",[250],"Hierarchical 3D Feature Learning",[],"h3",{"_key":408,"_type":111,"children":409,"markDefs":414,"style":119},"22efd7486e59",[410],{"_key":411,"_type":115,"marks":412,"text":413},"26123db28de3",[],"CoralBay extends DINO self-supervised learning to native 3D CT volumes. Using a hierarchical 3D Swin-Transformer, it captures long-range spatial relationships efficiently and learns multi-scale features ranging from organ-level anatomy to fine structures such as vessels.",[],{"_key":416,"_type":223,"caption":417,"image":418},"fc44b831cb37","Figure 6: CoralBay’s technical overviews of the training pipeline",{"_type":70,"asset":419},{"_ref":420,"_type":73},"image-dbb59b4f247bc0d580e639560876d93759308654-2983x830-png",{"_key":422,"_type":230,"text":423},"e24b6ad5a981",[424],{"_key":425,"_type":111,"children":426,"markDefs":431,"style":406},"b5b27746eff8",[427],{"_key":428,"_type":115,"marks":429,"text":430},"17448af3114b",[250],"Radiology-Specific Augmentations",[],{"_key":433,"_type":223,"caption":434,"image":435},"d5b91e32677e","Figure 7: HU ranges for pre-training data augmentation.",{"_type":70,"asset":436},{"_ref":437,"_type":73},"image-81151700e6702f95aea34f954a36249de02bdf7b-1871x662-png",{"_key":439,"_type":230,"text":440},"cef125988b64",[441,449,461,473],{"_key":442,"_type":111,"children":443,"markDefs":448,"style":119},"0183fefb6cf8",[444],{"_key":445,"_type":115,"marks":446,"text":447},"0aebd20a500b",[],"CoralBay replaces generic image augmentations with CT-specific transformations that reflect real-world imaging variability.",[],{"_key":450,"_type":111,"children":451,"level":82,"listItem":277,"markDefs":460,"style":119},"6cb032b32dc8",[452,456],{"_key":453,"_type":115,"marks":454,"text":455},"ed6a8d78c81c",[250],"Random HU Windowing:",{"_key":457,"_type":115,"marks":458,"text":459},"4950f307840f",[]," Samples from multiple clinically relevant HU ranges (e.g., lung, liver, brain, abdomen, and full CT) to learn features that are robust across viewing settings.",[],{"_key":462,"_type":111,"children":463,"level":82,"listItem":277,"markDefs":472,"style":119},"ce21392c405f",[464,468],{"_key":465,"_type":115,"marks":466,"text":467},"4773b4688429",[250],"Scanner robustness:",{"_key":469,"_type":115,"marks":470,"text":471},"05f3208926e3",[]," Gaussian smoothing and histogram shifts simulate differences in scanners and reconstruction protocols, improving generalization across sites.",[],{"_key":474,"_type":111,"children":475,"level":82,"listItem":277,"markDefs":484,"style":119},"2acdfbb129d2",[476,480],{"_key":477,"_type":115,"marks":478,"text":479},"cfec638253be",[250],"Local-to-global 3D context:",{"_key":481,"_type":115,"marks":482,"text":483},"e6d600c2cfe9",[]," Combines large and small 3D crops during training, helping the model link fine pathological details with their surrounding anatomy.\n\n",[],{"_key":486,"_type":230,"text":487},"15ad5094a40b",[488,496],{"_key":489,"_type":111,"children":490,"markDefs":495,"style":406},"be04f790ac1e",[491],{"_key":492,"_type":115,"marks":493,"text":494},"8da62902f193",[250],"Scan-Level Inference via Sliding Window",[],{"_key":497,"_type":111,"children":498,"markDefs":503,"style":119},"b01bb1b885eb",[499],{"_key":500,"_type":115,"marks":501,"text":502},"2e22e58a177a",[],"While the backbone is trained on 96×96×96 crops, inference uses a sliding-window approach: the full scan is divided into overlapping 3D patches, each encoded independently, then stitched together. For classification, pooling merges the patch-level features into a single scan-level vector. For segmentation, features are passed to a Swin-UNETR decoder with skip connections, which preserves fine boundary details for voxel-wise segmentation.",[],{"_key":505,"_type":223,"caption":506,"image":507},"b53c3ec1a4f0","Figure 8a-b: 3D scan-level classification (left) and segmentation (right) sliding window Inference",{"_type":70,"asset":508},{"_ref":509,"_type":73},"image-9c7d3b91caa94b4ccf1f486106b39d3adff64310-1080x608-gif",{"_key":511,"_type":317},"9d1f2386f430",{"_key":513,"_type":230,"text":514},"373492d8a965",[515,523,531],{"_key":516,"_type":111,"children":517,"markDefs":522,"style":265},"27742a556a73",[518],{"_key":519,"_type":115,"marks":520,"text":521},"49a15d39b191",[250],"How CoralBay Compares",[],{"_key":524,"_type":111,"children":525,"markDefs":530,"style":119},"d9730d5a19bc",[526],{"_key":527,"_type":115,"marks":528,"text":529},"5ef80e386a2f",[],"As the CT foundation model space matures, CoralBay can be understood along three axes: how well it transfers to downstream tasks with both frozen and fine-tuned encoders, how little pre-training data it requires, and how broadly it generalises across classification and segmentation tasks.",[],{"_key":532,"_type":111,"children":533,"markDefs":538,"style":119},"edf6185361e2",[534],{"_key":535,"_type":115,"marks":536,"text":537},"9a9870265e97",[],"To test this, we evaluated CoralBay on 11 datasets spanning classification and