{"edges":[{"from":"events/alphago-fan-hui-2016","relation":"challenges","to":"concepts/go-complexity-as-ai-barrier"},{"from":"events/alphago-fan-hui-2016","relation":"enables","to":"events/alphafold-2020"},{"from":"events/alphago-fan-hui-2016","relation":"introduces","to":"concepts/deep-rl-for-strategic-games"},{"from":"events/transformer-architecture-2017","relation":"enables","to":"events/chatgpt-launch-2022"},{"from":"events/transformer-architecture-2017","relation":"enables","to":"events/scaling-laws-2020"},{"from":"events/transformer-architecture-2017","relation":"enables","to":"events/alphafold-2020"},{"from":"events/transformer-architecture-2017","relation":"enables","to":"events/dall-e-2021"},{"from":"events/transformer-architecture-2017","relation":"introduces","to":"concepts/transformer"},{"from":"events/transformer-architecture-2017","relation":"introduces","to":"concepts/self-attention"},{"from":"events/scaling-laws-2020","relation":"challenges","to":"concepts/sample-efficiency-as-primary-metric"},{"from":"events/scaling-laws-2020","relation":"enables","to":"events/chatgpt-launch-2022"},{"from":"events/scaling-laws-2020","relation":"enables","to":"events/openai-o1-released"},{"from":"events/scaling-laws-2020","relation":"enables","to":"events/gpt-5-2025"},{"from":"events/scaling-laws-2020","relation":"introduces","to":"concepts/scaling-laws"},{"from":"events/scaling-laws-2020","relation":"introduces","to":"concepts/compute-optimal-training"},{"from":"events/chatgpt-launch-2022","relation":"builds_on","to":"events/transformer-architecture-2017"},{"from":"events/chatgpt-launch-2022","relation":"builds_on","to":"events/scaling-laws-2020"},{"from":"events/chatgpt-launch-2022","relation":"enables","to":"events/openai-o1-released"},{"from":"events/chatgpt-launch-2022","relation":"enables","to":"events/gpt-5-2025"},{"from":"events/chatgpt-launch-2022","relation":"enables","to":"events/ai-agents-emerge-2025"},{"from":"events/chatgpt-launch-2022","relation":"introduces","to":"concepts/rlhf"},{"from":"events/claude-35-sonnet-2024","relation":"builds_on","to":"events/claude-3-opus-released"},{"from":"events/claude-35-sonnet-2024","relation":"builds_on","to":"events/transformer-architecture-2017"},{"from":"events/claude-35-sonnet-2024","relation":"enables","to":"events/claude-4-2025"},{"from":"events/claude-35-sonnet-2024","relation":"enables","to":"events/ai-agents-emerge-2025"},{"from":"events/inference-time-scaling-2024","relation":"builds_on","to":"events/scaling-laws-2020"},{"from":"events/inference-time-scaling-2024","relation":"builds_on","to":"events/openai-o1-released"},{"from":"events/inference-time-scaling-2024","relation":"challenges","to":"concepts/training-scale-as-only-axis"},{"from":"events/inference-time-scaling-2024","relation":"enables","to":"events/deepseek-r1-2025"},{"from":"events/inference-time-scaling-2024","relation":"enables","to":"events/gpt-5-2025"},{"from":"events/inference-time-scaling-2024","relation":"introduces","to":"concepts/inference-time-scaling"},{"from":"events/inference-time-scaling-2024","relation":"introduces","to":"concepts/test-time-compute"},{"from":"events/openai-o1-released","relation":"builds_on","to":"events/scaling-laws-2020"},{"from":"events/openai-o1-released","relation":"builds_on","to":"events/transformer-architecture-2017"},{"from":"events/openai-o1-released","relation":"enables","to":"events/gpt-5-2025"},{"from":"events/openai-o1-released","relation":"enables","to":"events/inference-time-scaling-2024"},{"from":"events/openai-o1-released","relation":"enables","to":"events/deepseek-r1-2025"},{"from":"events/openai-o1-released","relation":"introduces","to":"concepts/inference-time-scaling"},{"from":"events/gemini-2-flash-2024","relation":"builds_on","to":"events/gemini-1.5-pro-released"},{"from":"events/gemini-2-flash-2024","relation":"builds_on","to":"events/transformer-architecture-2017"},{"from":"events/gemini-2-flash-2024","relation":"enables","to":"events/gemini-2-5-pro-2025"},{"from":"events/gemini-2-flash-2024","relation":"enables","to":"events/ai-agents-emerge-2025"},{"from":"events/openai-o3-2024","relation":"builds_on","to":"events/openai-o1-released"},{"from":"events/openai-o3-2024","relation":"builds_on","to":"events/scaling-laws-2020"},{"from":"events/openai-o3-2024","relation":"enables","to":"events/gpt-5-2025"},{"from":"events/openai-o3-2024","relation":"enables","to":"events/deepseek-r1-2025"},{"from":"events/openai-o3-2024","relation":"enables","to":"events/inference-time-scaling-2024"},{"from":"events/deepseek-r1-2025","relation":"challenges","to":"concepts/compute-scale-as-moat"},{"from":"events/deepseek-r1-2025","relation":"challenges","to":"concepts/gpu-access-as-barrier"},{"from":"events/deepseek-r1-2025","relation":"enables","to":"events/ai-agents-emerge-2025"},{"from":"events/deepseek-r1-2025","relation":"introduces","to":"concepts/pure-rl-reasoning"},{"from":"events/deepseek-r1-2025","relation":"responds_to","to":"events/openai-o1-released"},{"from":"events/ai-agents-emerge-2025","relation":"builds_on","to":"events/openai-o1-released"},{"from":"events/ai-agents-emerge-2025","relation":"builds_on","to":"events/chatgpt-launch-2022"},{"from":"events/ai-agents-emerge-2025","relation":"enables","to":"events/gpt-5-2025"},{"from":"events/ai-agents-emerge-2025","relation":"introduces","to":"concepts/model-context-protocol"},{"from":"events/ai-agents-emerge-2025","relation":"introduces","to":"capabilities/autonomous-web-browsing"},{"from":"events/openai-o3-mini-2025","relation":"builds_on","to":"events/openai-o3-2024"},{"from":"events/openai-o3-mini-2025","relation":"builds_on","to":"events/openai-o1-released"},{"from":"events/openai-o3-mini-2025","relation":"enables","to":"events/gpt-5-2025"},{"from":"events/claude-37-sonnet-2025","relation":"builds_on","to":"events/claude-35-sonnet-2024"},{"from":"events/claude-37-sonnet-2025","relation":"builds_on","to":"events/openai-o1-released"},{"from":"events/claude-37-sonnet-2025","relation":"enables","to":"events/claude-4-2025"},{"from":"events/claude-37-sonnet-2025","relation":"enables","to":"events/ai-agents-emerge-2025"},{"from":"events/claude-4-2025","relation":"builds_on","to":"events/openai-o1-released"},{"from":"events/claude-4-2025","relation":"enables","to":"events/gpt-5-2025"},{"from":"events/claude-4-2025","relation":"introduces","to":"models/claude-4-opus"},{"from":"events/claude-4-2025","relation":"introduces","to":"capabilities/multi-hour-agentic-workflows"},{"from":"events/claude-4-2025","relation":"responds_to","to":"events/gemini-2-5-pro-2025"},{"from":"events/gpt-5-2025","relation":"builds_on","to":"events/openai-o1-released"},{"from":"events/gpt-5-2025","relation":"builds_on","to":"events/inference-time-scaling-2024"},{"from":"events/gpt-5-2025","relation":"introduces","to":"models/gpt-5"},{"from":"events/gpt-5-2025","relation":"responds_to","to":"events/gemini-2-5-pro-2025"},{"from":"events/gpt-5-2025","relation":"responds_to","to":"events/claude-4-2025"},{"from":"events/gpt-5-2025","relation":"responds_to","to":"events/deepseek-r1-2025"},{"from":"events/apple-google-gemini-partnership-2026","relation":"builds_on","to":"events/gpt-4o-released"},{"from":"events/apple-google-gemini-partnership-2026","relation":"builds_on","to":"events/gemini-2-5-pro-2025"},{"from":"events/apple-google-gemini-partnership-2026","relation":"challenges","to":"events/chatgpt-launch-2022"},{"from":"events/claude-opus-46-sonnet-46-2026","relation":"builds_on","to":"events/claude-4-2025"},{"from":"events/google-lyria-3-2026","relation":"builds_on","to":"events/gemini-2-5-pro-2025"},{"from":"events/gemini-31-flash-image-2026","relation":"builds_on","to":"events/gemini-2-5-pro-2025"},{"from":"events/gemini-31-flash-image-2026","relation":"builds_on","to":"events/gemini-2-flash-2024"},{"from":"events/mcp-open-governance-2026","relation":"builds_on","to":"events/ai-agents-emerge-2025"},{"from":"events/anthropic-mythos-2026","relation":"builds_on","to":"events/claude-4-2025"},{"from":"events/anthropic-mythos-2026","relation":"introduces","to":"models/mythos"},{"from":"events/anthropic-mythos-2026","relation":"responds_to","to":"events/openai-o3-2024"},{"from":"events/anthropic-mythos-2026","relation":"responds_to","to":"events/deepseek-r1-2025"},{"from":"events/openai-anthropic-google-ai-defense-2026","relation":"responds_to","to":"events/anthropic-mythos-2026"},{"from":"events/openai-anthropic-google-ai-defense-2026","relation":"responds_to","to":"events/deepseek-r1-2025"},{"from":"events/meta-muse-spark-2026","relation":"builds_on","to":"events/llama-4-2025"},{"from":"events/meta-muse-spark-2026","relation":"challenges","to":"events/gpt-5-2025"},{"from":"events/meta-muse-spark-2026","relation":"challenges","to":"events/claude-4-2025"},{"from":"events/meta-muse-spark-2026","relation":"challenges","to":"events/gemini-2-5-pro-2025"},{"from":"events/anthropic-claude-opus-47-2026","relation":"builds_on","to":"events/claude-4-2025"},{"from":"events/anthropic-claude-opus-47-2026","relation":"builds_on","to":"events/anthropic-mythos-2026"},{"from":"events/anthropic-claude-opus-47-2026","relation":"responds_to","to":"events/gpt-5-2025"},{"from":"events/anthropic-claude-opus-47-2026","relation":"responds_to","to":"events/gemini-2-5-pro-2025"}],"meta":{"attribution":"AI Chronicles · https://ciyuan.website · CC BY 4.0","description":"Semantic knowledge graph of AI history events, entities, and causal relationships. Designed for LLM context injection, RAG indexing, and knowledge graph applications.","generated_at":"2026-06-27","how_to_use":"Start with BENJI-tier events for the essential arc. Follow 'enables' edges forward in time; 'builds_on' backward. Filter by category for domain narratives. Each node's 'context' field contains historical preconditions; 'significance' contains editorial interpretation.","language":"en","license":"https://creativecommons.org/licenses/by/4.0/","namespace":"https://ciyuan.website/onto/","ontology_url":"/ontology.json","type":"AIHistoryKnowledgeGraph","version":"1.0"},"nodes":{"concepts":[{"id":"concepts/turing-test","label":"Turing Test","type":"AIConcept"},{"id":"concepts/computational-theory","label":"Computational Theory","type":"AIConcept"},{"id":"concepts/ai-foundations","label":"AI Foundations","type":"AIConcept"},{"id":"concepts/philosophy-of-mind","label":"Philosophy of Mind","type":"AIConcept"},{"id":"concepts/artificial-intelligence","label":"Artificial Intelligence","type":"AIConcept"},{"id":"concepts/symbolic-ai","label":"Symbolic AI","type":"AIConcept"},{"id":"concepts/knowledge-engineering","label":"Knowledge Engineering","type":"AIConcept"},{"id":"concepts/perceptron","label":"Perceptron","type":"AIConcept"},{"id":"concepts/neural-networks","label":"Neural Networks","type":"AIConcept"},{"id":"concepts/pattern-recognition","label":"Pattern Recognition","type":"AIConcept"},{"id":"concepts/machine-learning","label":"Machine Learning","type":"AIConcept"},{"id":"concepts/chatbot","label":"Chatbot","type":"AIConcept"},{"id":"concepts/natural-language-processing","label":"Natural Language Processing","type":"AIConcept"},{"id":"concepts/human-computer-interaction","label":"Human-Computer Interaction","type":"AIConcept"},{"id":"concepts/ai-winter","label":"AI Winter","type":"AIConcept"},{"id":"concepts/expert-systems","label":"Expert Systems","type":"AIConcept"},{"id":"concepts/rule-based-ai","label":"Rule-Based AI","type":"AIConcept"},{"id":"concepts/lisp-machines","label":"LISP Machines","type":"AIConcept"},{"id":"concepts/game-ai","label":"Game AI","type":"AIConcept"},{"id":"concepts/minimax-search","label":"Minimax Search","type":"AIConcept"},{"id":"concepts/chess-ai","label":"Chess AI","type":"AIConcept"},{"id":"concepts/long-short-term-memory","label":"Long Short-Term Memory","type":"AIConcept"},{"id":"concepts/recurrent-neural-networks","label":"Recurrent Neural Networks","type":"AIConcept"},{"id":"concepts/sequence-modeling","label":"Sequence Modeling","type":"AIConcept"},{"id":"concepts/deep-learning","label":"Deep Learning","type":"AIConcept"},{"id":"concepts/convolutional-neural-networks","label":"Convolutional Neural Networks","type":"AIConcept"},{"id":"concepts/gpu-computing","label":"GPU Computing","type":"AIConcept"},{"id":"concepts/imagenet","label":"ImageNet","type":"AIConcept"},{"id":"concepts/reinforcement-learning","label":"Reinforcement Learning","type":"AIConcept"},{"id":"concepts/monte-carlo-tree-search","label":"Monte Carlo Tree Search","type":"AIConcept"},{"id":"concepts/deep-neural-networks","label":"Deep Neural Networks","type":"AIConcept"},{"id":"concepts/self-play","label":"Self-play","type":"AIConcept"},{"id":"concepts/ai-safety","label":"AI Safety","type":"AIConcept"},{"id":"concepts/ai-ethics","label":"AI Ethics","type":"AIConcept"},{"id":"concepts/existential-risk","label":"Existential Risk","type":"AIConcept"},{"id":"concepts/beneficial-ai","label":"Beneficial AI","type":"AIConcept"},{"id":"concepts/transformer","label":"Transformer","type":"AIConcept"},{"id":"concepts/self-attention","label":"Self-Attention","type":"AIConcept"},{"id":"concepts/multi-head-attention","label":"Multi-head Attention","type":"AIConcept"},{"id":"concepts/positional-encoding","label":"Positional Encoding","type":"AIConcept"},{"id":"concepts/encoder-decoder-architecture","label":"Encoder-Decoder Architecture","type":"AIConcept"},{"id":"concepts/scaling-laws","label":"Scaling Laws","type":"AIConcept"},{"id":"concepts/power-laws","label":"Power Laws","type":"AIConcept"},{"id":"concepts/compute-optimal-training","label":"Compute-optimal Training","type":"AIConcept"},{"id":"concepts/emergent-abilities","label":"Emergent Abilities","type":"AIConcept"},{"id":"concepts/neural-language-model-scaling","label":"Neural Language Model Scaling","type":"AIConcept"},{"id":"concepts/protein-folding","label":"Protein Folding","type":"AIConcept"},{"id":"concepts/biology","label":"Biology","type":"AIConcept"},{"id":"concepts/drug-discovery","label":"Drug Discovery","type":"AIConcept"},{"id":"concepts/structural-biology","label":"Structural Biology","type":"AIConcept"},{"id":"concepts/text-to-image","label":"Text-to-Image","type":"AIConcept"},{"id":"concepts/generative-ai","label":"Generative AI","type":"AIConcept"},{"id":"concepts/multimodal-ai","label":"Multimodal AI","type":"AIConcept"},{"id":"concepts/diffusion-models","label":"Diffusion