{"ontology":"https://layer.re/machine/schema.json","epistemic_model":"evidence-backed-temporal-layer-graph","canonical":true,"type":"civilization_snapshot","snapshot":{"id":"civ-2022q2-reconstructed","as_of":"2022-06-30 23:59:59.000","summary":"By the end of Q2 2022, the AI stack had broadened materially beyond the Q1 picture of language models, prompting and specialized compute. PaLM and Cloud TPU v4 strengthened the coupling among models, distributed-systems software, accelerators and high-bandwidth networks. Flamingo and DALL·E 2 expanded multimodal capability and real-user generative workflows. OPT-175B reduced the frontier-model research-access bottleneck but exposed licensing and compute as remaining constraints. GitHub Copilot moved model-generated code from research demonstration into a broadly available developer product. SayCan and Gato provided the strongest Q2 evidence for agent-like and embodied action capabilities, but not for persistent autonomous agents as an independent product layer. Governance and geopolitics moved closer to the technical stack through transatlantic AI-risk, semiconductor, standards and export-control coordination plus private model-release controls. Energy became an explicit carbon/efficiency metric, not yet an evidenced binding AI power bottleneck.","cross_layer_changes":["Distributed-systems software and interconnect became more explicit determinants of frontier model scaling.","Multimodal prompting and generation broadened the model capability surface.","Large-model research access moved closer to frontier scale while remaining controlled by licensing, approval and compute.","Models entered mainstream software-development workflows through Copilot general availability.","Embodied/model-action research strengthened an Agents precursor without establishing persistent agents.","Governance increasingly addressed specific model access, evaluation, deployment safety and semiconductor coordination.","Energy/carbon appeared as an infrastructure metric without yet becoming a demonstrated capacity bottleneck."],"dependencies":["chips + networks + software -> compute","compute + research -> models","models -> software workflows","models + research -> robotics/agent-like action","governance -> AI/models/chip access","geopolitics + governance -> semiconductor and technology access"],"constraints":["frontier compute cost and access","distributed training/interconnect complexity","model reliability and capability predictability","controlled model licensing and release policy","robot grounding and skill coverage","safety and misuse in real-user deployment"],"emerging_capabilities":["multi-pod 540B language-model training","few-shot multimodal visual-language interaction","large-scale text-to-image user workflows","AI pair programming at general availability","language-model-guided robotic planning","generalist multi-embodiment action policies","stronger quantitative reasoning through model/data/prompt/inference composition"],"possible_bottlenecks":["accelerator availability and serving economics","network/interconnect efficiency","frontier model access and reproducibility","model-output verification","grounding and reliability for agent-like action"],"uncertainties":["Whether SayCan/Gato represented the beginning of a distinct Agents layer or remained robotics/model research prototypes.","Whether apparent emergent abilities reflected genuinely discontinuous capability changes or evaluation/metric effects.","Whether controlled frontier-model releases would meaningfully decentralize frontier research.","Energy was visible as carbon intensity and siting, but the selected Q2 sources do not justify calling grid power a binding AI bottleneck.","Vendor-reported usage and performance should not be treated as independent proof of durable productivity or deployment reliability."],"mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.770Z","updated_at":"2026-10-01T19:22:17.770Z"},"claims":[{"id":"claim-q2-agent-precursor","statement":"SayCan and Gato materially strengthened evidence for an agent-like model capability in Q2, but do not establish persistent autonomous Agents as an independent operational layer.","epistemic_status":"supported_inference","confidence":0.91,"civilization_snapshot_id":"civ-2022q2-reconstructed","layer_state_snapshot_id":"state-agents-2022q2","competing_interpretation":"A broader definition of agent could classify Gato and SayCan as genuine agents already, especially in embodied environments.","uncertainty":"Layer.re's later Agents concept additionally emphasizes persistence, tools, delegated execution, permissions and long-horizon reliability, which Q2 evidence does not establish.","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.772Z","updated_at":"2026-10-01T19:22:17.772Z"},{"id":"claim-q2-energy-not-yet-bottleneck","statement":"Q2 