segmentation across diverse anatomical targets, with two model variants: CoralBayU96B (53.2M) and CoralBayU96H (847M).",[],{"_key":540,"_type":230,"text":541},"546419008d33",[542],{"_key":543,"_type":111,"children":544,"markDefs":549,"style":265},"454658f66a2e",[545],{"_key":546,"_type":115,"marks":547,"text":548},"000230fac4c6",[],"Quantitative Results",[],{"_key":551,"_type":223,"caption":552,"image":553},"7a5cce16e0ad","Table A: Quantitative performance across classification (Multi-class Accuracy/Binary AUROC) and segmentation (Dice score) tasks, as evaluated via the eva framework.",{"_type":70,"asset":554},{"_ref":555,"_type":73},"image-d3bd456a4f8a0285bf1592543ff451c5d2655910-3881x1864-png",{"_key":557,"_type":230,"text":558},"f0202c030fc1",[559,571],{"_key":560,"_type":111,"children":561,"markDefs":570,"style":119},"444f6936a4f7",[562,566],{"_key":563,"_type":115,"marks":564,"text":565},"0a2fa2c6c444",[250],"- Classification:",{"_key":567,"_type":115,"marks":568,"text":569},"246c3c863c32",[]," CoralBayU96H achieves the best or tied-best frozen-encoder results across all four classification benchmarks — organ identification (OrganMNIST3D), lung nodule malignancy (NoduleMNIST3D, LUNA25), and COVID-19 classification (CC-CCII).",[],{"_key":572,"_type":111,"children":573,"markDefs":582,"style":119},"72429b2441da",[574,578],{"_key":575,"_type":115,"marks":576,"text":577},"65e6b282dfa5",[250],"- Segmentation:",{"_key":579,"_type":115,"marks":580,"text":581},"66e3b870c73b",[]," With a frozen encoder and a lightweight 22.8M-parameter decoder, CoralBay performs comparably to VoCo across seven segmentation benchmarks despite the significant data gap.\n\n",[],{"_key":584,"_type":230,"text":585},"11e22cc37247",[586,594,602,614,626],{"_key":587,"_type":111,"children":588,"markDefs":593,"style":265},"038664d34317",[589],{"_key":590,"_type":115,"marks":591,"text":592},"a88282b576bf",[],"Qualitative Analysis",[],{"_key":595,"_type":111,"children":596,"markDefs":601,"style":119},"6e4b26c4cad3",[597],{"_key":598,"_type":115,"marks":599,"text":600},"eb8962ac6119",[],"Visualising the model's performance (Figures 1-4) reveals its precision across diverse anatomical challenges:",[],{"_key":603,"_type":111,"children":604,"level":82,"listItem":277,"markDefs":613,"style":119},"ff3b8ca7829f",[605,609],{"_key":606,"_type":115,"marks":607,"text":608},"bb4643edcecb",[250],"Multi-Organ Segmentation:",{"_key":610,"_type":115,"marks":611,"text":612},"9b40d7d71e77",[]," On benchmarks like BTCV and FLARE22, the model accurately delineates complex abdominal structures including the liver, spleen, and kidneys with high spatial consistency.",[],{"_key":615,"_type":111,"children":616,"level":82,"listItem":277,"markDefs":625,"style":119},"24bc2f23c9f5",[617,621],{"_key":618,"_type":115,"marks":619,"text":620},"b15a6cd2934f",[250],"Lesion and tumour boundaries:",{"_key":622,"_type":115,"marks":623,"text":624},"b5f90a41d68e",[]," Evaluation on LiTS17 and MSD Task 7 demonstrates its ability to resolve fine-grained tumor boundaries within the liver and pancreas, critical for clinical diagnostics.",[],{"_key":627,"_type":111,"children":628,"level":82,"listItem":277,"markDefs":637,"style":119},"e824eaf2088d",[629,633],{"_key":630,"_type":115,"marks":631,"text":632},"7101abc76ded",[250],"Reliability:",{"_key":634,"_type":115,"marks":635,"text":636},"93edb776f4ca",[]," The difference maps highlight minimal variance between CoralBay segmentations and ground-truth annotations, confirming robust feature capture in low-contrast regions.",[],{"_key":639,"_type":317},"98cc8424d8fb",{"_key":641,"_type":230,"text":642},"28f603846357",[643,651,659,679,699],{"_key":644,"_type":111,"children":645,"markDefs":650,"style":265},"2958a8aa097b",[646],{"_key":647,"_type":115,"marks":648,"text":649},"5eb56301d4d0",[250],"What the Ablations Reveal",[],{"_key":652,"_type":111,"children":653,"markDefs":658,"style":119},"624b97cc39b6",[654],{"_key":655,"_type":115,"marks":656,"text":657},"e79ff7601a16",[],"Ablation studies in the paper confirm that CoralBay’s performance arises from a combination of structural and data-driven innovations:",[],{"_key":660,"_type":111,"children":661,"level":82,"listItem":277,"markDefs":678,"style":119},"9786e5156e10",[662,666,670,674],{"_key":663,"_type":115,"marks":664,"text":665},"aabe76e00128",[250],"Native 3D Dominance:",{"_key":667,"_type":115,"marks":668,"text":669},"adce4cd9683d",[]," Switching from the native dino 2D to CoralBay 3D spatial modeling provides a substantial improvement of ",{"_key":671,"_type":115,"marks":672,"text":673},"12e0269e2fba",[250],"20%",{"_key":675,"_type":115,"marks":676,"text":677},"3a46a6f7539b",[]," in Dice score (BTCV), proving that consistent 3D inductive biases are essential for medical volumes.",[],{"_key":680,"_type":111,"children":681,"level":82,"listItem":277,"markDefs":698,"style":119},"fe64aed92a06",[682,686,690,694],{"_key":683,"_type":115,"marks":684,"text":685},"e7647ed0f39d",[250],"Effective