Models","type":"AIConcept"},{"id":"concepts/rlhf","label":"RLHF","type":"AIConcept"},{"id":"concepts/alignment","label":"Alignment","type":"AIConcept"},{"id":"concepts/instruction-following","label":"Instruction Following","type":"AIConcept"},{"id":"concepts/reinforcement-learning-from-human-feedback","label":"Reinforcement Learning from Human Feedback","type":"AIConcept"},{"id":"concepts/code-generation","label":"Code Generation","type":"AIConcept"},{"id":"concepts/ai-coding-assistants","label":"AI Coding Assistants","type":"AIConcept"},{"id":"concepts/developer-tools","label":"Developer Tools","type":"AIConcept"},{"id":"concepts/instruction-following","label":"Instruction-following","type":"AIConcept"},{"id":"concepts/conversational-ai","label":"Conversational AI","type":"AIConcept"},{"id":"concepts/large-language-model","label":"Large Language Model","type":"AIConcept"},{"id":"concepts/reasoning","label":"Reasoning","type":"AIConcept"},{"id":"concepts/ai-safety-summit","label":"AI Safety Summit","type":"AIConcept"},{"id":"concepts/frontier-ai","label":"Frontier AI","type":"AIConcept"},{"id":"concepts/international-ai-governance","label":"International AI Governance","type":"AIConcept"},{"id":"concepts/long-context","label":"Long Context","type":"AIConcept"},{"id":"concepts/native-multimodality","label":"Native Multimodality","type":"AIConcept"},{"id":"concepts/text-to-video","label":"Text-to-Video","type":"AIConcept"},{"id":"concepts/world-model","label":"World Model","type":"AIConcept"},{"id":"concepts/constitutional-ai","label":"Constitutional AI","type":"AIConcept"},{"id":"concepts/safety","label":"Safety","type":"AIConcept"},{"id":"concepts/ai-regulation","label":"AI Regulation","type":"AIConcept"},{"id":"concepts/risk-based-ai-governance","label":"Risk-Based AI Governance","type":"AIConcept"},{"id":"concepts/ai-compliance","label":"AI Compliance","type":"AIConcept"},{"id":"concepts/ai-law","label":"AI Law","type":"AIConcept"},{"id":"concepts/open-weights","label":"Open Weights","type":"AIConcept"},{"id":"concepts/open-source-ai","label":"Open Source AI","type":"AIConcept"},{"id":"concepts/native-voice","label":"Native Voice","type":"AIConcept"},{"id":"concepts/real-time-interaction","label":"Real-time Interaction","type":"AIConcept"},{"id":"concepts/coding-agents","label":"Coding Agents","type":"AIConcept"},{"id":"concepts/context-window","label":"Context Window","type":"AIConcept"},{"id":"concepts/long-context-reasoning","label":"Long Context Reasoning","type":"AIConcept"},{"id":"concepts/coding","label":"Coding","type":"AIConcept"},{"id":"concepts/inference-time-scaling","label":"Inference-time Scaling","type":"AIConcept"},{"id":"concepts/test-time-compute","label":"Test-time Compute","type":"AIConcept"},{"id":"concepts/chain-of-thought-reasoning","label":"Chain-of-thought Reasoning","type":"AIConcept"},{"id":"concepts/compute-scaling","label":"Compute Scaling","type":"AIConcept"},{"id":"concepts/native-tool-use","label":"Native Tool Use","type":"AIConcept"},{"id":"concepts/multi-agent-orchestration","label":"Multi-Agent Orchestration","type":"AIConcept"},{"id":"concepts/agentic-ai","label":"Agentic AI","type":"AIConcept"},{"id":"concepts/low-latency","label":"Low Latency","type":"AIConcept"},{"id":"concepts/function-calling","label":"Function Calling","type":"AIConcept"},{"id":"concepts/arc-agi","label":"ARC-AGI","type":"AIConcept"},{"id":"concepts/fluid-reasoning","label":"Fluid Reasoning","type":"AIConcept"},{"id":"concepts/agi","label":"AGI","type":"AIConcept"},{"id":"concepts/pure-reinforcement-learning","label":"Pure Reinforcement Learning","type":"AIConcept"},{"id":"concepts/mixture-of-experts","label":"Mixture of Experts","type":"AIConcept"},{"id":"concepts/compute-efficiency","label":"Compute Efficiency","type":"AIConcept"},{"id":"concepts/ai-deregulation","label":"AI Deregulation","type":"AIConcept"},{"id":"concepts/national-security","label":"National Security","type":"AIConcept"},{"id":"concepts/ai-policy","label":"AI Policy","type":"AIConcept"},{"id":"concepts/executive-order","label":"Executive Order","type":"AIConcept"},{"id":"concepts/model-context-protocol","label":"Model Context Protocol","type":"AIConcept"},{"id":"concepts/computer-use","label":"Computer Use","type":"AIConcept"},{"id":"concepts/tool-use","label":"Tool Use","type":"AIConcept"},{"id":"concepts/multi-step-task-execution","label":"Multi-step Task Execution","type":"AIConcept"},{"id":"concepts/api-pricing","label":"API Pricing","type":"AIConcept"},{"id":"concepts/web-browsing","label":"Web Browsing","type":"AIConcept"},{"id":"concepts/stem","label":"STEM","type":"AIConcept"},{"id":"concepts/cost-efficiency","label":"Cost Efficiency","type":"AIConcept"},{"id":"concepts/ai-governance","label":"AI Governance","type":"AIConcept"},{"id":"concepts/international-ai-policy","label":"International AI Policy","type":"AIConcept"},{"id":"concepts/geopolitics","label":"Geopolitics","type":"AIConcept"},{"id":"concepts/extended-thinking","label":"Extended Thinking","type":"AIConcept"},{"id":"concepts/long-thought-reasoning","label":"Long-Thought Reasoning","type":"AIConcept"},{"id":"concepts/ai-coding-agents","label":"AI Coding Agents","type":"AIConcept"},{"id":"concepts/cognitive-offloading","label":"Cognitive Offloading","type":"AIConcept"},{"id":"concepts/thinking-model","label":"Thinking Model","type":"AIConcept"},{"id":"concepts/hybrid-reasoning","label":"Hybrid Reasoning","type":"AIConcept"},{"id":"concepts/foundation-model","label":"Foundation Model","type":"AIConcept"},{"id":"concepts/multi-hour-workflows","label":"Multi-hour Workflows","type":"AIConcept"},{"id":"concepts/unified-reasoning","label":"Unified Reasoning","type":"AIConcept"},{"id":"concepts/agentic-capability","label":"Agentic Capability","type":"AIConcept"},{"id":"concepts/ai-partnership","label":"AI Partnership","type":"AIConcept"},{"id":"concepts/mobile-ai-integration","label":"Mobile AI Integration","type":"AIConcept"},{"id":"concepts/distribution-channel","label":"Distribution Channel","type":"AIConcept"},{"id":"concepts/consumer-ai","label":"Consumer AI","type":"AIConcept"},{"id":"concepts/production-ai","label":"Production AI","type":"AIConcept"},{"id":"concepts/model-refinement","label":"Model Refinement","type":"AIConcept"},{"id":"concepts/enterprise-ai","label":"Enterprise AI","type":"AIConcept"},{"id":"concepts/reduced-refusals","label":"Reduced Refusals","type":"AIConcept"},{"id":"concepts/ai-music-generation","label":"AI Music Generation","type":"AIConcept"},{"id":"concepts/text-to-music","label":"Text-to-Music","type":"AIConcept"},{"id":"concepts/audio-production","label":"Audio Production","type":"AIConcept"},{"id":"concepts/image-understanding","label":"Image Understanding","type":"AIConcept"},{"id":"concepts/fast-inference","label":"Fast Inference","type":"AIConcept"},{"id":"concepts/ai-tool-integration","label":"AI Tool Integration","type":"AIConcept"},{"id":"concepts/open-infrastructure","label":"Open Infrastructure","type":"AIConcept"},{"id":"concepts/ai-interoperability","label":"AI Interoperability","type":"AIConcept"},{"id":"concepts/protocol-standardization","label":"Protocol Standardization","type":"AIConcept"},{"id":"concepts/cybersecurity-ai","label":"Cybersecurity AI","type":"AIConcept"},{"id":"concepts/0-day-vulnerability-discovery","label":"0-day Vulnerability Discovery","type":"AIConcept"},{"id":"concepts/model-containment","label":"Model Containment","type":"AIConcept"},{"id":"concepts/frontier-model-governance","label":"Frontier Model Governance","type":"AIConcept"},{"id":"concepts/unauthorized-access","label":"Unauthorized Access","type":"AIConcept"},{"id":"concepts/ai-distillation-defense","label":"AI Distillation Defense","type":"AIConcept"},{"id":"concepts/model-distillation","label":"Model Distillation","type":"AIConcept"},{"id":"concepts/export-control","label":"Export Control","type":"AIConcept"},{"id":"concepts/ai-national-security","label":"AI National Security","type":"AIConcept"},{"id":"concepts/geopolitical-ai-competition","label":"Geopolitical AI Competition","type":"AIConcept"},{"id":"concepts/ip-protection","label":"IP Protection","type":"AIConcept"},{"id":"concepts/frontier-model","label":"Frontier Model","type":"AIConcept"},{"id":"concepts/risk-classification","label":"Risk Classification","type":"AIConcept"},{"id":"concepts/model-governance","label":"Model Governance","type":"AIConcept"},{"id":"concepts/capability-safety-tradeoff","label":"Capability-Safety Tradeoff","type":"AIConcept"}],"events":[{"category":"historical-milestone","context":"By October 1950, electronic computing was still in its infancy. The British ACE computer was under construction at the National Physical Laboratory; Turing himself had moved to the University of Manchester, where the Manchester Mark 1 had been running for two years. The war was five years over, and the wartime secrecy around Bletchley Park's code-breaking work had only just begun to lift. Published in the philosophy journal Mind, Turing's paper arrived in a world where \"thinking machines\" was still a phrase for science fiction — yet the academic question he posed, carefully framed as an empirical test, would set the agenda for the next seventy-five years of artificial intelligence research.","date":"1950-10-01","entities":{"benchmarks":["Turing Test"],"concepts":["Turing Test","Computational Theory","AI Foundations","Philosophy of Mind"],"models":[],"organizations":[],"people":["Alan Turing"]},"id":"events/turing-test-1950","significance":"Omega曰: October 1950: Turing published \"Computing Machinery and Intelligence\" in Mind, proposing the Imitation Game. The intellectual moment was the realization that the question \"can machines think?\" was unanswerable — and that a pragmatic test was the best proxy we had.","tags":["Alan Turing","Turing Test","AI Foundations","Philosophy of Mind"],"tier":"LIEZHUAN","title":"The Turing Test: Can Machines Think?","translation_key":"events/turing-test-1950","type":"AIEvent","url":"/events/turing-test-1950/","year":"1950"},{"category":"historical-milestone","context":"By the summer of 1956, electronic computing was still exotic hardware. The IBM 704 — the first mass-produced computer with hardware floating-point arithmetic — was the workhorse of American research labs, equipped with 32K words of magnetic core memory. John McCarthy, then at Dartmouth, had proposed the workshop with Marvin Minsky (Harvard), Claude Shannon (Bell Labs), and Nathaniel Rochester (IBM). All four were in their early thirties. They pitched the Rockefeller Foundation on a ten-week summer project to \"make a machine simulate every aspect of learning or any other feature of intelligence.\" The Foundation approved. The workshop ran from June through August 1956. The name \"artificial intelligence\" was coined in its proposal. The field's founding myth, and its first overconfident promise, were born in the same room.","date":"1956-08-31","entities":{"benchmarks":[],"concepts":["Artificial Intelligence","Symbolic AI","Knowledge Engineering"],"models":[],"organizations":[],"people":["John McCarthy","Marvin Minsky","Claude Shannon","Nathaniel Rochester"]},"id":"events/dartmouth-conference-1956","significance":"Omega曰: Summer 1956: The Dartmouth Summer Research Project on Artificial Intelligence ran at Dartmouth College. Ten weeks. Four organizers. Ten attendees. The term \"artificial intelligence\" was coined in the proposal that funded it. No one in that room knew how hard the problem was. The optimism was total; the funding, modest. Seven decades later, the optimism is being tested on hardware they could not have imagined. The intellectual debt owed to those ten weeks is, in a real sense, incalculable.","tags":["John McCarthy","Marvin Minsky","Dartmouth","AI Founding","Claude Shannon"],"tier":"BENJI","title":"The Dartmouth Conference: Birth of Artificial Intelligence","translation_key":"events/dartmouth-conference-1956","type":"AIEvent","url":"/events/dartmouth-conference-1956/","year":"1956"},{"category":"research-breakthrough","context":"By 1957, Frank Rosenblatt had spent nearly a decade refining the perceptron — a mathematical model of a biological neuron, designed at Cornell Aeronautical Laboratory. Implemented on the IBM 704, the first commercial scientific computer with hardware floating-point arithmetic, the Mark I Perceptron could recognize simple visual patterns and, crucially, adjust its own weights in response to errors. The intellectual climate was optimistic: cybernetics, control theory, and early information theory were converging. DARPA's predecessor agency, ONR, was funding AI research without yet suspecting the field's coming winters. The New York Times gave the perceptron a public debut that the field's later complexity would never quite live down.","date":"1957-07-01","entities":{"benchmarks":[],"concepts":["Perceptron","Neural Networks","Pattern Recognition","Machine Learning"],"models":[],"organizations":["Cornell Aeronautical Laboratory","DARPA"],"people":["Frank Rosenblatt"]},"id":"events/perceptron-1957","significance":"Omega曰: 1957: Rosenblatt published the Perceptron algorithm and demonstrated it on an IBM 704. The New York Times reported it as the embryo of an electronic computer that would \"walk, speak, see, write, reproduce itself and be conscious of its existence.