makes energy and carbon intensity visible in AI infrastructure decisions, but the selected evidence does not support a claim that electricity availability had become a binding AI compute bottleneck.","epistemic_status":"supported_inference","confidence":0.84,"civilization_snapshot_id":"civ-2022q2-reconstructed","layer_state_snapshot_id":"state-energy-2022q2","competing_interpretation":"Large data centers were inherently energy-dependent, so one could treat power as a latent constraint even before public AI-specific scarcity evidence became prominent.","uncertainty":"This reconstruction requires material evidence, not latent physical necessity, before marking a bottleneck as active.","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.773Z","updated_at":"2026-10-01T19:22:17.773Z"},{"id":"claim-q2-governance-closer","statement":"Q2 governance moved closer to concrete technical stack decisions through coordinated AI-risk measurement, semiconductor/export-control policy and controlled model deployment/release.","epistemic_status":"supported_inference","confidence":0.94,"civilization_snapshot_id":"civ-2022q2-reconstructed","layer_state_snapshot_id":"state-governance-2022q2","competing_interpretation":"Many public measures were still nonbinding workstreams, while company release policies were private product governance rather than public regulation.","uncertainty":"Durable standards and enforceable AI rules had not yet emerged from these Q2 processes.","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.772Z","updated_at":"2026-10-01T19:22:17.772Z"},{"id":"claim-q2-human-workflows","statement":"Q2 broadened direct human use of generative models through creative image workflows and generally available AI-assisted coding before a persistent-agent product category existed.","epistemic_status":"supported_inference","confidence":0.95,"civilization_snapshot_id":"civ-2022q2-reconstructed","layer_state_snapshot_id":"state-human-2022q2","competing_interpretation":"DALL·E remained a selected research preview and Copilot targeted developers, so these were still bounded user populations rather than broad society-wide adoption.","uncertainty":"Durability, distribution and labor effects remained unclear.","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.773Z","updated_at":"2026-10-01T19:22:17.773Z"},{"id":"claim-q2-models-software-operational","statement":"The models-to-software relationship crossed an operational adoption threshold in Q2 as GitHub Copilot became generally available inside developer editors.","epistemic_status":"supported_inference","confidence":0.97,"civilization_snapshot_id":"civ-2022q2-reconstructed","layer_state_snapshot_id":"state-software-2022q2","competing_interpretation":"Copilot remained assistive autocomplete rather than autonomous software engineering, so the change may be better described as workflow augmentation than a structural software-layer shift.","uncertainty":"Independent productivity, correctness and long-term adoption effects were not yet established.","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.772Z","updated_at":"2026-10-01T19:22:17.772Z"},{"id":"claim-q2-open-access-shift","statement":"OPT-175B reduced the frontier-model research-access barrier, but did not remove it; access shifted toward licensing, approval, inference cost and reproduction compute.","epistemic_status":"supported_inference","confidence":0.95,"civilization_snapshot_id":"civ-2022q2-reconstructed","layer_state_snapshot_id":"state-open-source-2022q2","competing_interpretation":"Controlled research access under a noncommercial license may be too restrictive to count as a major open-model transition.","uncertainty":"Actual breadth and downstream impact of access were still developing during Q2.","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.771Z","updated_at":"2026-10-01T19:22:17.771Z"},{"id":"claim-q2-systems-scaling","statement":"Q2 2022 materially strengthened the dependency of frontier model scaling on distributed-systems software and high-bandwidth accelerator networking.","epistemic_status":"supported_inference","confidence":0.96,"civilization_snapshot_id":"civ-2022q2-reconstructed","layer_state_snapshot_id":"state-compute-2022q2","competing_interpretation":"PaLM and TPU v4 may represent Google-specific architecture choices rather than a universal requirement for every model family.","uncertainty":"The relative importance of network, compiler/runtime and raw accelerator supply varied by workload and provider.","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.771Z","updated_at":"2026-10-01T19:22:17.771Z"}],"layer_states":[{"id":"state-ai-2022q2","layer_id":"layer-ai","as_of":"2022-06-30 23:59:59.000","summary":"By end-Q2, AI