Scaling",{"_key":687,"_type":115,"marks":688,"text":689},"195dd081c52b",[],": Segmentation accuracy improves consistently as both the model size and the pre-training dataset grow, resulting in an overall ",{"_key":691,"_type":115,"marks":692,"text":693},"80af9713004a",[250],"15% improvement",{"_key":695,"_type":115,"marks":696,"text":697},"5d5a95f56819",[]," (LiTS17). This demonstrates the framework’s ability to benefit from larger data pools and suggests that training with even more data could further strengthen the model.",[],{"_key":700,"_type":111,"children":701,"level":82,"listItem":277,"markDefs":710,"style":119},"bdce70c68041",[702,706],{"_key":703,"_type":115,"marks":704,"text":705},"eefc607d4e17",[250],"Superior Label Efficiency:",{"_key":707,"_type":115,"marks":708,"text":709},"5b3b36cdcf4e",[]," On challenging tumor tasks, self-supervised pre-training acts as a powerful prior, outperforming heavily tuned models in low-data settings.",[],{"_key":712,"_type":317},"8c07382648bb",{"_key":714,"_type":230,"text":715},"fcdb71887c26",[716,724,758,766,774,782,794,818,830],{"_key":717,"_type":111,"children":718,"markDefs":723,"style":265},"6da8f2643ea9",[719],{"_key":720,"_type":115,"marks":721,"text":722},"509fc72c3ffe",[250],"Why this matters for kaiko",[],{"_key":725,"_type":111,"children":726,"markDefs":757,"style":119},"39467400a5c4",[727,730,734,737,741,745,749,753],{"_key":728,"_type":115,"marks":729,"text":251},"c0638d99be1f",[250],{"_key":731,"_type":115,"marks":732,"text":733},"f43458ba626d",[]," establishes a powerful 3D foundation for radiological data. Today, it masters native spatial reasoning and physics-informed intensity invariance, delivering highly data-efficient, clinically robust models for classification and segmentation across organs, pathologies, and scanners. This raises the ceiling for high‑resolution pathology detection while also providing a standard, open benchmark through the 3D radiology leaderboard. \n\nLooking ahead, ",{"_key":735,"_type":115,"marks":736,"text":251},"941e381f5bd5",[250],{"_key":738,"_type":115,"marks":739,"text":740},"4614f3441733",[]," serves as a visual anchor for ",{"_key":742,"_type":115,"marks":743,"text":744},"23837e082ca5",[250],"Multimodal Medical Intelligence",{"_key":746,"_type":115,"marks":747,"text":748},"6b0544eff428",[],", where AI systems jointly interpret imaging, text, lab values, and other clinical signals in a unified clinical context. Built on a strong 3D visual backbone, it moves toward an ",{"_key":750,"_type":115,"marks":751,"text":752},"641b143a9e24",[250],"agentic vision paradigm",{"_key":754,"_type":115,"marks":755,"text":756},"469ed812d54a",[],", where models go beyond interpretation to actively reason over scans, plan multi-step analyses, and coordinate tool use across modalities and time.",[],{"_key":759,"_type":111,"children":760,"markDefs":765,"style":119},"f884f3cd83a7",[761],{"_key":762,"_type":115,"marks":763,"text":764},"71631a68b7ae",[],"",[],{"_key":767,"_type":111,"children":768,"markDefs":773,"style":265},"f17bcd4148c4",[769],{"_key":770,"_type":115,"marks":771,"text":772},"3506c3bc6681",[250],"Open Science",[],{"_key":775,"_type":111,"children":776,"markDefs":781,"style":119},"8ad16fdfcb89",[777],{"_key":778,"_type":115,"marks":779,"text":780},"b2767c4e8ed4",[],"CoralBay promotes open, reproducible evaluation in medical AI through core contributions:",[],{"_key":783,"_type":111,"children":784,"level":82,"listItem":277,"markDefs":793,"style":119},"e8a7a8dcb4c7",[785,789],{"_key":786,"_type":115,"marks":787,"text":788},"1aa0531f239e",[250],"Comprehensive Whitepaper:",{"_key":790,"_type":115,"marks":791,"text":792},"ae32bdf52146",[]," We provide a detailed whitepaper describing the model, training process, datasets, evaluation setup, and results.",[],{"_key":795,"_type":111,"children":796,"level":82,"listItem":277,"markDefs":814,"style":119},"f4b10baeca14",[797,801,805,810],{"_key":798,"_type":115,"marks":799,"text":800},"11e7b8fa0e9e",[250],"Standardized Benchmarking:",{"_key":802,"_type":115,"marks":803,"text":804},"17b681ae0a2b",[]," We’re scaling 🔗 ",{"_key":806,"_type":115,"marks":807,"text":809},"556e5f7ffd40",[250,808],"ca96badbe275","eva",{"_key":811,"_type":115,"marks":812,"text":813},"14d64945d831",[]," — our open-source evaluation framework — to include comprehensive support for 3D radiology. This includes standardized data loaders, model backbones, and a public leaderboard designed to make 3D medical AI research transparent, reproducible, and easily comparable.",[815],{"_key":808,"_type":816,"href":817},"link","https://github.com/kaiko-ai/eva",{"_key":819,"_type":111,"children":820,"level":82,"listItem":277,"markDefs":829,"style":119},"9057a5c47442",[821,825],{"_key":822,"_type":115,"marks":823,"text":824},"71e126998f48",[250],"Open Access:",{"_key":826,"_type":115,"marks":827,"text":828},"f3b605346cff",[]," Model weights are available on GitHub and