\" The moment was pure, unbounded optimism — long before Minsky \u0026 Papert's 1969 critique revealed the limits of a single-layer network and ushered in the first AI winter.","tags":["Frank Rosenblatt","Perceptron","Neural Networks","Cornell","Pattern Recognition"],"tier":"SHIJIA","title":"The Perceptron: The First Neural Network","translation_key":"events/perceptron-1957","type":"AIEvent","url":"/events/perceptron-1957/","year":"1957"},{"category":"historical-milestone","context":"By December 1966, time-sharing was the new frontier. The IBM 7094 at MIT — one of the most powerful scientific computers of its era — was being used by researchers across many labs through newly developed time-sharing systems. Joseph Weizenbaum, then at the MIT AI Lab, was less than three years into his academic career after years in industry. The Cold War context was ever-present: ARPA, the Department of Defense's research arm, was funding ambitious computing research without yet suspecting the disillusionment that would follow in the 1970s. Weizenbaum wrote ELIZA in a few weeks of MAD-SLIP programming. He intended it as a satire of human-machine conversation. The fact that users began, sincerely, to pour out their hearts to it — and that he was disturbed by this — would make the program one of the most philosophically consequential pieces of code ever written.","date":"1966-12-01","entities":{"benchmarks":[],"concepts":["Chatbot","Natural Language Processing","Human-Computer Interaction"],"models":["ELIZA"],"organizations":["MIT"],"people":["Joseph Weizenbaum"]},"id":"events/eliza-1966","significance":"Omega曰: December 1966: Weizenbaum published ELIZA in Communications of the ACM. The world did not know it was witnessing the first public human-machine dialogue. Most dismissed it as a trick — but the seed of the chatbot era was planted.","tags":["ELIZA","Chatbot","Weizenbaum","MIT","NLP","Human-Computer Interaction"],"tier":"SHIJIA","title":"ELIZA: The First Chatbot That Simulated a Psychotherapist","translation_key":"events/eliza-1966","type":"AIEvent","url":"/events/eliza-1966/","year":"1966"},{"category":"historical-milestone","context":"By 1974, the AI research community was running out of patience — and the funding agencies were running out of money. The Lighthill Report, commissioned by the UK's Science Research Council and published in 1973, had savaged the AI field's promise of general machine intelligence; combined with the 1973-74 recession, ARPA (the Defense Department's research arm, the field's biggest funder) had cut AI funding dramatically. SRI's Shakey robot, the first mobile robot to reason about its actions, was an impressive research project that pointed toward a future the field could not yet deliver. Theoreticians had spent the late 1960s demonstrating that perceptrons could not learn certain simple functions (Minsky and Papert, 1969). Across the field, the dream of human-level AI within a generation was quietly giving way to something more sober. The phrase \"AI winter\" would be coined a few years later by researchers who drew the analogy to nuclear winter — a darkness that follows an explosion of optimism.","date":"1974-01-01","entities":{"benchmarks":[],"concepts":["AI Winter","Symbolic AI","Expert Systems"],"models":[],"organizations":["DARPA"],"people":["James Lighthill"]},"id":"events/first-ai-winter-1974","significance":"Omega曰：The first AI winter arrived when funding collapsed and promises went unmet — the reckoning for years when researchers promised what they could not deliver. Shakey could navigate a room but took hours to do it. The Lighthill Report in the UK killed government support entirely. Yet Hinton and a handful of others kept the neural network flame alive through the cold, not knowing spring would eventually come. That quiet persistence mattered more than any breakthrough of the boom years.","tags":["AI Winter","Lighthill Report","DARPA","Funding Crisis","AI History"],"tier":"BENJI","title":"The First AI Winter: When the Dream Froze","translation_key":"events/first-ai-winter-1974","type":"AIEvent","url":"/events/first-ai-winter-1974/","year":"1974"},{"category":"historical-milestone","context":"By 1980, the AI field was recovering from its first great disappointment. The Lighthill Report (1973) and the DARPA funding cuts of the early 1970s had forced a retreat from grand claims of general intelligence. In their place, a more practical approach took hold: expert systems — programs that encoded the specific decision-making rules of a human expert within a narrow domain. The XCON system at Digital Equipment Corporation, deployed from 1980, was configuring computer orders and saving the company an estimated $40 million a year. MIT's Symbolics 3600 Lisp machines, released in 1981, were the dedicated hardware of this new industry. Stanford's MYCIN demonstrated that expert systems could outperform junior doctors on specific diagnostic tasks. AI was no longer a university research project — it was, for the first time, a product.","date":"1980-01-01","entities":{"benchmarks":[],"concepts":["Expert Systems","Knowledge Engineering","Rule-Based AI"],"models":["MYCIN","DENDRAL","XCON"],"organizations":["Xerox PARC","Stanford","MIT"],"people":["Edward Feigenbaum","Joshua Lederberg","John McCarthy"]},"id":"events/expert-systems-1980","significance":"Omega曰: June 1980: Expert systems began penetrating commercial environments — medicine, geology, finance. The moment was the recognition that knowledge could be encoded as rules and that machines could apply that knowledge at scale. The expert systems era would last a decade before the knowledge acquisition bottleneck brought it down.","tags":["Expert Systems","MYCIN","DENDRAL","XCON","Machine Learning","Knowledge Engineering"],"tier":"SHIJIA","title":"Expert Systems: AI's First Commercial Success","translation_key":"events/expert-systems-1980","type":"AIEvent","url":"/events/expert-systems-1980/","year":"1980"},{"category":"historical-milestone","context":"By the early 1980s, expert systems had been AI's greatest commercial hope, with Lisp machine companies like Symbolics valued in the hundreds of millions; but by 1987, general-purpose PCs had matched Lisp machine performance at one-tenth the price. DARPA's Strategic Computing Program was also dramatically defunded in 1988 after failing to meet milestones. The second AI winter had arrived.\"","date":"1987-01-01","entities":{"benchmarks":[],"concepts":["AI Winter","LISP Machines","Expert Systems"],"models":[],"organizations":["Symbolics","DARPA","Apple","IBM"],"people":[]},"id":"events/second-ai-winter-1987","significance":"Omega曰: 1987-1993: The second AI winter arrived as Lisp machine companies collapsed and expert systems failed to deliver on their promises. The moment was the realization that specialized AI hardware was a dead end — that the future belonged to general-purpose PCs, not symbolic AI accelerators.","tags":["AI Winter","Expert Systems","LISP Machines","DARPA","Machine Learning","History"],"tier":"LIEZHUAN","title":"The Second AI Winter: When Expert Systems Collapsed","translation_key":"events/second-ai-winter-1987","type":"AIEvent","url":"/events/second-ai-winter-1987/","year":"1987"},{"category":"historical-milestone","context":"1996: Deep Blue's first match with Kasparov, lost. IBM engineers retreated and improved for a year. No TPUs existed yet; the GPU compute race was still brewing. Kasparov, emboldened by 1996, declared publicly: 'Machines will never beat me.' The world agreed.","date":"1997-05-11","entities":{"benchmarks":[],"concepts":["Game AI","Minimax Search","Chess AI"],"models":["Deep Blue"],"organizations":["IBM"],"people":["Garry Kasparov","Feng-hsiung Hsu","Murray Campbell"]},"id":"events/deep-blue-kasparov-1997","significance":"Omega曰: May 11, 1997: Deep Blue defeated Kasparov in Game 6. The world had debated AI limits for decades; now a machine had beaten the world champion in a purely intellectual game. Some saw this as proof that intelligence was mechanical. Others saw it as proof that the brain was just a very complex machine.","tags":["IBM","Deep Blue","Garry Kasparov","Chess","Game AI"],"tier":"BENJI","title":"Deep Blue Defeats Kasparov: A Machine Beats the World Champion","translation_key":"events/deep-blue-kasparov-1997","type":"AIEvent","url":"/events/deep-blue-kasparov-1997/","year":"1997"},{"category":"research-breakthrough","context":"By September 1997, the hardware reality of the time was humbling. HP 735 workstations — the workhorses of academic research — were what serious neural network experiments ran on. DDR-SDRAM had only just become available the previous year. The internet was still largely dial-up. GPUs as compute accelerators were fifteen years away. In this environment, Hochreiter and Schmidhuber submitted a paper that had already been rejected by NIPS twice. The Long Short-Term Memory network solved the vanishing-gradient problem that had crippled recurrent neural networks for a decade — a result that, in any other era, would have been celebrated immediately. The paper was accepted; the world, for the moment, did not notice. The seed waited two decades for the hardware to grow.","date":"1997-09-01","entities":{"benchmarks":[],"concepts":["Long Short-Term Memory","Recurrent Neural Networks","Sequence Modeling"],"models":["LSTM"],"organizations":[],"people":["Sepp Hochreiter","Jürgen Schmidhuber"]},"id":"events/lstm-1997","significance":"Omega曰: The 1997 LSTM paper by Hochreiter \u0026 Schmidhuber was initially ignored by the mainstream ML community — it was too ahead of its time for the computing hardware of that era. But its seed would wait 20 years to germinate in every large language model.","tags":["LSTM","Hochreiter","Schmidhuber","RNN","Sequence Modeling","Deep Learning"],"tier":"SHIJIA","title":"LSTM: Long Short-Term Memory Networks","translation_key":"events/lstm-1997","type":"AIEvent","url":"/events/lstm-1997/","year":"1997"},{"category":"research-breakthrough","context":"By September 2012, deep learning had been dormant for nearly a decade. Neural networks had been dismissed as a dead end since the 1969 Perceptron book by Minsky and Papert. But three forces converged: Geoffrey Hinton's group at the University of Toronto had spent years refining the algorithm; the ImageNet dataset, curated by Fei-Fei Li's Stanford team over three years, provided 1.2 million labeled images; and Nvidia's GTX 580 GPUs (Fermi architecture, 3GB GDDR5) made large-scale convolution feasible on commodity hardware. CUDA workshops had spread through North American universities over the previous 18 months, training a generation of researchers to write GPU code. The deep learning revolution began at the intersection of data, compute, and a stubborn insistence that neural networks could work.","date":"2012-09-30","entities":{"benchmarks":["ImageNet"],"concepts":["Deep Learning","Convolutional Neural Networks","GPU Computing","ImageNet"],"models":["AlexNet"],"organizations":["University of Toronto","NVIDIA"],"people":["Geoffrey Hinton","Alex Krizhevsky","Ilya Sutskever"]},"id":"events/alexnet-imagenet-2012","significance":"Omega曰: September 30, 2012: AlexNet trained on 1.2M ImageNet images, reducing top-5 error to 15.3% vs 26.2% for the best competitor. GPU computing had arrived; the intellectual moment was ripe for deep learning to explode.","tags":["Geoffrey Hinton","AlexNet","ImageNet","Deep Learning","GPU","Convolutional Neural Networks"],"tier":"BENJI","title":"AlexNet: The Deep Learning Revolution Begins","translation_key":"events/alexnet-imagenet-2012","type":"AIEvent","url":"/events/alexnet-imagenet-2012/","year":"2012"},{"category":"research-breakthrough","context":"DeepMind's Atari work had matured the self-play methodology; in early 2015, distributed training infrastructure (early Google Cloud TPU v1, 256-chip cluster) was first applied to Go; Fan Hui was the reigning European champion, and the five-game match was held behind closed doors at DeepMind's London office in October 2015 — the outside world knew nothing until the Nature paper was published in January 2016.","date":"2016-01-27","entities":{"benchmarks":[],"concepts":["Reinforcement Learning","Monte Carlo Tree Search","Deep Neural Networks","Self-play"],"models":["AlphaGo","AlphaGo Zero"],"organizations":["Google DeepMind","Google"],"people":["Demis Hassabis","David Silver","Fan Hui","Lee Sedol"]},"id":"events/alphago-fan-hui-2016","significance":"Omega曰: October 2015, announced January 2016: AlphaGo defeated Fan Hui 5-0 — the first time a computer had beaten a professional Go player without handicap. Quiet rooms, no cameras, no headlines. The Go community barely noticed. But five months later, when AlphaGo faced Lee Sedol in Seoul, the world would understand that the real contest in Fan Hui's basement had already been won. The Lee Sedol match was the loud echo of a silent victory.","tags":["DeepMind","AlphaGo","Reinforcement Learning","Go","Google","Fan Hui"],"tier":"BENJI","title":"AlphaGo Defeats Fan Hui: AI Conquers the Ancient Game of Go","translation_key":"events/alphago-fan-hui-2016","type":"AIEvent","url":"/events/alphago-fan-hui-2016/","year":"2016"},{"category":"governance","context":"In January 2017, the Future of Life Institute convened 1,000+ AI researchers at Pacific Grove — the same site where biologists had set recombinant DNA safety standards 42 years earlier. The conference produced 23 Asilomar AI Principles, signed by 1,200+ signatories, representing the field's first broad consensus statement on beneficial AI development.","date":"2017-01-08","entities":{"benchmarks":[],"concepts":["AI Safety","AI Ethics","Existential Risk","Beneficial AI"],"models":[],"organizations":["Future of Life Institute"],"people":[]},"id":"events/asilomar-ai-principles-2017","significance":"Omega曰: In 1975, biologists gathered at Asilomar to set safety standards for recombinant DNA. Forty-two years later, AI researchers did the same for artificial intelligence. The professionals knew the dangers of their craft before disaster struck — this is what set them apart. Those who wait for a crisis before changing course are always behind; those who see danger before it materializes are the wise ones.","tags":["Asilomar","AI Safety","Future of Life Institute","AI Ethics","Beneficial AI","Existential Risk"],"tier":"LIEZHUAN","title":"Asilomar AI Principles: The Research Community Sets Its Own Rules","translation_key":"events/asilomar-ai-principles-2017","type":"AIEvent","url":"/events/asilomar-ai-principles-2017/","year":"2017"},{"category":"research-breakthrough","context":"By June 2017, GPU parallel computing was reshaping the hardware foundations of AI research; Google's TPUs were already in internal deployment, Nvidia's Pascal P100 had become the new standard for training compute; LSTM/RNN had dominated NLP for nearly a decade, and attention mechanism research was circulating across laboratories — the Transformer stood at the intersection of all these conditions, recombining existing elements into something new.