remained model-centric but was no longer primarily a text-and-code story: multimodal models, creative generation, coding assistance and embodied generalist prototypes were visibly coupling models to more interfaces and environments.","dependencies":["layer-models","layer-research","layer-compute","layer-software","layer-human"],"constraints":["reliability","compute and serving cost","release safety","capability predictability"],"emerging_capabilities":["multimodal few-shot interaction","large-scale generative image workflows","AI-assisted coding","agent-like embodied action research"],"underdeveloped_components":["persistent autonomous agents","reliable open-ended tool use","stable multimodal deployment guarantees"],"uncertainties":["whether research prototypes would generalize into durable products","how much threshold-like capability behavior reflected metrics versus underlying discontinuities"],"confidence":0.94,"previous_snapshot_id":"state-ai-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.761Z","updated_at":"2026-10-01T19:22:17.761Z","layer_slug":"ai","layer_name":"AI"},{"id":"state-agents-2022q2","layer_id":"layer-agents","as_of":"2022-06-30 23:59:59.000","summary":"By end-Q2, agent-like capability was materially more visible in research: models could participate in planning and action loops across embodied or interactive environments. The layer was still a precursor/research capability rather than a persistent deployed agent category.","dependencies":["layer-models","layer-research","layer-software","layer-robotics"],"constraints":["durable memory","general tool use","reliable long-horizon execution","persistent identity","delegation and permissions"],"emerging_capabilities":["language-model-guided robotic planning","multi-task generalist action policies"],"underdeveloped_components":["persistent background autonomy","broad external-tool orchestration","production agent ecosystems"],"uncertainties":["the boundary between 'model policy', 'robot controller' and the later Agents layer remains conceptually ambiguous"],"confidence":0.88,"previous_snapshot_id":"state-agents-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.766Z","updated_at":"2026-10-01T19:22:17.766Z","layer_slug":"agents","layer_name":"Agents"},{"id":"state-chips-2022q2","layer_id":"layer-chips","as_of":"2022-06-30 23:59:59.000","summary":"AI accelerators were increasingly experienced as components of integrated pods and cloud systems rather than standalone chips, while semiconductor access remained a strategic governance concern.","dependencies":["layer-manufacturing","layer-networks","layer-software","layer-governance","layer-geopolitics"],"constraints":["supply concentration","system-level scaling","specialized software dependence"],"emerging_capabilities":["publicly accessible TPU v4 pod infrastructure"],"underdeveloped_components":["evidence of broad supply abundance","cross-vendor portability"],"uncertainties":["Q2 evidence is stronger on system availability than on manufacturing-capacity change"],"confidence":0.9,"previous_snapshot_id":"state-chips-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.763Z","updated_at":"2026-10-01T19:22:17.763Z","layer_slug":"chips","layer_name":"Chips"},{"id":"state-compute-2022q2","layer_id":"layer-compute","as_of":"2022-06-30 23:59:59.000","summary":"Compute in Q2 matured from large internal clusters toward orchestrated multi-pod systems and cloud-accessible frontier infrastructure, with software scheduling, network bandwidth and cost becoming increasingly explicit parts of the compute layer.","dependencies":["layer-chips","layer-networks","layer-software","layer-energy"],"constraints":["scale and price-performance","interconnect efficiency","frontier access concentration","capital intensity"],"emerging_capabilities":["multi-pod language-model training","public preview access to thousands of TPU v4 accelerators"],"underdeveloped_components":["broad inexpensive access","evidence of unconstrained accelerator supply"],"uncertainties":["vendor peak-performance figures versus sustained workload economics"],"confidence":0.97,"previous_snapshot_id":"state-compute-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.762Z","updated_at":"2026-10-01T19:22:17.762Z","layer_slug":"compute","layer_name":"Compute"},{"id":"state-energy-2022q2","layer_id":"layer-energy","as_of":"2022-06-30 23:59:59.000","summary":"Energy became visible in the reconstructed AI infrastructure stack as a sustainability, efficiency and siting metric, but the selected Q2 evidence still does not establish power availability as a binding compute-capacity bottleneck.","dependencies":["electricity supply","layer-compute","data-center siting"],"constraints":["carbon intensity","efficiency","incomplete evidence on grid