Hugging Face.",[],{"_key":831,"_type":111,"children":832,"markDefs":864,"style":119},"402b38cf12a7",[833,837,842,846,851,855,860],{"_key":834,"_type":115,"marks":835,"text":836},"68c6ea89183c",[],"Explore the 🔗 ",{"_key":838,"_type":115,"marks":839,"text":841},"390e9ca438c4",[840],"3ca4e837b5ae","paper",{"_key":843,"_type":115,"marks":844,"text":845},"f9ffddb50552",[],", 🔗 ",{"_key":847,"_type":115,"marks":848,"text":850},"3633c31e6832",[849],"f06242a5f1f7","leaderboard",{"_key":852,"_type":115,"marks":853,"text":854},"aea443f75bfb",[],", and 🔗 ",{"_key":856,"_type":115,"marks":857,"text":859},"44c5f83a91c5",[858],"bd67fdf5c7ad","model weights",{"_key":861,"_type":115,"marks":862,"text":863},"e77c19555e00",[]," to learn more.",[865,867,869],{"_key":840,"_type":816,"href":866},"https://arxiv.org/abs/2606.03888",{"_key":849,"_type":816,"href":868},"https://github.com/kaiko-ai/eva#radiology",{"_key":858,"_type":816,"href":870},"https://huggingface.co/kaiko-ai/coralbay",{"_key":872,"_type":317},"906c8f045b1f",{"_key":874,"_type":230,"text":875},"7e8094857822",[876,885,892],{"_key":877,"_type":111,"children":878,"markDefs":883,"style":884},"08d8e875db33",[879],{"_key":880,"_type":115,"marks":881,"text":882},"d50b6c35f89e",[],"CoralBay is released under the MIT License and is intended solely for research purposes. This model has not been validated or certified for clinical use and must not be used to inform, guide, or replace clinical decision-making, diagnosis, treatment, or patient management.",[],"disclaimer",{"_key":886,"_type":111,"children":887,"markDefs":891,"style":884},"28407ab80679",[888],{"_key":880,"_type":115,"marks":889,"text":890},[],"\nCoralBay's capabilities — including tumor segmentation, volume estimation, lesion tracking, and scan-level diagnosis — have been evaluated in research settings only. Performance may not generalise across patient populations, scanner hardware, imaging protocols, or clinical environments. The model has not been reviewed or approved by any regulatory authority (including but not limited to the FDA or under the EU MDR).",[],{"_key":893,"_type":111,"children":894,"markDefs":898,"style":884},"85d8875c981d",[895],{"_key":880,"_type":115,"marks":896,"text":897},[],"\nUsers are solely responsible for assessing suitability for their intended use. Kaiko strongly recommends independent validation, ethical review, and compliance with applicable laws before any downstream application — particularly in healthcare AI contexts. CoralBay is provided \"as is\", without warranty of any kind. Kaiko accepts no liability for outcomes arising from its use.",[],"2026-06-10T09:00:00.000Z",{"_type":70,"asset":901},{"_ref":902,"_type":73},"image-0327c08822c6c84776553e305bd13275fb6dc913-2011x941-png",false,[905],{"_key":906,"_ref":907,"_type":73},"63b393e2b21e","e3bae32e-f001-4ba4-8f4f-e9e38af8cc33",{"_type":909,"current":910},"slug","CoralBay-3D-Foundation-Model","Introducing CoralBay: A fully open-source, native 3D foundation model for radiology that delivers state-of-the-art CT scan analysis with unmatched data efficiency.",[913],{"categories":914,"content":917,"date":1183,"featuredImage":1184,"slug":1186,"subHeading":1188,"title":1189,"topics":1190},[915],{"_key":916,"_ref":219,"_type":73},"61ec7569661d",[918,922,982,988,1037,1043,1054,1060,1158,1169],{"_key":919,"_type":920,"text":921},"97e89f0381c7","tldrSection","Due to generalisation failures in the current AI paradigm, capability is fundamentally shaped by its training environment. As a result, we bet against a single dominant AI and argue that the flywheel of domain-specific training loops & deployment will outcompete generalist frontier models and explain how this is feasible for strong vertical players under resource constraints.",{"_key":923,"_type":230,"text":924},"57c9d6526610",[925,933,941,966,974],{"_key":926,"_type":111,"children":927,"markDefs":932,"style":119},"766f142374c8",[928],{"_key":929,"_type":115,"marks":930,"text":931},"2157710823b2",[],"Frontier models can write poetry, generate working code, and prove mathematical theorems and then, without warning, become strangely incompetent at things that feel adjacent. They are “jagged”, which means their capability profile is not evenly distributed. ",[],{"_key":934,"_type":111,"children":935,"markDefs":940,"style":119},"cc27a2343632",[936],{"_key":937,"_type":115,"marks":938,"text":939},"abfd9eb6f201",[],"There are two ways to read this: One view says the jaggedness is temporary - scale and better methods will smooth it out. Capability converges, sufficiently powerful systems do everything well, training context becomes increasingly irrelevant, one dominant general intelligence emerges and all else is a wrapper around it. The other view says the jaggedness is structural and capability remains shaped by training context because the entire current AI paradigm shows insufficient generalisation ability. In this world general systems keep getting better