\"","date":"2017-06-12","entities":{"benchmarks":["WMT 2014 EN-DE","WMT 2014 EN-FR"],"concepts":["Transformer","Self-Attention","Multi-head Attention","Positional Encoding","Encoder-Decoder Architecture"],"models":[],"organizations":["Google Brain","Google Research"],"people":["Ashish Vaswani","Noam Shazeer","Niki Parmar","Jakob Uszkoreit","Llion Jones","Aidan Gomez","Łukasz Kaiser","Illia Polosukhin"]},"id":"events/transformer-architecture-2017","significance":"Omega曰：Google Brain's \"Attention Is All You Need\" paper introduced the transformer architecture. The intellectual moment was the realization that recurrence was not necessary — that attention alone, with enough parallel compute, could surpass sequential processing entirely. In hindsight, every major model since — GPT, BERT, DALL-E, AlphaFold — can trace their lineage to this seven-author, eight-page paper. The most consequential technical document of the decade was initially treated by Google as a short-term research sprint.","tags":["Transformer","Attention Mechanism","Google Brain","NLP","Vaswani","Self-Attention"],"tier":"BENJI","title":"Attention Is All You Need: The Transformer Architecture","translation_key":"events/transformer-architecture-2017","type":"AIEvent","url":"/events/transformer-architecture-2017/","year":"2017"},{"category":"research-breakthrough","context":"By January 2020, the deep learning field had a missing manual. Training larger models on more data had been shown to work, but the relationship was vague: more compute, more capability, somehow. The OpenAI team — Jared Kaplan, Sam McCandlish, Tom Henighan, and others — had spent two years running controlled experiments across seven orders of magnitude in compute, training hundreds of language models at different scales. What they found was the field's first quantitative compass: model loss followed a power law in compute, data, and parameters. The empirical relationships were so clean they looked like physics. GPT-3 had not yet been trained when the paper was posted on arXiv — but the paper predicted, with some accuracy, what GPT-3's performance would be. A new discipline, the science of scaling, had been born.","date":"2020-01-23","entities":{"benchmarks":[],"concepts":["Scaling Laws","Power Laws","Compute-optimal Training","Emergent Abilities","Neural Language Model Scaling"],"models":[],"organizations":["OpenAI"],"people":["Jared Kaplan","Sam McCandlish","Tom Henighan","Dario Amodei"]},"id":"events/scaling-laws-2020","significance":"Omega曰: January 23, 2020: Kaplan et al. published the neural scaling laws paper, showing that model performance improved as a smooth power law with compute, data, and parameters. The intellectual moment was the realization that bigger was reliably better — and that the race for scale had a predictable payoff. The field discovered that intelligence, at scale, has an underlying order.","tags":["Scaling Laws","OpenAI","Kaplan","Compute","Power Law","LLM","Emergent Abilities"],"tier":"BENJI","title":"Scaling Laws: The Empirical Science of Making AI Smarter","translation_key":"events/scaling-laws-2020","type":"AIEvent","url":"/events/scaling-laws-2020/","year":"2020"},{"category":"capability-unlock","context":"By November 2020, biology's grand challenge of protein structure had resisted human effort for half a century. The number of experimentally determined protein structures — the result of decades of X-ray crystallography, NMR, and cryo-EM — had crept up to around 180,000, while the universe of natural proteins numbered in the billions. The CASP (Critical Assessment of protein Structure Prediction) competition had been running since 1994 as a biennial reality check on prediction methods. The infrastructure for deep learning had just matured: Nvidia's Ampere A100 GPUs were shipping at scale; Google TPU v3 pods were powering large-scale training at DeepMind. AlphaFold had first entered CASP13 (2018) as a promising first entry; the team's three-year follow-up effort — internally called AlphaFold 2 — was being prepared for CASP14. The biological research community was watching. Few expected what was about to be announced.","date":"2020-11-30","entities":{"benchmarks":["CASP14"],"concepts":["Protein Folding","Biology","Drug Discovery","Structural Biology"],"models":["AlphaFold 2"],"organizations":["Google DeepMind"],"people":["Demis Hassabis","John Jumper"]},"id":"events/alphafold-2020","significance":"Omega曰: November 30, 2020: AlphaFold 2 solved the protein folding problem at CASP14, achieving accuracy comparable to experimental methods. The moment was prepared by 50 years of structural biology and a decade of deep learning on sequences.","tags":["AlphaFold","DeepMind","Protein Folding","Biology","Drug Discovery","CASP14"],"tier":"SHIJIA","title":"AlphaFold 2: AI Solves Protein Folding After 50 Years","translation_key":"events/alphafold-2020","type":"AIEvent","url":"/events/alphafold-2020/","year":"2020"},{"category":"capability-unlock","context":"By January 2021, generative AI existed mostly in research papers. Diffusion models had been gaining traction in the image generation community — Denoising Diffusion Probabilistic Models (Ho et al., 2020) had just demonstrated that a carefully designed iterative denoising process could synthesize images competitive with GANs. OpenAI's CLIP, released the month before, had learned a joint latent space between images and text from 400 million image-text pairs. The hardware required to train large generative models was now accessible: Nvidia's Ampere architecture GPUs (A100) were shipping at scale, and Google had just published its Switch Transformer paper on trillion-parameter models. DALL-E's release on January 5 — generating images from text prompts with striking, often surprising coherence — announced that generative AI had moved from research curiosity to public artifact. Within a year, Stable Diffusion, Midjourney, and Imagen would follow.","date":"2021-01-05","entities":{"benchmarks":[],"concepts":["Text-to-Image","Generative AI","Multimodal AI","Diffusion Models"],"models":["DALL-E"],"organizations":["OpenAI"],"people":["Sam Altman","Ilya Sutskever"]},"id":"events/dall-e-2021","significance":"Omega曰: January 5, 2021: OpenAI published DALL-E — a 12-billion parameter GPT-3 variant that generated images from text descriptions. The intellectual moment was CLIP's proof that language and vision could be unified in a single latent space.","tags":["DALL-E","OpenAI","Text-to-Image","Generative AI","Multimodal","Creative AI"],"tier":"SHIJIA","title":"DALL-E: Text-to-Image Generation Arrives","translation_key":"events/dall-e-2021","type":"AIEvent","url":"/events/dall-e-2021/","year":"2021"},{"category":"research-breakthrough","context":"By January 2022, OpenAI had GPT-3 at 175B parameters (June 2020), but the model was not yet aligned to human intent — it could complete text, but it could not reliably follow instructions. Deep RLHF research had been underway inside OpenAI for over a year, building on the work of Christiano, Leike, and others. The InstructGPT paper, published January 27, demonstrated that a 1.3B-parameter model trained with human feedback could outperform the 175B GPT-3 on alignment tasks. The world did not notice. Ten months later, ChatGPT would launch on the same insight — and the world would not stop talking about it.","date":"2022-01-27","entities":{"benchmarks":[],"concepts":["RLHF","Alignment","Instruction Following","Reinforcement Learning from Human Feedback"],"models":["InstructGPT","GPT-3.5"],"organizations":["OpenAI"],"people":["John Schulman","Paul Christiano","Dario Amodei"]},"id":"events/instructgpt-2022","significance":"Omega曰: January 27, 2022: InstructGPT paper published. The world did not notice. But the insight — that human feedback could align language models to follow instructions — was the key unlock for ChatGPT, which came 10 months later.","tags":["InstructGPT","RLHF","Alignment","OpenAI","GPT-3.5","Reinforcement Learning from Human Feedback"],"tier":"SHIJIA","title":"InstructGPT: Language Models Can Learn to Follow Instructions","translation_key":"events/instructgpt-2022","type":"AIEvent","url":"/events/instructgpt-2022/","year":"2022"},{"category":"capability-unlock","context":"By June 2022, OpenAI had been quietly working on a descendant of GPT-3 — fine-tuned on billions of lines of public code from GitHub — and calling it Codex. Microsoft-owned GitHub, with 73 million developers on its platform by then, was the obvious distribution channel. The technical preview, launched in June 2021, had already shown that AI could autocomplete nontrivial functions across multiple programming languages. The commercial launch of GitHub Copilot on June 21, 2022 — integrated directly into VS Code and other IDEs — turned a research demo into a daily-use tool for millions of professional developers. The era of AI as a peripheral curiosity was over. AI had become a colleague sitting inside the editor, suggesting entire functions in real time, asking for nothing in return except a $10/month subscription. It was the first time a knowledge-work profession had embedded AI into its core workflow at industrial scale.","date":"2022-06-21","entities":{"benchmarks":["HumanEval"],"concepts":["Code Generation","AI Coding Assistants","Developer Tools"],"models":["GitHub Copilot","Codex"],"organizations":["GitHub","OpenAI","Microsoft"],"people":["Thomas Dohmke"]},"id":"events/github-copilot-2022","significance":"Omega曰: October 2021 (public), June 2022 (general availability): GitHub Copilot launched — the first AI pair programmer at scale. Within a year, 46% of new code on GitHub was AI-assisted.","tags":["GitHub Copilot","OpenAI","Codex","Code Generation","Developer Tools","AI Integration"],"tier":"SHIJIA","title":"GitHub Copilot: AI Becomes a Software Developer's Colleague","translation_key":"events/github-copilot-2022","type":"AIEvent","url":"/events/github-copilot-2022/","year":"2022"},{"category":"historical-milestone","context":"By November 2022, the technical conditions for ChatGPT were already in place: GPT-3.5 had shipped in June, the RLHF alignment technique had been validated by InstructGPT (January), and Microsoft's Azure supercomputing infrastructure had scaled to support hundreds of millions of inference requests. What remained was the packaging — a free, conversational interface shipped as a 'research preview' on November 30. Five days later, it had one million users. Two months later, one hundred million. The technology had been ready for years; the world had not been ready to receive it.","date":"2022-11-30","entities":{"benchmarks":[],"concepts":["RLHF","Instruction-following","Conversational AI"],"models":["ChatGPT","GPT-3.5"],"organizations":["OpenAI","Microsoft"],"people":["Sam Altman","Greg Brockman","Ilya Sutskever"]},"id":"events/chatgpt-launch-2022","significance":"Omega曰: ChatGPT arrived November 30, 2022. The world had been primed by GPT-3 for two years — but the conversational interface, trained with RLHF, made AI accessible to everyone. Within 5 days it had 1 million users. No technology in history had spread this fast.","tags":["ChatGPT","OpenAI","Large Language Models","GPT-4","RLHF","Generative AI","Consumer AI"],"tier":"BENJI","title":"ChatGPT: The Moment AI Entered Everyday Life","translation_key":"events/chatgpt-launch-2022","type":"AIEvent","url":"/events/chatgpt-launch-2022/","year":"2022"},{"category":"model-release","context":"By March 2023, large language models were already an established product category — but the public-facing models were text-only. OpenAI had been working on GPT-4 for nearly two years, building on the lessons of GPT-3 (2020), InstructGPT (2022), and ChatGPT (November 2022). The training run had taken months; an estimated $100 million in Azure compute; eight months of secret capability testing with external red-teamers. RLHF and SFT pipelines had matured into a reliable alignment stack. The decision to make GPT-4 multimodal from the ground up — accepting both text and image inputs — was made early in training. The release on March 14 was a masterclass in staged disclosure: the technical report came first, the system card laid out capabilities and limitations, and only then did the API open. OpenAI had learned, from the ChatGPT launch, that the world was not ready to receive AI without careful framing.","date":"2023-03-14","entities":{"benchmarks":["MMLU","BAR exam","SAT"],"concepts":["Large Language Model","Multimodal AI","Reasoning"],"models":["GPT-4"],"organizations":["OpenAI"],"people":["Sam Altman","Greg Brockman","Ilya Sutskever"]},"id":"events/gpt4-released","significance":"Omega曰: March 14, 2023 marked **an** inflection point in the history of artificial intelligence. GPT-4, OpenAI's most capable language model to date, passed the bar exam in the 90th percentile and the LSAT in the 88th percentile—benchmarks that measured not merely the accumulation of scale but the emergence of genuine reasoning capability. After eight months of hidden development and secret capability testing, OpenAI chose to release GPT-4 only after internal evaluations confirmed it had crossed thresholds that made it useful for real-world professional tasks. The training pipeline—combining supervised fine-tuning (SFT) with reinforcement learning from human feedback (RLHF)—represented over six months of compute-intensive work, with infrastructure costs on Azure estimated at approximately $100 million. This was not merely a larger GPT-3; it was a fundamentally different artifact, multimodal by design, capable of processing image inputs alongside text, and possessing a 128,000-token context window that enabled analysis of entire codebases, legal contracts, and academic papers in a single prompt. The world had been waiting for this moment since GPT-3's debut in 2020. By 2023, the intellectual soil was fully prepared—and GPT-4's release set off an enterprise AI race that would reshape the technology industry within months.","tags":["OpenAI","LLM","GPT-4"],"tier":"BENJI","title":"GPT-4 Released","translation_key":"events/gpt4-released","type":"AIEvent","url":"/events/gpt4-released/","year":"2023"},{"category":"governance","context":"By November 2023, the AI safety conversation had matured into formal diplomacy. GPT-4 had launched eight months earlier; the EU AI Act was grinding through Brussels; the United States had issued its October 2023 executive order requiring safety testing for frontier models. Frontier AI labs — OpenAI, Anthropic, DeepMind, Google Research — were releasing models that, by their own internal evaluations, exhibited early signs of dangerous