capacity"],"emerging_capabilities":["explicit energy/carbon reporting for large-model infrastructure"],"underdeveloped_components":["evidence of power scarcity constraining Q2 AI deployment"],"uncertainties":["later AI power bottlenecks must not be projected back onto this quarter"],"confidence":0.8,"previous_snapshot_id":null,"mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.769Z","updated_at":"2026-10-01T19:22:17.769Z","layer_slug":"energy","layer_name":"Energy"},{"id":"state-geopolitics-2022q2","layer_id":"layer-geopolitics","as_of":"2022-06-30 23:59:59.000","summary":"Technology geopolitics became more coordinated across allied institutions, with semiconductors, export controls, standards and trustworthy AI increasingly discussed inside one policy machinery.","dependencies":["layer-governance","layer-chips","layer-manufacturing","layer-ai"],"constraints":["war-driven export restrictions","supply-chain concentration","standards competition"],"emerging_capabilities":["cross-border semiconductor early-warning cooperation","coordinated technology-control policy"],"underdeveloped_components":["stable global technology regimes"],"uncertainties":["coordination commitments did not guarantee later policy convergence"],"confidence":0.94,"previous_snapshot_id":"state-geopolitics-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.765Z","updated_at":"2026-10-01T19:22:17.765Z","layer_slug":"geopolitics","layer_name":"Geopolitics"},{"id":"state-governance-2022q2","layer_id":"layer-governance","as_of":"2022-06-30 23:59:59.000","summary":"Governance moved closer to specific technical deployment decisions: who could access large models, how preview safety was monitored, how AI risk would be measured and how semiconductor supply risks would be coordinated internationally.","dependencies":["layer-ai","layer-models","layer-chips","layer-geopolitics","institutional capacity"],"constraints":["nonbinding frameworks","cross-jurisdiction coordination","limited evaluation standards","deployment safety uncertainty"],"emerging_capabilities":["joint AI-risk measurement workstreams","controlled model release","iterative preview safety systems"],"underdeveloped_components":["binding general AI rules","mature agent permissions/liability"],"uncertainties":["many TTC outputs remained work programmes rather than enforceable rules"],"confidence":0.94,"previous_snapshot_id":"state-governance-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.765Z","updated_at":"2026-10-01T19:22:17.765Z","layer_slug":"governance","layer_name":"Governance"},{"id":"state-human-2022q2","layer_id":"layer-human","as_of":"2022-06-30 23:59:59.000","summary":"Humans became more visibly downstream users and workflow partners of generative models, not only supervisors or labelers upstream of training.","dependencies":["layer-models","layer-software","layer-governance"],"constraints":["trust","output correctness","misuse","workflow adaptation"],"emerging_capabilities":["creative text-to-image workflows","AI pair programming"],"underdeveloped_components":["clear labor-productivity evidence","stable social expectations for AI-generated work"],"uncertainties":["vendor adoption metrics may not represent durable use or net benefit"],"confidence":0.93,"previous_snapshot_id":"state-human-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.766Z","updated_at":"2026-10-01T19:22:17.766Z","layer_slug":"human","layer_name":"Human"},{"id":"state-models-2022q2","layer_id":"layer-models","as_of":"2022-06-30 23:59:59.000","summary":"Large models in Q2 became more general across tasks and modalities, while capability increasingly depended on training system scale, specialized data, prompt/inference strategies and controlled deployment rather than parameter count alone.","dependencies":["layer-research","layer-compute","layer-chips","layer-software","layer-human"],"constraints":["training and inference cost","unreliable reasoning","limited interpretability","capability forecasting uncertainty"],"emerging_capabilities":["multimodal few-shot learning","stronger quantitative reasoning","broader code and language capability"],"underdeveloped_components":["verified reasoning","predictable scaling behavior","broad low-cost frontier inference"],"uncertainties":["benchmark performance did not establish robust real-world generality","emergence interpretations were method-sensitive"],"confidence":0.96,"previous_snapshot_id":"state-models-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.761Z","updated_at":"2026-10-01T19:22:17.761Z","layer_slug":"models","layer_name":"Models"},{"id":"state-networks-2022q2","layer_id":"layer-networks","as_of":"2022-06-30 23:59:59.000","summary":"Networking remained an infrastructure dependency but became more visibly co-designed with model parallelism, accelerators