at general tasks, but domains with genuinely different constraints keep rewarding specialized training because the error costs, feedback loops, workflows & data are different.",[],{"_key":942,"_type":111,"children":943,"markDefs":965,"style":119},"8ae1662483a6",[944,948,953,957,961],{"_key":945,"_type":115,"marks":946,"text":947},"48524ddfd38c",[],"Most AI discourse implicitly assumes the first view. The \"models are commoditizing, so bet on the application layer\" argument is a version of it. But that argument contains a contradiction. If models are commoditizing because capability is converging, and convergence continues to its natural conclusion: A god-like model that excels at everything (the public-facing stance of current frontier labs",{"_key":949,"_type":115,"marks":950,"text":952},"ff454deadf51",[951],"sup","1",{"_key":954,"_type":115,"marks":955,"text":956},"0da0e52aad75",[],"), then the application layer necessarily compresses too. The model does ",{"_key":958,"_type":115,"marks":959,"text":960},"87f57185ac2e",[250],"everything",{"_key":962,"_type":115,"marks":963,"text":964},"8223cf9048a3",[]," well, hence executes entire workflows correctly and reliably. Integration, governance and monitoring become standardized plumbing around this oracle with no differentiation between application players.",[],{"_key":967,"_type":111,"children":968,"markDefs":973,"style":119},"1b248f40f5a2",[969],{"_key":970,"_type":115,"marks":971,"text":972},"f57288ec6a88",[],"The \"bet on apps\" advice therefore must implicitly assume convergence is real enough to commoditize models but somehow stops before it commoditizes applications, horizontal convergence happens but vertical convergence does not but that's not a coherent position. Either capability converges all the way, and the model layer captures all verticals and the entire stack within those verticals, or the environment keeps mattering and specialized training keeps producing differentiated capability. You can believe one or the other, but not both.",[],{"_key":975,"_type":111,"children":976,"markDefs":981,"style":119},"fc176f1b5f1f",[977],{"_key":978,"_type":115,"marks":979,"text":980},"b5f696cb4d93",[],"Within the current paradigm, we're betting on structural. The reason comes from how training works. The story of 2020–2024 was pre-training at scale and involved more data, more compute, and capabilities improving together. But pre-training is only the foundation. Everything after, continued training on domain distributions, instruction tuning, and reinforcement learning, works differently. These later stages don't lift all boats because they make choices that privilege some capabilities while suppressing others.",[],{"_key":983,"_type":223,"caption":984,"image":985},"e3651bb5cacb","Figure 1. Human intelligence (black) vs AI intelligence (coral) – both are jagged, but differently so, fundamentally due to having evolved in different environments. Adapted from Karpathy [2]",{"_type":70,"asset":986},{"_ref":987,"_type":73},"image-32a32327d1af402c6c9e5d56486ac30f0c096ea3-902x722-png",{"_key":989,"_type":230,"text":990},"c21aac851e87",[991,999,1006,1014,1021],{"_key":992,"_type":111,"children":993,"markDefs":998,"style":119},"7a449c4c03f0",[994],{"_key":995,"_type":115,"marks":996,"text":997},"cf3e0d8a2e33",[],"Once you move past generic next-token prediction and start optimizing for particular outcomes, you're shaping what kind of intelligence the model becomes. The data distributions you choose, the reward signals you design and the benchmarks you evaluate against are selection pressures.",[],{"_key":1000,"_type":111,"children":1001,"markDefs":1005,"style":119},"81fdc8450f04",[1002],{"_key":1003,"_type":115,"marks":1004,"text":764},"7a0f8af8fcf4",[],[],{"_key":1007,"_type":111,"children":1008,"markDefs":1013,"style":119},"f2545127dc68",[1009],{"_key":1010,"_type":115,"marks":1011,"text":1012},"76425903d53f",[],"In clinical medicine the gap between \"benchmark performance\" and \"operational capability\" is particularly stark. On radiology quiz questions, models can look strong but when you move to messier clinical work such as integrating patient history, handling rare cases and staying calibrated about uncertainty performance falls apart in ways that would get a resident pulled from the rotation. These aren't properties you currently get for free from pre-training on internet text. Most of the value, and most of the danger, lives in this gap.",[],{"_key":1015,"_type":111,"children":1016,"markDefs":1020,"style":119},"8265fcddebe4",[1017],{"_key":1018,"_type":115,"marks":1019,"text":764},"9188c7060978",[],[],{"_key":1022,"_type":111,"children":1023,"markDefs":1036,"style":119},"f6a337aac2e4",[1024,1028,1032],{"_key":1025,"_type":115,"marks":1026,"text":1027},"c4ac34956e2d",[],"This is why there's a third viable configuration beyond \"frontier lab\" or \"API wrapper\" in certain verticals. Companies that start from open-weight foundations