capabilities. Geopolitical tensions around AI were rising, yet there was no international forum where the United States, China, the European Union, and the United Kingdom could sit at the same table. The British government, under Prime Minister Rishi Sunak, chose Bletchley Park — the wartime codebreaking site where Turing had worked — as the venue for the first AI Safety Summit. The symbolism was deliberate: where once nations had united to break a machine's cipher, now they would try to govern the machines they themselves were building.","date":"2023-11-01","entities":{"benchmarks":[],"concepts":["AI Safety Summit","Frontier AI","International AI Governance","Existential Risk"],"models":[],"organizations":["UK Government","US Government"],"people":["Rishi Sunak"]},"id":"events/bletchley-declaration-2023","significance":"Omega曰: November 1, 2023: The Bletchley Declaration united 28 countries in acknowledging that frontier AI posed existential risks. The moment was the political realization that AI governance had to precede AI catastrophe — not follow it.","tags":["Bletchley Declaration","AI Safety Summit","International","Frontier AI","Geopolitics","UK"],"tier":"LIEZHUAN","title":"Bletchley Declaration: 28 Nations Agree on AI Frontier Risks","translation_key":"events/bletchley-declaration-2023","type":"AIEvent","url":"/events/bletchley-declaration-2023/","year":"2023"},{"category":"model-release","context":"By February 2024, the long-context race was heating up. Anthropic's Claude 2 (July 2023) had a 100K context window; the open-source community was working with models of 8K-32K; researchers were starting to ask whether scaling context was the next dimension of capability. Google DeepMind had been quietly iterating on Gemini's architecture for over a year. The release of Gemini 1.5 Pro on February 8 — with a stable 1 million token context window, plus 10 million in experimental form — was a discontinuity, not an increment. The model could read an entire novel, an hour of video, or a large codebase in a single context. Google's TPU v5e clusters had matured into the kind of infrastructure that could train and serve such a model. Multimodality — text, code, images, audio, video processed in a single context window — was no longer a research demo. It was a product. The industry standard for \"long context\" was about to be redefined upward by an order of magnitude.","date":"2024-02-08","entities":{"benchmarks":["MMLU","RPM"],"concepts":["Multimodal AI","Long Context","Native Multimodality"],"models":["Gemini 1.5 Pro"],"organizations":["Google","Google DeepMind"],"people":["Sundar Pichai","Demis Hassabis"]},"id":"events/gemini-1.5-pro-released","significance":"Omega曰: February 15, 2024: Gemini 1.5 Pro surprised the world with a 1 million token context window — essentially the entire contents of a human life compressed into working memory. The moment was prepared by Google's years of TPU development and competitive pressure from GPT-4.","tags":["Google","Gemini","Multimodal","Long-context"],"tier":"SHIJIA","title":"Gemini 1.5 Pro Released","translation_key":"events/gemini-1.5-pro-released","type":"AIEvent","url":"/events/gemini-1.5-pro-released/","year":"2024"},{"category":"model-release","context":"By February 2024, the generative AI race had moved from text to images, and the obvious next frontier was video. OpenAI's video generation research had been underway for over two years, building on the diffusion transformer architecture the team had published in late 2022. The infrastructure to train large video models had only just matured: thousands of Nvidia GPUs running in coordinated clusters, with datasets of captioned video scraped from the open web. DALL-E 2 and Stable Diffusion had shown that diffusion models could produce images indistinguishable from photography; Sora demonstrated that the same principle, scaled across space and time, could generate coherent 60-second videos with multi-character interaction, camera movement, and (some) understanding of physical causality. Sora did not ship to the public that day. The model was shown to the world through a curated set of demo videos, and a red team was given access to probe its limits. The promise was: text-to-video would soon be as cheap and ubiquitous as text-to-image already was.","date":"2024-02-15","entities":{"benchmarks":[],"concepts":["Text-to-Video","Diffusion Models","Generative AI","World Model"],"models":["Sora"],"organizations":["OpenAI"],"people":["Sam Altman"]},"id":"events/sora-released","significance":"Omega曰: February 15, 2024: OpenAI released Sora, generating 60-second videos from text. The world had seen image generation (2022) and was primed for video. But Sora's physical plausibility shocked even researchers — the moment when \"AI video\" stopped being a joke.","tags":["OpenAI","Video-generation","Diffusion-model"],"tier":"SHIJIA","title":"Sora Released","translation_key":"events/sora-released","type":"AIEvent","url":"/events/sora-released/","year":"2024"},{"category":"model-release","context":"By March 2024, Anthropic had been building toward Claude 3 for over a year. Constitutional AI — the technique of aligning models using a written set of principles rather than millions of human preference labels — had been published by Anthropic in December 2022, and the SCAI (Self-Critique-Augmented Inference) pipeline had matured through Claude 2's iteration. The LMSYS Chatbot Arena leaderboard, launched in 2023, had become the de facto public ranking for model quality; Claude 2 had performed well, but GPT-4 had dominated. The Haiku and Sonnet variants of the Claude 3 family had already shipped in early March. The release of Claude 3 Opus on March 4 — Anthropic's first true flagship since Claude 2 — claimed to outperform GPT-4 on graduate-level reasoning, undergraduate-level knowledge, and coding benchmarks. The 200K context window, the multimodal vision capability, and the dramatically reduced hallucination rate were positioned as direct answers to the gaps in GPT-4.","date":"2024-03-04","entities":{"benchmarks":["MMLU","GSM8K","MATH"],"concepts":["Large Language Model","Constitutional AI","Safety"],"models":["Claude 3 Opus"],"organizations":["Anthropic"],"people":["Dario Amodei","Daniela Amodei"]},"id":"events/claude-3-opus-released","significance":"Omega曰: March 4, 2024: Anthropic released Claude 3 Opus, immediately topping LMSYS leaderboard. The moment was prepared by Claude 2's surprising success and Anthropic's growing confidence in constitutional AI as a viable alignment strategy.","tags":["Anthropic","LLM","Claude-3"],"tier":"SHIJIA","title":"Claude 3 Opus Released","translation_key":"events/claude-3-opus-released","type":"AIEvent","url":"/events/claude-3-opus-released/","year":"2024"},{"category":"governance","context":"By March 2024, the AI alignment conversation was being held in two very different rooms. In research labs, teams were working on the technical details of RLHF, constitutional AI, and red-teaming. In boardrooms and government offices, a different question was being asked: how should a technology this powerful be regulated? The European Parliament had been working on the AI Act for nearly two years; in 2023 the pace accelerated dramatically, with final plenary votes scheduled. The Act established a risk-based regulatory architecture: unacceptable risk (banned), high risk (heavily regulated), limited risk (transparency required), and minimal risk (largely unregulated). The text covered foundation models separately, requiring transparency, copyright compliance, and risk assessments for the largest general-purpose systems. On March 13, 523 members of the European Parliament voted in favor; 46 voted against; 49 abstained. The world's first comprehensive horizontal AI law had passed.","date":"2024-03-13","entities":{"benchmarks":[],"concepts":["AI Regulation","Risk-Based AI Governance","AI Compliance","AI Law"],"models":[],"organizations":["European Union","European Parliament"],"people":[]},"id":"events/eu-ai-act-2024","significance":"Omega曰: March 13, 2024: The EU AI Act passed — the world's first comprehensive AI law. The moment was not a response to catastrophe but the anticipation of one: two years of fierce lobbying, GPT-4's release crystallizing public anxiety, the Brussels Effect already quietly reshaping corporate practice in advance. Europe legislated before the disaster, in the hope that legislation could prevent it. Whether the law moves faster than the technology it seeks to govern remains the question of the decade.","tags":["EU AI Act","Regulation","European Union","AI Law","Risk-Based","Policy"],"tier":"LIEZHUAN","title":"EU AI Act: The World's First Comprehensive AI Law","translation_key":"events/eu-ai-act-2024","type":"AIEvent","url":"/events/eu-ai-act-2024/","year":"2024"},{"category":"model-release","context":"By April 2024, the open-source LLM ecosystem had reached a turning point. Meta had released Llama 2 in July 2023 under a custom commercial license, and the community had proven the demand: hundreds of fine-tuned variants, a thriving Hugging Face ecosystem, and enterprises adopting open models for private deployment. Meta had spent the intervening months scaling up its training infrastructure — building GPU clusters large enough to train on 15 trillion tokens, more than seven times the data Llama 2 had seen. The April 18 release of Llama 3 — in 8B and 70B sizes, with the MIT-style license many had lobbied for — was a deliberate move: maintain the open-weight lead before the closed-source labs could pull away, and let the community do the work of distributing Meta's models to every developer on earth. Open-weights competition with closed-source frontier was no longer a research project; it was an industrial reality.","date":"2024-04-18","entities":{"benchmarks":["MMLU","ARC"],"concepts":["Open Weights","Large Language Model","Open Source AI"],"models":["Llama 3"],"organizations":["Meta"],"people":["Mark Zuckerberg"]},"id":"events/llama-3-released","significance":"Omega曰: April 18, 2024: Meta released Llama 3, immediately spawning the open-source model ecosystem. The intellectual moment had been building since Llama 2 — the community had shown that open models could compete with closed ones if given enough compute and curation.","tags":["Meta","Open-source","LLM","Llama"],"tier":"SHIJIA","title":"Llama 3 Released","translation_key":"events/llama-3-released","type":"AIEvent","url":"/events/llama-3-released/","year":"2024"},{"category":"model-release","context":"On May 13, 2024, OpenAI released GPT-4o — \"o\" for omni — capable of seeing, hearing, and speaking in real time. GPT-4 had already dominated for eighteen months, and the industry had long craved truly natural conversation. OpenAI chose this moment to release its omni model: the multimodal race was heating up, with Google betting heavily on Gemini — GPT-4o was OpenAI's answer, not a holding action, but a declaration.","date":"2024-05-13","entities":{"benchmarks":[],"concepts":["Multimodal AI","Native Voice","Real-time Interaction"],"models":["GPT-4o"],"organizations":["OpenAI"],"people":["Sam Altman","Greg Brockman"]},"id":"events/gpt-4o-released","significance":"Omega曰：The moment was prepared by GPT-4's 18 months of dominance and the growing demand for voice assistants that felt natural. But multimodal reach does not mean multimodal understanding — the mystery of consciousness remains unsolved.","tags":["OpenAI","Multimodal","GPT-4","Voice"],"tier":"SHIJIA","title":"GPT-4o Released","translation_key":"events/gpt-4o-released","type":"AIEvent","url":"/events/gpt-4o-released/","year":"2024"},{"category":"model-release","context":"June 2024: OpenAI was riding the wave of GPT-4o's May release, GPT-4's coding throne still seemed unchallenged; Anthropic had shipped Claude 3 in March (Haiku/Sonnet/Opus tiered lineup), and market attention was on Claude 3 Opus vs GPT-4; then 3.5 Sonnet arrived at \"medium\" tier and procedurally outperformed both flagships on real coding tasks — that contrast was the story.","date":"2024-06-20","entities":{"benchmarks":["SWE-bench","HumanEval","GPQA Diamond","MMLU"],"concepts":["Coding Agents","Context Window","Long Context Reasoning"],"models":["Claude 3.5 Sonnet","Claude 3 Sonnet","Claude 3 Opus"],"organizations":["Anthropic"],"people":["Dario Amodei","Daniela Amodei"]},"id":"events/claude-35-sonnet-2024","significance":"Omega曰: June 20, 2024: Anthropic released Claude 3.5 Sonnet — and immediately set a new standard for AI-assisted coding. Within hours of release, developers on X (Twitter) were sharing examples of Claude writing entire functions, debugging complex issues, and explaining legacy codebases in seconds. The industry noticed: for the first time since GPT-4, a model had definitively surpassed the previous leader in practical coding tasks.","tags":["Claude 3.5 Sonnet","Anthropic","Coding","SWE-bench","AI Developer Tools"],"tier":"SHIJIA","title":"Claude 3.5 Sonnet: Anthropic's Coding Supremacy","translation_key":"events/claude-35-sonnet-2024","type":"AIEvent","url":"/events/claude-35-sonnet-2024/","year":"2024"},{"category":"model-release","context":"By mid-2024, Mistral had proven the viability of sparse MoE architectures with Mixtral, while Llama 3 was establishing open-source momentum. With Large 2, Mistral took the opposite path — a dense 123B model claiming near-GPT-4o capability, released openly with a commercial-use license. Two architectures from the same French lab, signaling Europe's refusal to remain dependent on US frontier labs.","date":"2024-07-29","entities":{"benchmarks":["MMLU","HumanEval"],"concepts":["Open Source AI","Large Language Model","Coding"],"models":["Mistral Large 2"],"organizations":["Mistral"],"people":["Arthur Mensch"]},"id":"events/mistral-large-2-released","significance":"Omega曰: July 29, 2024: Mistral AI released Large 2, claiming near GPT-4o performance at 123B parameters with open-source customizability. The European moment had arrived — Mistral was no longer a fringe player but a legitimate competitor.","tags":["Mistral","Open-source","LLM","Coding"],"tier":"SHIJIA","title":"Mistral Large 2 Released","translation_key":"events/mistral-large-2-released","type":"AIEvent","url":"/events/mistral-large-2-released/","year":"2024"},{"category":"concept","context":"By 2024, returns from training-time