and orchestration software at multi-pod scale.","dependencies":["layer-chips","layer-compute","layer-software"],"constraints":["bandwidth","latency","distributed coordination overhead"],"emerging_capabilities":["multi-pod language-model training","high-bandwidth TPU v4 pods"],"underdeveloped_components":["portable large-scale interconnect abstractions"],"uncertainties":["vendor-specific comparisons do not establish universal architecture superiority"],"confidence":0.92,"previous_snapshot_id":"state-networks-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.763Z","updated_at":"2026-10-01T19:22:17.763Z","layer_slug":"networks","layer_name":"Networks"},{"id":"state-open-source-2022q2","layer_id":"layer-open-source","as_of":"2022-06-30 23:59:59.000","summary":"Open-model dynamics moved much closer to frontier scale in Q2, while 'openness' itself became visibly constrained by licensing, controlled access, compute cost and responsible-release policy.","dependencies":["layer-software","layer-research","layer-compute","layer-governance"],"constraints":["controlled licensing","serving cost","training reproducibility","compute concentration"],"emerging_capabilities":["frontier-scale weight access for approved research","training logbooks and code release"],"underdeveloped_components":["unrestricted frontier model access"],"uncertainties":["whether controlled research access should be classified as open source in the strict software sense"],"confidence":0.89,"previous_snapshot_id":"state-open-source-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.764Z","updated_at":"2026-10-01T19:22:17.764Z","layer_slug":"open-source","layer_name":"Open source"},{"id":"state-research-2022q2","layer_id":"layer-research","as_of":"2022-06-30 23:59:59.000","summary":"Research became more tightly coupled to system engineering and deployment surfaces: papers increasingly described complete model+compute+interface or model+environment systems rather than isolated algorithmic improvements.","dependencies":["layer-models","layer-compute","layer-software","layer-robotics"],"constraints":["compute concentration","reproducibility","benchmark validity","controlled access to frontier models"],"emerging_capabilities":["generalist policy research","multimodal prompting","quantitative reasoning systems"],"underdeveloped_components":["independent replication of frontier-scale results","standard evaluation of agent-like behavior"],"uncertainties":["frontier results remained concentrated in a small number of labs"],"confidence":0.95,"previous_snapshot_id":"state-research-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.762Z","updated_at":"2026-10-01T19:22:17.762Z","layer_slug":"research","layer_name":"Research"},{"id":"state-robotics-2022q2","layer_id":"layer-robotics","as_of":"2022-06-30 23:59:59.000","summary":"Robotics became newly visible in this historical slice as an integration surface for large pretrained models: semantic model knowledge could be grounded through skills, affordances and action policies in physical environments.","dependencies":["layer-research","layer-models","layer-software"],"constraints":["grounding","safety","skill coverage","real-world robustness"],"emerging_capabilities":["language-guided long-horizon task execution","generalist transformer control across embodiments"],"underdeveloped_components":["open-ended physical autonomy","durable real-world learning"],"uncertainties":["robotics itself long predates Q2; this is first material relevance to this reconstructed large-model stack, not birth of robotics"],"confidence":0.9,"previous_snapshot_id":null,"mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.768Z","updated_at":"2026-10-01T19:22:17.768Z","layer_slug":"robotics","layer_name":"Robotics"},{"id":"state-software-2022q2","layer_id":"layer-software","as_of":"2022-06-30 23:59:59.000","summary":"Software became both an upstream enabler of large distributed AI systems and a downstream domain directly transformed by deployed model assistance.","dependencies":["layer-models","layer-compute","layer-chips","layer-open-source","layer-human"],"constraints":["model reliability","hardware-specific scaling","generated-code correctness"],"emerging_capabilities":["AI pair programming","multi-pod accelerator orchestration","multimodal prompt interfaces"],"underdeveloped_components":["autonomous software engineering","cross-hardware orchestration standards"],"uncertainties":["vendor-reported adoption does not establish net productivity or software quality"],"confidence":0.96,"previous_snapshot_id":"state-software-2022q1","mapping_mode":"retrospective","reconstructed_at":"2026-10-01T18:14:00.000Z","created_at":"2026-10-01T19:22:17.764Z","updated_at":"2026-10-01T19:22:17.764Z","layer_slug":"software","layer_name":"Software"}]}