and build domain capability through heavy post-training tied to actual deployment. Frontier open-weight models are close enough on broad benchmarks that the foundation is good, trailing closed models by roughly three to six months on average",{"_key":1029,"_type":115,"marks":1030,"text":1031},"5640c5a6a068",[951],"3,4",{"_key":1033,"_type":115,"marks":1034,"text":1035},"e78baf9709da",[],", but lack domain-specific capability. What they offer is the ability to train, push domain knowledge into the weights and build RL environments around real workflows to create a loop where deployment feedback becomes training signal. The compute is substantial, hundreds of top-end GPUs, but not 10,000+. Not frontier lab scale.",[],{"_key":1038,"_type":223,"caption":1039,"image":1040},"f40d9bf99507","Figure 2. Time to train DeepSeek-V3[5] on 100B tokens at 32k context (H200 GPUs). Shaded region indicates vertical lab scale (128-1024 H200 equivalent GPUs). At 384 GPUs, supervised fine-tuning completes in under two weeks. Modelling RL as SFT plus on-policy generation of equal volume (as in GRPO), full iterations complete in approximately one month.",{"_type":70,"asset":1041},{"_ref":1042,"_type":73},"image-8f3889eeb92e7d98b5e2165282b349baf64d92aa-902x572-png",{"_key":1044,"_type":230,"text":1045},"7dec0a2184d4",[1046],{"_key":1047,"_type":111,"children":1048,"markDefs":1053,"style":119},"8f20ccb04ce5",[1049],{"_key":1050,"_type":115,"marks":1051,"text":1052},"28f103acf552",[],"What makes this defensible is how deployment and training reinforce each other. A model deployed in a real clinical environment generates feedback that doesn't exist in any public dataset and shapes the next training iteration as environment design. Performance improves, deeper deployment becomes possible, richer feedback follows and the loop compounds. But it compounds under constraints because data must be heavily protected, labels are expensive, site variance is a concern, and continuous learning collides with change control. You don't get to ship weekly behaviour changes into regulated clinical pathways without a validation story and a as a result the flywheel is overall slower and more disciplined than in consumer software.",[],{"_key":1055,"_type":223,"caption":1056,"image":1057},"16abad836a45","Figure 3. The deployment-training flywheel. Models deployed in clinical environments generate feedback that does not exist in public datasets. This feedback shapes the training environment for the next iteration, improving capability and enabling deeper integration. The loop compounds over time, creating differentiated capability that is difficult to replicate without equivalent deployment depth.",{"_type":70,"asset":1058},{"_ref":1059,"_type":73},"image-872a54e494ff335b360c87cb90ab455dc717f8de-902x516-png",{"_key":1061,"_type":230,"text":1062},"0645c1ee1129",[1063,1071,1078,1085,1100,1123,1130,1137,1144,1151],{"_key":1064,"_type":111,"children":1065,"markDefs":1070,"style":119},"c6491f8826c2",[1066],{"_key":1067,"_type":115,"marks":1068,"text":1069},"be420e6ddc2a",[],"Two companies can start from the same open-weight base and, after enough iterations, end up with meaningfully different systems. The moat becomes less about the base model and more about the RL environment, the domain data pipelines, the regulatory approvals and trust relationships that took years to build.",[],{"_key":1072,"_type":111,"children":1073,"markDefs":1077,"style":119},"2841c95655f6",[1074],{"_key":1067,"_type":115,"marks":1075,"text":1076},[],"There's also a capability barrier that's easy to underestimate. Full fine-tuning and RL at scale requires deep expertise in training, evaluation design, data & systems engineering, and the operational machinery to iterate quickly. The gap between \"we fine-tune models\" and \"we train at frontier open-weight scale within clinical RL environments\" is large. Most companies that want to be in this category can't assemble the technical capability to execute it.",[],{"_key":1079,"_type":111,"children":1080,"markDefs":1084,"style":119},"724cd1459674",[1081],{"_key":1067,"_type":115,"marks":1082,"text":1083},[],"Frontier labs can train powerful general models and enter any domain they choose. But they are unlikely to match the deployment depth of a focused vertical player across hundreds of regulated domains simultaneously.",[],{"_key":1086,"_type":111,"children":1087,"markDefs":1099,"style":119},"999658055dc2",[1088,1091,1095],{"_key":1067,"_type":115,"marks":1089,"text":1090},[],"Cohere provides some validation here: their CEO Aidan Gomez reported that revenue more than doubled in 2025, with The Information reporting $200 million in projected annualized revenue by mid-year, and a $500 million funding round at $7 billion valuation followed shortly after",{"_key":1092,"_type":115,"marks":1093,"text":1094},"9039188f76e3",[951],"6,7",{"_key":1096,"_type":115,"marks":1097,"text":1098},"2e3df07b0721",[],". Their pivot