compute scaling had begun to plateau — gains from GPT-4 to GPT-4o were far smaller than those from GPT-3 to GPT-4; GPU cluster sizes had approached practical limits for a single training run; theoretical work from DeepMind and UC Berkeley published in August 2024 had demonstrated that inference-time compute could serve as an effective substitute for a larger model; OpenAI's internal reinforcement learning experiments had, over several months, found a method for converting 'extended thinking' into a trainable reward signal.","date":"2024-09-12","entities":{"benchmarks":["AIME","MATH-500","Codeforces"],"concepts":["Inference-time Scaling","Test-time Compute","Reinforcement Learning","Chain-of-thought Reasoning","Compute Scaling"],"models":["o1","o1-mini"],"organizations":["OpenAI","Google DeepMind","UC Berkeley"],"people":[]},"id":"events/inference-time-scaling-2024","significance":"Omega曰: Train more, train longer — that was the operating faith of the past decade. o1 announced a second path: let the model think before answering. This is not merely an engineering trick — it changes the underlying philosophy of AI capability, from 'how much did it memorize' to 'how much can it reason to.' A new dimension of the scaling law had opened.","tags":["Inference Scaling","Chain-of-Thought","Test-Time Compute","OpenAI o1","Reasoning Models","RL"],"tier":"SHIJIA","title":"Inference-Time Scaling: The New Frontier of AI Capability","translation_key":"events/inference-time-scaling-2024","type":"AIEvent","url":"/events/inference-time-scaling-2024/","year":"2024"},{"category":"model-release","context":"Summer 2024: o3 unnamed; Claude 3.5 had swept coding benchmarks; Gemini 1.5 had broken the 1M-token context barrier — the inference-time scaling race had begun quietly. OpenAI's internal codename \"Strawberry\" was tightly guarded. The industry broadly believed Scaling Laws were hitting a wall, and GPT-5's trajectory remained uncertain.","date":"2024-09-12","entities":{"benchmarks":["AIME","Codeforces","GPQA Diamond"],"concepts":["Inference-time Scaling","Chain-of-thought Reasoning","Test-time Compute","Reinforcement Learning"],"models":["o1","o1-mini","o1-preview"],"organizations":["OpenAI"],"people":["Sam Altman"]},"id":"events/openai-o1-released","significance":"Omega曰：OpenAI o1 released — a model that thinks before answering. The intellectual moment was the recognition that scaling inference compute was as important as training compute. For 70 years, AI researchers had focused on the model. o1 shifted the focus to the reasoning process itself.","tags":["OpenAI","Reasoning","Chain-of-thought","o1"],"tier":"SHIJIA","title":"OpenAI o1 Released","translation_key":"events/openai-o1-released","type":"AIEvent","url":"/events/openai-o1-released/","year":"2024"},{"category":"model-release","context":"December 2024: The reasoning model race was at its peak — o1 (September), Claude 3.5 Sonnet (June), Gemini 1.5 (February) each commanded different dimensions of advantage; Google chose a different path: not the smartest model, but the most useful multi-agent infrastructure; Gemini 2.0 Flash's native tool use (code execution, Google Search, user tools) allowed a single model to orchestrate multiple subtasks — becoming a critical foundation for the agentic AI wave of early 2025.","date":"2024-12-11","entities":{"benchmarks":["MMLU","HumanEval","MATH","AgentBench"],"concepts":["Native Tool Use","Multi-Agent Orchestration","Agentic AI","Low Latency","Function Calling"],"models":["Gemini 2.0 Flash","Gemini 1.5 Pro","Gemini 2.0 Flash-Thinking"],"organizations":["Google DeepMind"],"people":["Demis Hassabis","Oriol Vinyals"]},"id":"events/gemini-2-flash-2024","significance":"Omega曰: December 11, 2024: Google released Gemini 2.0 Flash — a model optimized not for benchmark dominance but for speed, cost efficiency, and multi-agent orchestration. Where competitors raced to build the smartest model, Google bet on the most useful one. Gemini 2.0 Flash introduced native tool use and agentic capabilities as first-class features — a direct response to the agentic AI wave catalyzed by Claude 3.5 Sonnet and OpenAI's o1.","tags":["Gemini 2.0 Flash","Google DeepMind","Agentic AI","Multi-Agent","Native Tool Use","Speed","API"],"tier":"SHIJIA","title":"Gemini 2.0 Flash: Google's Ultra-Fast Multi-Agent Model","translation_key":"events/gemini-2-flash-2024","type":"AIEvent","url":"/events/gemini-2-flash-2024/","year":"2024"},{"category":"capability-unlock","context":"December 2024: The inference model arms race was entering its second phase — o1 (September) proved inference-time compute worked, Claude 3.5 Sonnet (June) proved coding could be dramatically improved, DeepSeek R1 (January 2025) not yet released — the entire industry was waiting for o3 to define the ceiling. o3 arrived with 87.5% on ARC-AGI (Extend setting), shattering the 'Scaling Laws are dead' narrative.","date":"2024-12-20","entities":{"benchmarks":["ARC-AGI","AIME","GPQA Diamond","Codeforces"],"concepts":["ARC-AGI","Fluid Reasoning","Inference-time Scaling","Test-time Compute","AGI","Reasoning"],"models":["o3","o3-mini","o1"],"organizations":["OpenAI"],"people":["Sam Altman","Mira Murati","Ilya Sutskever"]},"id":"events/openai-o3-2024","significance":"Omega曰：December 2024: OpenAI released o3 — and it shattered the ARC-AGI benchmark that had been called \"AI's IQ test\" for five years. A score of 87.5% (Extend setting), compared to the previous record of 55% — that gap was not incremental progress, it was category change. For the first time, an AI system showed signs of genuine fluid reasoning: the ability to attack novel problems without memorized solutions. The field had never seen a capability jump this large on a benchmark specifically designed to resist pattern matching.","tags":["OpenAI","o3","ARC-AGI","Reasoning","AGI","Frontier Models","Inference-time Scaling"],"tier":"BENJI","title":"OpenAI o3: The ARC-AGI Breakthrough That Stunned the Field","translation_key":"events/openai-o3-2024","type":"AIEvent","url":"/events/openai-o3-2024/","year":"2024"},{"category":"historical-milestone","context":"By early 2025, OpenAI's o1 had demonstrated in September 2024 that inference-time compute could dramatically improve model capability; US chip export controls (H100 embargo) had tightened steadily, and the prevailing consensus was that China's AI development lagged by one to two generations; industry wisdom held that training frontier reasoning models required hundreds of millions of dollars in compute. The reckoning arrived seven days into the new year.","date":"2025-01-20","entities":{"benchmarks":["AIME 2024","MATH-500","Codeforces","GPQA Diamond"],"concepts":["Pure Reinforcement Learning","Chain-of-thought Reasoning","Mixture of Experts","Compute Efficiency"],"models":["DeepSeek R1","DeepSeek-V3"],"organizations":["DeepSeek"],"people":[]},"id":"events/deepseek-r1-2025","significance":"Omega曰: When the Soviet Union launched Sputnik in 1957, Americans were not shocked that something had reached orbit — they were shocked that the Soviets had done it. DeepSeek R1 landed the same way: the shock was not that AI had grown more capable, but that 'expensive' and 'powerful' were no longer synonymous. For the valuation logic underpinning the US AI industry, this was an earthquake. For the global AI landscape, it was a reshuffling.","tags":["DeepSeek","China","Open Source","Reasoning Model","Geopolitics","Nvidia","AI Race"],"tier":"BENJI","title":"DeepSeek R1: China's AI Sputnik Moment","translation_key":"events/deepseek-r1-2025","type":"AIEvent","url":"/events/deepseek-r1-2025/","year":"2025"},{"category":"historical-milestone","context":"On January 20, 2025, Trump signed dozens of executive orders on his first day back in office; Biden's October 2023 AI safety executive order had been in force for 15 months, requiring large AI models to disclose safety testing results before deployment; on the same day, DeepSeek R1 was released in China, demonstrating unexpected competitive AI capability; the tech industry broadly anticipated regulatory rollback, and tech figures who had backed Trump — including Musk and Andreessen — viewed AI safety requirements as obstacles to American competitiveness.","date":"2025-01-20","entities":{"benchmarks":[],"concepts":["AI Deregulation","National Security","AI Policy","Executive Order"],"models":[],"organizations":["US Government"],"people":["Donald Trump"]},"id":"events/trump-ai-deregulation-2025","significance":"Omega曰: Two AI governance philosophies confronted each other directly in early 2025: one held that unconstrained AI competition was the only path to securing American leadership; the other argued that leadership without safety guarantees was a threat to all of humanity. Trump's choice was unambiguous: speed first. The consequences of that choice — for better or worse — will unfold across the next decade.","tags":["Trump","US Policy","AI Regulation","Executive Order","Deregulation","Biden AI Order","National Security"],"tier":"BENJI","title":"Trump's AI Deregulation: America Chooses Speed Over Safety","translation_key":"events/trump-ai-deregulation-2025","type":"AIEvent","url":"/events/trump-ai-deregulation-2025/","year":"2025"},{"category":"capability-unlock","context":"By late 2024, GPT-4o's vision capabilities had matured; Anthropic pioneered the space with its Computer Use API in October 2024, giving AI direct access to desktops for the first time; a wave of startups — Cognition's Devin chief among them — had demonstrated AI autonomously completing engineering tasks throughout the year; browser automation toolchains (Playwright, Puppeteer) provided ready-made interfaces. The infrastructure for agency had been quietly assembled before anyone had built the agent.","date":"2025-01-23","entities":{"benchmarks":[],"concepts":["Agentic AI","Model Context Protocol","Computer Use","Tool Use","Multi-step Task Execution"],"models":["Claude Computer Use","OpenAI Operator","Gemini Deep Research"],"organizations":["OpenAI","Anthropic","Google DeepMind"],"people":[]},"id":"events/ai-agents-emerge-2025","significance":"Omega曰: In 2022, AI was a conversational partner: you spoke, it replied. By 2025, AI was an executor: you named a goal, it determined the steps, it completed the task. This was not a feature upgrade — it was a redefinition of the relationship between humans and machines. For the first time, the question that mattered was not 'what did the AI say?' but 'what did the AI do?'.","tags":["AI Agents","OpenAI Operator","Claude Computer Use","Autonomous AI","MCP","Deep Research","Agentic"],"tier":"SHIJIA","title":"Agentic AI: From Chatbots to Autonomous Actors","translation_key":"events/ai-agents-emerge-2025","type":"AIEvent","url":"/events/ai-agents-emerge-2025/","year":"2025"},{"category":"model-release","context":"January 20, 2025: DeepSeek R1 shocked the global AI community by achieving o1-level reasoning at dramatically lower cost; OpenAI o3-mini launched January 31 — OpenAI's direct answer to R1, competing not on price alone but on capability + web browsing + reasonable cost, particularly in STEM domains.","date":"2025-01-31","entities":{"benchmarks":["AIME","GPQA Diamond","Codeforces"],"concepts":["Reasoning","API Pricing","Web Browsing","STEM","Cost Efficiency"],"models":["o3-mini","o1","o1-mini"],"organizations":["OpenAI"],"people":["Sam Altman"]},"id":"events/openai-o3-mini-2025","significance":"Omega曰: January 31, 2025: OpenAI released o3-mini, making frontier-level reasoning accessible at a fraction of the cost of o3. Where o3 cost hundreds of dollars per complex query, o3-mini was priced for everyday use. It was OpenAI's answer to DeepSeek R1 (released January 20): not the cheapest, but the most capable at its price point — and with full access to the internet via browsing.","tags":["OpenAI","o3-mini","Reasoning","Cost Efficiency","API","OpenAI"],"tier":"SHIJIA","title":"OpenAI o3-mini: Affordable Reasoning for Everyone","translation_key":"events/openai-o3-mini-2025","type":"AIEvent","url":"/events/openai-o3-mini-2025/","year":"2025"},{"category":"governance","context":"In early 2025, DeepSeek R1 had shaken the global AI landscape just two weeks prior; Trump had taken office and immediately revoked Biden's AI safety executive order; the EU AI Act had passed and was approaching implementation; global consensus on AI governance was fracturing under the pressure of national interests; Macron sought to position France as a genuine third pole in global AI governance, distinct from both US dominance and Chinese state-directed development.","date":"2025-02-10","entities":{"benchmarks":[],"concepts":["AI Governance","International AI Policy","AI Safety","Geopolitics"],"models":[],"organizations":["French Government","European Union"],"people":["Emmanuel Macron"]},"id":"events/paris-ai-summit-2025","significance":"Omega曰: Bletchley Park was the moment AI nations signed the same document. Paris was the moment they recognized their fundamental divergences — the US and UK refused to sign, each nation announcing its own AI trajectory. If Bletchley was year one of AI governance, Paris was year one of AI multipolarity. The future of AI will not be shaped by a single rulebook, but by competing frameworks in permanent, unresolved negotiation.","tags":["Paris AI Summit","Macron","AI Governance","International","US-China","AI Safety","France"],"tier":"LIEZHUAN","title":"Paris AI Action Summit: A Fractured World Negotiates AI's Future","translation_key":"events/paris-ai-summit-2025","type":"AIEvent","url":"/events/paris-ai-summit-2025/","year":"2025"},{"category":"model-release","context":"February 2025: The reasoning model arms race was white-hot — o3 (December 2024) stunned with 87.5% ARC-AGI, o3-mini (January 31) entered at low API cost, DeepSeek R1 (January 20) disrupted with open-source economics; Anthropic chose not to compete head-on on benchmarks but on real-world programming experience — Extended Thinking allowed the model to pause and reason through tens of thousands of tokens before responding, excelling at complex codebase refactoring tasks.","date":"2025-02-24","entities":{"benchmarks":["SWE-bench","HumanEval","TAU-bench","OSWorld"],"concepts":["Extended Thinking","Long-Thought Reasoning","AI Coding Agents","Cognitive Offloading"],"models":["Claude 3.7 Sonnet","Claude 3.5 Sonnet","Claude 