was explicit, away from foundation model scale, toward tailored enterprise models across several industries. They're horizontal rather than vertical, but the economics work.",[],{"_key":1101,"_type":111,"children":1102,"markDefs":1122,"style":119},"cda6031033fe",[1103,1106,1110,1114,1118],{"_key":1096,"_type":115,"marks":1104,"text":1105},[],"Cursor makes the case more sharply. They crossed $1 billion in annualized revenue in 2025 and raised at a $29.3 billion valuation",{"_key":1107,"_type":115,"marks":1108,"text":1109},"e5107a77c3dd",[951],"8,9",{"_key":1111,"_type":115,"marks":1112,"text":1113},"7a45fa4549e2",[],". They built Composer, their own model trained through reinforcement learning inside real coding environments, co-designed with the IDE, trained on hundreds of thousands of concurrent sandboxes, optimized for the specific feedback loops of software development. The result is frontier-level capability at four times the speed. They're vertical where Cohere is horizontal, but the logic is the same. Healthcare is where we're testing the thesis. It's one of the largest verticals for specialized AI investment, with spending reaching $1.5 billion in 2025",{"_key":1115,"_type":115,"marks":1116,"text":1117},"88033d919669",[951],"10",{"_key":1119,"_type":115,"marks":1120,"text":1121},"f089d20c7ced",[],".",[],{"_key":1124,"_type":111,"children":1125,"markDefs":1129,"style":119},"615db4f70519",[1126],{"_key":1119,"_type":115,"marks":1127,"text":1128},[],"For European healthcare systems there's also the question whether core clinical reasoning should run on models controlled by US providers, under different legal frameworks. For some institutions, the ability to audit and continuously validate the full system is a requirement that determines what they can deploy at all.",[],{"_key":1131,"_type":111,"children":1132,"markDefs":1136,"style":119},"299b33003473",[1133],{"_key":1067,"_type":115,"marks":1134,"text":1135},[],"kaiko.ai is building on this thesis: a multimodal assistant for clinical experts that integrates pathology, radiology, genomics, and clinical text, alongside the execution environment, integrations & safety mechanisms that lets an agent function in a clinic. We're training from frontier-grade open foundations, which gives us commoditized base capabilities plus model access to close domain-specific gaps. Our work on Vortex-1, a frontier pathology VLM releasing in the coming weeks, suggests competitive performance is achievable through domain-focused training choices. This is our bet: that clinical capability requires clinical training, that the deployment-training loop creates compounding advantage and that the jagged capability profile in frontier models reflects something structural about how the current systems learn rather than a temporary gap that scale will automatically close.",[],{"_key":1138,"_type":111,"children":1139,"markDefs":1143,"style":119},"300235f5cf10",[1140],{"_key":1067,"_type":115,"marks":1141,"text":1142},[],"We might be wrong. If we discover a better paradigm with more favourable generalisation properties compared to today's LLM stack, the jaggedness might smooth out faster than we expect. But if we're right about how capability develops in the mid-term, the winners in healthcare AI will be whoever closes clinical gaps through clinical deployment and training.",[],{"_key":1145,"_type":111,"children":1146,"markDefs":1150,"style":119},"53f83d38af11",[1147],{"_key":1067,"_type":115,"marks":1148,"text":1149},[],"The AI market will consequently start to fragment more. General models will keep improving and serve broad applications well. But in verticals where feedback loops require deployment depth and error costs create real stakes, specialized training will produce significantly differentiated capability & real self-reinforcing moats.",[],{"_key":1152,"_type":111,"children":1153,"markDefs":1157,"style":119},"7e679df58965",[1154],{"_key":1067,"_type":115,"marks":1155,"text":1156},[],"We're betting against the machine God, against the view that a single general system will do everything well and that training environment stops mattering at sufficient scale.",[],{"_key":1159,"_type":230,"text":1160},"e94862590617",[1161],{"_key":1162,"_type":111,"children":1163,"markDefs":1168,"style":884},"08b218f15cde",[1164],{"_key":1165,"_type":115,"marks":1166,"text":1167},"6a233d2c68aa",[],"Our systems are currently developed and tested in research and validation settings. Some of our products focus on supporting administrative and workflow processes and are not medical devices. Any future clinical decision support use would be subject to regulatory approval and compliance with applicable medical device frameworks, including the EU Medical Device Regulation (MDR).",[],{"_key":1170,"_type":1171,"references":1172},"249ec668f259","referencesSection",[1173,1174,1175,1176,1177,1178,1179,1180,1181,1182],"Noah Smith, \"What if AI succeeds but OpenAI fails?