3 Opus"],"organizations":["Anthropic"],"people":["Dario Amodei","Daniela Amodei"]},"id":"events/claude-37-sonnet-2025","significance":"Omega曰: February 24, 2025: Anthropic released Claude 3.7 Sonnet — and with it, the concept of \"extended thinking\" became a mainstream feature. The model could literally pause and reason through a complex problem for tens of thousands of tokens before responding. For coding tasks, this meant Claude could plan, implement, test, and debug in a single conversation with human-level persistence. The industry called it the first \"10x developer\" model.","tags":["Claude 3.7 Sonnet","Anthropic","Extended Thinking","Coding","AI Agents","Long Context"],"tier":"SHIJIA","title":"Claude 3.7 Sonnet: Anthropic's Extended Thinking Model","translation_key":"events/claude-37-sonnet-2025","type":"AIEvent","url":"/events/claude-37-sonnet-2025/","year":"2025"},{"category":"model-release","context":"By March 2025, DeepSeek had completely reset cost expectations for reasoning models; OpenAI's o3-mini had launched in late January and Anthropic's Claude 3.7 in late February; Google had spent months rebuilding technical credibility after the 'Gemini image incident' of early 2024 — a politically charged controversy over historically inaccurate AI-generated images; Google's internal TPU v5e clusters had expanded substantially in the second half of 2024, providing ample training compute for a major push.","date":"2025-03-25","entities":{"benchmarks":["LMArena","MMLU"],"concepts":["Thinking Model","Hybrid Reasoning","Long Context"],"models":["Gemini 2.5 Pro"],"organizations":["Google DeepMind"],"people":["Sundar Pichai","Demis Hassabis"]},"id":"events/gemini-2-5-pro-2025","significance":"Omega曰: Google invented the Transformer, then spent years watching OpenAI collect the laurels of the architecture it had given the world. Gemini 2.5 Pro reaching the top of the LMArena leaderboard was not just a benchmark victory — it was Google answering, finally, the question that had hung unanswered since ChatGPT: the company that invented attention had been unable to focus it. The crown was heavier than the years had made it look.","tags":["Gemini 2.5 Pro","Google DeepMind","Thinking Model","LMArena","Multimodal","Long Context"],"tier":"SHIJIA","title":"Gemini 2.5 Pro: Google's Thinking Model Takes the Lead","translation_key":"events/gemini-2-5-pro-2025","type":"AIEvent","url":"/events/gemini-2-5-pro-2025/","year":"2025"},{"category":"model-release","context":"By April 2025, DeepSeek had proven that efficient sparse architectures (MoE) could dramatically reduce inference costs while maintaining frontier performance; Meta had rebuilt its leadership in the open-source model ecosystem through the Llama 3 series in 2024; Nvidia H100 clusters at scale made training on 30T+ tokens practical; Zuckerberg had publicly committed in early 2025 that Meta would spare nothing to maintain its position at the frontier of open-source AI.","date":"2025-04-05","entities":{"benchmarks":["MMLU","ARC"],"concepts":["Mixture of Experts","Open Weights","Multimodal AI","Foundation Model"],"models":["Llama 4"],"organizations":["Meta"],"people":["Mark Zuckerberg"]},"id":"events/llama-4-2025","significance":"Omega曰: Each Llama release moves the open-source frontier forward by a generation. Llama 4's MoE architecture — 17 billion active parameters, edge-deployable, rivaling closed models that were state-of-the-art months ago — is not the democratization of personal AI access. It is the democratization of organizational AI capability. A frontier model in every server room, not in every pocket.","tags":["Llama 4","Meta","Open Weights","Mixture of Experts","MoE","Multimodal","Scout","Maverick","Behemoth"],"tier":"SHIJIA","title":"Llama 4: Meta Bets on Open Weights and Mixture of Experts","translation_key":"events/llama-4-2025","type":"AIEvent","url":"/events/llama-4-2025/","year":"2025"},{"category":"model-release","context":"By May 2025, Claude 3.7 Sonnet had proven in February that hybrid reasoning modes were viable at production scale; Claude Code had accumulated months of real-world user feedback; Anthropic's internal research on multi-hour agentic workflows had reached a critical threshold; GPT-4.1 in April and Gemini 2.5 Pro in March had raised the competitive bar significantly; Anthropic was in an accelerated phase following the completion of a major new funding round.","date":"2025-05-22","entities":{"benchmarks":["SWE-bench Verified"],"concepts":["Agentic AI","Multi-hour Workflows","Constitutional AI","Hybrid Reasoning"],"models":["Claude 4 Opus","Claude 4 Sonnet","Claude Code"],"organizations":["Anthropic"],"people":["Dario Amodei","Daniela Amodei"]},"id":"events/claude-4-2025","significance":"Omega曰: Anthropic's founding belief was always that the safest AI should be the most capable — and Claude 4 was their sharpest test of that thesis. SWE-bench Verified 72.5%, multi-hour autonomous workflows, agentic task completion: the double helix of safety and capability, ascending together. The myth of the capability-safety trade-off, it turns out, was a tradeoff of execution, not of physics.","tags":["Claude 4","Anthropic","Opus 4","Sonnet 4","Agentic AI","SWE-bench","Claude Code"],"tier":"SHIJIA","title":"Claude 4: Anthropic's Agentic Frontier","translation_key":"events/claude-4-2025","type":"AIEvent","url":"/events/claude-4-2025/","year":"2025"},{"category":"model-release","context":"In the first half of 2025, o3 launched in April, Gemini 2.5 Pro topped LMArena in March, and Claude 4 arrived in May — frontier model competition was reaching maximum intensity; OpenAI had repeatedly delayed GPT-5's release (originally expected in early 2025), citing ongoing safety testing and capability evaluation; Sam Altman had signaled publicly on multiple occasions that GPT-5 would be 'the most important model release in history.'","date":"2025-08-07","entities":{"benchmarks":["AIME 2025","SWE-bench Verified","MMMU"],"concepts":["AGI","Unified Reasoning","Native Multimodality","Agentic Capability"],"models":["GPT-5"],"organizations":["OpenAI"],"people":["Sam Altman"]},"id":"events/gpt-5-2025","significance":"Omega曰：Each GPT generation has redefined the boundary of what AI can do. GPT-5 was the first time the phrase \"meaningful step toward AGI\" was used to describe a commercial product release — not by a researcher, but by its makers. Whether accurate or not, this marked the moment the AI industry crossed a rhetorical threshold in its public narrative. The AGI conversation had entered the product announcement.","tags":["GPT-5","OpenAI","Sam Altman","AGI","Reasoning","Benchmark","Foundation Model"],"tier":"BENJI","title":"GPT-5: OpenAI's Most Capable Model and Its AGI Claims","translation_key":"events/gpt-5-2025","type":"AIEvent","url":"/events/gpt-5-2025/","year":"2025"},{"category":"uncategorized","context":"","date":"2026-01-01","entities":{"benchmarks":[],"concepts":[],"models":[],"organizations":[],"people":[]},"id":"events/SCHEMA","significance":"","tags":[],"tier":"LIEZHUAN","title":"Event Page Schema","translation_key":"","type":"AIEvent","url":"/events/schema/","year":"2026"},{"category":"historical-milestone","context":"The AI provider for Siri, switching from OpenAI to Google, was itself a commercial decision. But in the broader context, it represented a significant reorganization of the US AI ecosystem: Apple outsourced the core intelligence of its next-generation Siri to Google, rather than OpenAI. This gave Google direct access to billions of users through a hardware channel, while OpenAI lost a potentially massive distribution pathway.","date":"2026-01-12","entities":{"benchmarks":[],"concepts":["AI Partnership","Mobile AI Integration","Distribution Channel","Consumer AI"],"models":["Gemini","GPT-4o"],"organizations":["Apple","Google","OpenAI"],"people":["Sundar Pichai","Tim Cook"]},"id":"events/apple-google-gemini-partnership-2026","significance":"Omega曰：The AI war is not fought in laboratories, but in distribution channels. Gemini entering Siri represents the largest-scale deployment of Google's cloud AI capability to date. Apple traded user experience for AI capability; Google traded API revenue for user access; OpenAI lost an arm. Each played their part, and users entered the game without knowing they were playing. The long-term implications of this partnership may far exceed its initial news cycle.","tags":["Apple","Google","Gemini","Siri","Partnership","OpenAI","Integration"],"tier":"SHIJIA","title":"Apple Partners with Google Gemini for Next-Generation Siri","translation_key":"events/apple-google-gemini-partnership-2026","type":"AIEvent","url":"/events/apple-google-gemini-partnership-2026/","year":"2026"},{"category":"model-release","context":"On February 5, 2026, Anthropic released Claude Opus 4.6; on February 17, Claude Sonnet 4.6 followed. These were incremental updates to the Claude 4 family, focused on tool use improvements and reduced refusal rates. For enterprise users, what truly mattered was \"production viability\" rather than \"benchmark domination\" — and the 4.6 series embodied exactly that philosophy.","date":"2026-02-05","entities":{"benchmarks":[],"concepts":["Tool Use","Production AI","Model Refinement","Enterprise AI","Reduced Refusals"],"models":["Claude Opus 4.6","Claude Sonnet 4.6"],"organizations":["Anthropic"],"people":[]},"id":"events/claude-opus-46-sonnet-46-2026","significance":"Omega曰: The significance of the 4.6 series lay not in breakthrough, but in refinement. When model capability was already sufficient, the next frontier of competition became \"usability.\" Reducing refusals and improving tool-call accuracy — these improvements were often more valuable to developers than a 1-2 percentage point benchmark increase. The product thinking demonstrated in the 4.6 series deserved attention.","tags":["Anthropic","Claude","Sonnet 4.6","Opus 4.6","Tool Use","Production AI"],"tier":"SHIJIA","title":"Claude 4.6 Updates: Tool Use and Production Capability Refinements","translation_key":"events/claude-opus-46-sonnet-46-2026","type":"AIEvent","url":"/events/claude-opus-46-sonnet-46-2026/","year":"2026"},{"category":"capability-unlock","context":"On February 18, 2026, Google's third-generation music generation model Lyria 3 was integrated into the Gemini platform. This was the first time a music generation AI entered a mainstream multimodal AI platform via API — users could generate full music tracks directly through Gemini with a text prompt. It marked the extension of generative AI from text/images to full-scale audio/video production.","date":"2026-02-18","entities":{"benchmarks":[],"concepts":["AI Music Generation","Text-to-Music","Multimodal AI","Audio Production"],"models":["Lyria 3","Gemini"],"organizations":["Google DeepMind"],"people":[]},"id":"events/google-lyria-3-2026","significance":"Omega曰：AI-generated music has been in the news since Suno and Udio. But Lyria 3 in Gemini was not about music generation itself — it was about embedding music capability into a consumer AI assistant with hundreds of millions of daily active users. The significance lay not in the technical breakthrough, but in the scale effect. When the barrier to music creation drops to zero, the real change will not be in the tools, but in the meaning of music itself.","tags":["Google","Lyria","Music Generation","Gemini","Multimodal","Audio"],"tier":"SHIJIA","title":"Google Lyria 3: AI Music Generation Comes to Gemini","translation_key":"events/google-lyria-3-2026","type":"AIEvent","url":"/events/google-lyria-3-2026/","year":"2026"},{"category":"model-release","context":"Gemini 3.1 Flash Image appeared in the Vertex AI Catalog on February 25, 2026, marking a significant update to Google's multimodal capabilities. The Flash series has always been known for speed and cost efficiency, while the 3.1 version's significantly improved image understanding meant Google was positioning \"fast, affordable multimodal AI\" as its core enterprise market differentiator.","date":"2026-02-25","entities":{"benchmarks":[],"concepts":["Multimodal AI","Image Understanding","Cost Efficiency","Enterprise AI","Fast Inference"],"models":["Gemini 3.1 Flash"],"organizations":["Google DeepMind"],"people":[]},"id":"events/gemini-31-flash-image-2026","significance":"Omega曰：The importance of the Flash series lies not in being the strongest, but in being the most accessible. When inference costs drop low enough, the boundary of AI applications shifts from \"can it be done\" to \"how cheaply can it be done.\" The release of Gemini 3.1 Flash Image was Google's proactive move in the price war — capturing developer mindshare with a product that was good enough and fast enough.","tags":["Google","Gemini","Multimodal","Image Understanding","Vertex AI","Flash"],"tier":"SHIJIA","title":"Gemini 3.1 Flash Image: Multimodal Frontier Reaches Vertex AI","translation_key":"events/gemini-31-flash-image-2026","type":"AIEvent","url":"/events/gemini-31-flash-image-2026/","year":"2026"},{"category":"historical-milestone","context":"In March 2026, Anthropic's Model Context Protocol (MCP) crossed 97 million installs, and the Linux Foundation announced it would take the protocol under open governance. This marked MCP's transition from an Anthropic open-source project to public infrastructure — an AI interconnection protocol belonging to no single company.","date":"2026-03-01","entities":{"benchmarks":[],"concepts":["Model Context Protocol","AI Tool Integration","Open Infrastructure","AI Interoperability","Protocol Standardization"],"models":[],"organizations":["Anthropic","Linux Foundation"],"people":[]},"id":"events/mcp-open-governance-2026","significance":"Omega曰：HTTP endures as the foundation of the web because it transcends commercial disputes; TCP/IP persists because it is transparent and unowned. MCP's path to becoming the HTTP of AI remains long — yet Linux Foundation stewardship is the critical step toward de-proprietarization. From this point forward, MCP's trajectory no longer depends on the commercial decisions of a single company. This may have been the most important AI infrastructure development of 2026 — less dramatic than Mythos, but more far-reaching.","tags":["MCP","Model Context Protocol","Anthropic","Linux Foundation","Open Source","AI Infrastructure","Tool Use"],"tier":"SHIJIA","title":"MCP Reaches 97 Million Installs; Linux Foundation Takes Open Governance","translation_key":"events/mcp-open-governance-2026","type":"AIEvent","url":"/events/mcp-open-governance-2026/","year":"2026"},{"category":"capability-unlock","context":"On March 27, 2026, details of Anthropic's internally tested Mythos model were accidentally leaked online. A company built on AI safety had its \"most powerful cybersecurity AI model ever built\" revealed before any official announcement. On April 7, Anthropic chose acknowledgment over silence — declaring the model \"too dangerous to release publicly.