\" (January 2025). https://www.noahpinion.blog/p/what-if-ai-succeeds-but-openai-fails ","Andrej Karpathy, \"Year in Review 2025\" (December 2025). https://karpathy.bearblog.dev/year-in-review-2025/","Epoch AI, \"Open-weights vs Closed-weights Models\" (2025). https://epoch.ai/data-insights/open-weights-vs-closed-weights-models","https://artificialanalysis.ai/models/open-source","DeepSeek-AI, \"DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models\" (December 2025). https://arxiv.org/abs/2512.02556","Bloomberg, \"Cohere Has Doubled Sales So Far This Year, CEO Aidan Gomez Says\" (May 2025). https://www.bloomberg.com/news/articles/2025-05-20/ai-startup-cohere-has-doubled-sales-so-far-this-year-ceo-says","The Information, \"AI Startup Cohere Projects $200 Million Revenue Pace as New Funding Nears\" (July 2025). https://www.theinformation.com/articles/ai-startup-cohere-projects-200-million-revenue-pace-new-funding-nears","CNBC, \"AI startup Cursor raises $2.3 billion funding round at $29.3 billion valuation\" (November 2025). https://www.cnbc.com/2025/11/13/cursor-ai-startup-funding-round-valuation.html","Cursor, \"Composer: Building a fast frontier model with RL\" (October 2025). https://cursor.com/blog/composer","Menlo Ventures, \"2025: The State of Generative AI in the Enterprise\" (2025). https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/","2026-02-18T13:28:56.322Z",{"_type":70,"asset":1185},{"_ref":987,"_type":73},{"_type":909,"current":1187},"betting-Machine-God","Why specialized AI training beats waiting for general intelligence, and what it means for healthcare.","Betting against the Machine God ",[1191,1194,1197],{"_key":1192,"_ref":1193,"_type":73},"d7bd44f3f726","0e7a76c9-607b-4b29-b0f5-48940b955680",{"_key":1195,"_ref":1196,"_type":73},"fc4e403e40d5","2846bb95-e33f-42f0-b0d3-005c3ccf71d1",{"_key":1198,"_ref":1199,"_type":73},"93ebf4ec7d28","c76081b0-8b05-4804-8c56-cec432389183","CoralBay: The 3D Foundation Model for Radiology CT-Scans by kaiko",[1202,1205,1208],{"_key":1203,"_ref":1204,"_type":73},"ab8ee8167c09","87071b12-f012-4898-9928-e78a41e6f88e",{"_key":1206,"_ref":1207,"_type":73},"8e1606b99a64","ffc879d7-9be3-47d1-bda5-5817c9d7cf8c",{"_key":1209,"_ref":1210,"_type":73},"bdc5cc8677ce","fd4ff9cb-2a4f-41a0-a2af-371ea8baafe4",{"documents":1212,"paths":1215,"mappings":1227},[1213,1214],{"_id":203,"_type":208},{"_id":907,"_type":208},[126,127,128,129,130,131,1216,1217,1218,1219,1220,1221,1222,1223,1224,1225,1226],"$['area']","$['authors']","$['categories']","$['content']","$['date']","$['featuredImage']","$['relatedPosts']","$['slug']","$['subHeading']","$['title']","$['topics']",{"$['_createdAt']":1228,"$['_id']":1230,"$['_rev']":1232,"$['_system']":1234,"$['_type']":1236,"$['_updatedAt']":1238,"$['area']":1240,"$['authors']":1242,"$['categories']":1244,"$['content']":1246,"$['date']":1248,"$['featuredImage']":1250,"$['relatedPosts']":1252,"$['slug']":1254,"$['subHeading']":1256,"$['suggestedPosts'][0]['categories']":1258,"$['suggestedPosts'][0]['content']":1260,"$['suggestedPosts'][0]['date']":1262,"$['suggestedPosts'][0]['featuredImage']":1264,"$['suggestedPosts'][0]['slug']":1266,"$['suggestedPosts'][0]['subHeading']":1268,"$['suggestedPosts'][0]['title']":1270,"$['suggestedPosts'][0]['topics']":1272,"$['title']":1274,"$['topics']":1276},{"source":1229,"type":149},{"document":147,"path":147,"type":148},{"source":1231,"type":149},{"document":147,"path":82,"type":148},{"source":1233,"type":149},{"document":147,"path":154,"type":148},{"source":1235,"type":149},{"document":147,"path":157,"type":148},{"source":1237,"type":149},{"document":147,"path":160,"type":148},{"source":1239,"type":149},{"document":147,"path":163,"type":148},{"source":1241,"type":149},{"document":147,"path":166,"type":148},{"source":1243,"type":149},{"document":147,"path":169,"type":148},{"source":1245,"type":149},{"document":147,"path":172,"type":148},{"source":1247,"type":149},{"document":147,"path":175,"type":148},{"source":1249,"type":149},{"document":147,"path":178,"type":148},{"source":1251,"type":149},{"document":147,"path":181,"type":148},{"source":1253,"type":149},{"document":147,"path":184,"type":148},{"source":1255,"type":149},{"document":147,"path":187,"type":148},{"source":1257,"type":149},{"document":147,"path":190,"type":148},{"source":1259,"type":149},{"document":82,"path":172,"type":148},{"source":1261,"type":149},{"document":82,"path":175,"type":148},{"source":1263,"type":149},{"document":82,"path":178,"type":148},{"source":1265,"type":149},{"document":82,"path":181,"type":148},{"source":1267,"type":149},{"document":82,"path":187,"type":148},{"source":1269,"type":149},{"document":82,"path":190,"type":148},{"source":1271,"type":149},{"document":82,"path":193,"type":148},{"source":1273,"type":149},{"document":82,"path":199,"type":148},{"source":1275,"type":149},{"document":147,"path":193,"type":148},{"source":1277,"type":149},{"document":147,"path":199,"type":148},1781083461884]