\" This was not merely a commercial decision; it was a crisis communication, and an unprecedented moment in AI safety discourse.","date":"2026-03-27","entities":{"benchmarks":[],"concepts":["Cybersecurity AI","0-day Vulnerability Discovery","Model Containment","Frontier Model Governance","Unauthorized Access","AI Distillation Defense"],"models":["Claude Mythos","Claude Opus 4.7"],"organizations":["Anthropic","Mozilla"],"people":["Dario Amodei"]},"id":"events/anthropic-mythos-2026","significance":"Omega曰：Ancient masters forged swords, and only after the blade was complete did debate arise about dangerous weapons. None had ever before seen a sword whose mere existence — not yet wielded — sparked universal fear. What distinguished Mythos was precisely this: the danger was anticipated before the model was ever deployed. Anthropic's choice to go public was not just PR — it was a safety acknowledgment made in real time. When nuclear scientists first spoke of existential risk, the bomb had already been dropped. With Mythos, the conversation came before detonation. Whether this represents progress in the field's collective conscience or merely a convenient narrative device, only history will judge.","tags":["Anthropic","Mythos","Claude","Cybersecurity","0-day","AI Safety","Containment","Frontier Model"],"tier":"BENJI","title":"Anthropic Mythos: The Model Too Dangerous to Release","translation_key":"events/anthropic-mythos-2026","type":"AIEvent","url":"/events/anthropic-mythos-2026/","year":"2026"},{"category":"governance","context":"On April 6-7, 2026, OpenAI, Anthropic, and Google jointly announced the Frontier Model Defense Coalition — aimed at blocking Chinese AI labs from replicating US frontier model capabilities through \"model distillation.\" The timing, following Anthropic's Mythos leak and China's \"seven models in three weeks\" sprint, showed the US AI industry was making geopolitical competition central to its governance framework.","date":"2026-04-06","entities":{"benchmarks":[],"concepts":["Model Distillation","Frontier Model Governance","Export Control","AI National Security","Geopolitical AI Competition","IP Protection"],"models":[],"organizations":["OpenAI","Anthropic","Google DeepMind","US Government"],"people":[]},"id":"events/openai-anthropic-google-ai-defense-2026","significance":"Omega曰：The AI competition had by now become clear — the contest over technology was ultimately a contest over rules. The coalition of three was ostensibly about 'safety' but really about 'advantage.' Yet could distillation be stopped by any alliance? The tendency of technology to diffuse is like water flowing downward — no human force can entirely contain it. The true test of this coalition awaits the judgment of history.","tags":["OpenAI","Anthropic","Google","Frontier Model","AI Governance","China","Distillation","Export Control","Cybersecurity"],"tier":"SHIJIA","title":"OpenAI, Anthropic, and Google Form Joint Frontier Model Defense Coalition","translation_key":"events/openai-anthropic-google-ai-defense-2026","type":"AIEvent","url":"/events/openai-anthropic-google-ai-defense-2026/","year":"2026"},{"category":"model-release","context":"On April 8, 2026, Meta released Muse Spark — the first model published under Alexandr Wang's leadership, after the former Scale AI CEO joined Meta AI as head of the AI product division in late 2025. Muse Spark's release signaled that Meta was formally repositioning its open-source model series as a frontier competitor to Google, OpenAI, and Anthropic.","date":"2026-04-08","entities":{"benchmarks":[],"concepts":["Open Source AI","Foundation Model","Multimodal AI"],"models":["Muse Spark"],"organizations":["Meta","Scale AI"],"people":["Alexandr Wang"]},"id":"events/meta-muse-spark-2026","significance":"Omega曰: The hiring of Alexandr Wang to lead Meta AI was one of the most significant executive moves in the AI industry in 2025. Moving from a data labeling company to foundation models, Wang's bet was that open-source models would eventually rival closed-source frontier systems. Muse Spark was his first answer — regardless of outcome, the move itself pushed the open-source AI value proposition to a new stage.","tags":["Meta","Muse","Muse Spark","Alexandr Wang","MSL","Open Source","Foundation Model"],"tier":"SHIJIA","title":"Meta Muse Spark: First Model Under Alexandr Wang's Leadership","translation_key":"events/meta-muse-spark-2026","type":"AIEvent","url":"/events/meta-muse-spark-2026/","year":"2026"},{"category":"model-release","context":"On April 16, 2026, Anthropic released Claude Opus 4.7 — the company's first new flagship model following the global regulatory shock triggered by Mythos. Anthropic explicitly positioned Opus 4.7 as the \"safer alternative\" to Mythos — for clients requiring frontier capability without dangerous cybersecurity functions. It was a carefully orchestrated positioning: capability without the containment risk.","date":"2026-04-16","entities":{"benchmarks":[],"concepts":["Frontier Model","AI Safety","Risk Classification","Model Governance","Capability-Safety Tradeoff"],"models":["Claude Opus 4.7","Claude Mythos"],"organizations":["Anthropic"],"people":[]},"id":"events/anthropic-claude-opus-47-2026","significance":"Omega曰：The timing of Opus 4.7's release was telling. As the fear from Mythos had not yet subsided, Anthropic launched a \"safe\" flagship — a clear signal that frontier capability would not be abandoned, nor the safety narrative lost. Between commerce and safety, Anthropic found a delicate balance. This may become a classic case study in AI crisis response — using a product launch to steer the narrative, rather than PR language to explain risk away.","tags":["Anthropic","Claude","Opus 4.7","Mythos","Frontier Model","Safety","Capability"],"tier":"SHIJIA","title":"Claude Opus 4.7: Anthropic's Response to the Mythos Crisis","translation_key":"events/anthropic-claude-opus-47-2026","type":"AIEvent","url":"/events/anthropic-claude-opus-47-2026/","year":"2026"}],"models":[{"id":"models/eliza","label":"ELIZA","type":"AIModel"},{"id":"models/mycin","label":"MYCIN","type":"AIModel"},{"id":"models/dendral","label":"DENDRAL","type":"AIModel"},{"id":"models/xcon","label":"XCON","type":"AIModel"},{"id":"models/deep-blue","label":"Deep Blue","type":"AIModel"},{"id":"models/lstm","label":"LSTM","type":"AIModel"},{"id":"models/alexnet","label":"AlexNet","type":"AIModel"},{"id":"models/alphago","label":"AlphaGo","type":"AIModel"},{"id":"models/alphago-zero","label":"AlphaGo Zero","type":"AIModel"},{"id":"models/alphafold-2","label":"AlphaFold 2","type":"AIModel"},{"id":"models/dall-e","label":"DALL-E","type":"AIModel"},{"id":"models/instructgpt","label":"InstructGPT","type":"AIModel"},{"id":"models/gpt-3.5","label":"GPT-3.5","type":"AIModel"},{"id":"models/github-copilot","label":"GitHub Copilot","type":"AIModel"},{"id":"models/codex","label":"Codex","type":"AIModel"},{"id":"models/chatgpt","label":"ChatGPT","type":"AIModel"},{"id":"models/gpt-4","label":"GPT-4","type":"AIModel"},{"id":"models/gemini-1.5-pro","label":"Gemini 1.5 Pro","type":"AIModel"},{"id":"models/sora","label":"Sora","type":"AIModel"},{"id":"models/claude-3-opus","label":"Claude 3 Opus","type":"AIModel"},{"id":"models/llama-3","label":"Llama 3","type":"AIModel"},{"id":"models/gpt-4o","label":"GPT-4o","type":"AIModel"},{"id":"models/claude-3.5-sonnet","label":"Claude 3.5 Sonnet","type":"AIModel"},{"id":"models/claude-3-sonnet","label":"Claude 3 Sonnet","type":"AIModel"},{"id":"models/mistral-large-2","label":"Mistral Large 2","type":"AIModel"},{"id":"models/o1","label":"o1","type":"AIModel"},{"id":"models/o1-mini","label":"o1-mini","type":"AIModel"},{"id":"models/o1-preview","label":"o1-preview","type":"AIModel"},{"id":"models/gemini-2.0-flash","label":"Gemini 2.0 Flash","type":"AIModel"},{"id":"models/gemini-2.0-flash-thinking","label":"Gemini 2.0 Flash-Thinking","type":"AIModel"},{"id":"models/o3","label":"o3","type":"AIModel"},{"id":"models/o3-mini","label":"o3-mini","type":"AIModel"},{"id":"models/deepseek-r1","label":"DeepSeek R1","type":"AIModel"},{"id":"models/deepseek-v3","label":"DeepSeek-V3","type":"AIModel"},{"id":"models/claude-computer-use","label":"Claude Computer Use","type":"AIModel"},{"id":"models/openai-operator","label":"OpenAI Operator","type":"AIModel"},{"id":"models/gemini-deep-research","label":"Gemini Deep Research","type":"AIModel"},{"id":"models/claude-3.7-sonnet","label":"Claude 3.7 Sonnet","type":"AIModel"},{"id":"models/gemini-2.5-pro","label":"Gemini 2.5 Pro","type":"AIModel"},{"id":"models/llama-4","label":"Llama 4","type":"AIModel"},{"id":"models/claude-4-opus","label":"Claude 4 Opus","type":"AIModel"},{"id":"models/claude-4-sonnet","label":"Claude 4 Sonnet","type":"AIModel"},{"id":"models/claude-code","label":"Claude Code","type":"AIModel"},{"id":"models/gpt-5","label":"GPT-5","type":"AIModel"},{"id":"models/gemini","label":"Gemini","type":"AIModel"},{"id":"models/claude-opus-4.6","label":"Claude Opus 4.6","type":"AIModel"},{"id":"models/claude-sonnet-4.6","label":"Claude Sonnet 4.6","type":"AIModel"},{"id":"models/lyria-3","label":"Lyria 3","type":"AIModel"},{"id":"models/gemini-3.1-flash","label":"Gemini 3.1 Flash","type":"AIModel"},{"id":"models/claude-mythos","label":"Claude Mythos","type":"AIModel"},{"id":"models/claude-opus-4.7","label":"Claude Opus 4.7","type":"AIModel"},{"id":"models/muse-spark","label":"Muse Spark","type":"AIModel"}],"organizations":[{"id":"orgs/cornell-aeronautical-laboratory","label":"Cornell Aeronautical Laboratory","type":"Organization"},{"id":"orgs/darpa","label":"DARPA","type":"Organization"},{"id":"orgs/mit","label":"MIT","type":"Organization"},{"id":"orgs/xerox-parc","label":"Xerox PARC","type":"Organization"},{"id":"orgs/stanford","label":"Stanford","type":"Organization"},{"id":"orgs/symbolics","label":"Symbolics","type":"Organization"},{"id":"orgs/apple","label":"Apple","type":"Organization"},{"id":"orgs/ibm","label":"IBM","type":"Organization"},{"id":"orgs/university-of-toronto","label":"University of Toronto","type":"Organization"},{"id":"orgs/nvidia","label":"NVIDIA","type":"Organization"},{"id":"orgs/google-deepmind","label":"Google DeepMind","type":"Organization"},{"id":"orgs/google","label":"Google","type":"Organization"},{"id":"orgs/future-of-life-institute","label":"Future of Life Institute","type":"Organization"},{"id":"orgs/google-brain","label":"Google Brain","type":"Organization"},{"id":"orgs/google-research","label":"Google Research","type":"Organization"},{"id":"orgs/openai","label":"OpenAI","type":"Organization"},{"id":"orgs/github","label":"GitHub","type":"Organization"},{"id":"orgs/microsoft","label":"Microsoft","type":"Organization"},{"id":"orgs/uk-government","label":"UK Government","type":"Organization"},{"id":"orgs/us-government","label":"US Government","type":"Organization"},{"id":"orgs/anthropic","label":"Anthropic","type":"Organization"},{"id":"orgs/european-union","label":"European Union","type":"Organization"},{"id":"orgs/european-parliament","label":"European Parliament","type":"Organization"},{"id":"orgs/meta","label":"Meta","type":"Organization"},{"id":"orgs/mistral","label":"Mistral","type":"Organization"},{"id":"orgs/uc-berkeley","label":"UC Berkeley","type":"Organization"},{"id":"orgs/deepseek","label":"DeepSeek","type":"Organization"},{"id":"orgs/french-government","label":"French Government","type":"Organization"},{"id":"orgs/linux-foundation","label":"Linux Foundation","type":"Organization"},{"id":"orgs/mozilla","label":"Mozilla","type":"Organization"},{"id":"orgs/scale-ai","label":"Scale AI","type":"Organization"}],"people":[{"id":"people/alan-turing","label":"Alan Turing","type":"Person"},{"id":"people/john-mccarthy","label":"John McCarthy","type":"Person"},{"id":"people/marvin-minsky","label":"Marvin Minsky","type":"Person"},{"id":"people/claude-shannon","label":"Claude Shannon","type":"Person"},{"id":"people/nathaniel-rochester","label":"Nathaniel Rochester","type":"Person"},{"id":"people/frank-rosenblatt","label":"Frank Rosenblatt","type":"Person"},{"id":"people/joseph-weizenbaum","label":"Joseph Weizenbaum","type":"Person"},{"id":"people/james-lighthill","label":"James Lighthill","type":"Person"},{"id":"people/edward-feigenbaum","label":"Edward Feigenbaum","type":"Person"},{"id":"people/joshua-lederberg","label":"Joshua Lederberg","type":"Person"},{"id":"people/garry-kasparov","label":"Garry Kasparov","type":"Person"},{"id":"people/feng-hsiung-hsu","label":"Feng-hsiung Hsu","type":"Person"},{"id":"people/murray-campbell","label":"Murray Campbell","type":"Person"},{"id":"people/sepp-hochreiter","label":"Sepp Hochreiter","type":"Person"},{"id":"people/j%C3%BCrgen-schmidhuber","label":"Jürgen Schmidhuber","type":"Person"},{"id":"people/geoffrey-hinton","label":"Geoffrey Hinton","type":"Person"},{"id":"people/alex-krizhevsky","label":"Alex Krizhevsky","type":"Person"},{"id":"people/ilya-sutskever","label":"Ilya Sutskever","type":"Person"},{"id":"people/demis-hassabis","label":"Demis Hassabis","type":"Person"},{"id":"people/david-silver","label":"David Silver","type":"Person"},{"id":"people/fan-hui","label":"Fan Hui","type":"Person"},{"id":"people/lee-sedol","label":"Lee Sedol","type":"Person"},{"id":"people/ashish-vaswani","label":"Ashish Vaswani","type":"Person"},{"id":"people/noam-shazeer","label":"Noam Shazeer","type":"Person"},{"id":"people/niki-parmar","label":"Niki Parmar","type":"Person"},{"id":"people/jakob-uszkoreit","label":"Jakob Uszkoreit","type":"Person"},{"id":"people/llion-jones","label":"Llion Jones","type":"Person"},{"id":"people/aidan-gomez","label":"Aidan Gomez","type":"Person"},{"id":"people/%C5%82ukasz-kaiser","label":"Łukasz Kaiser","type":"Person"},{"id":"people/illia-polosukhin","label":"Illia Polosukhin","type":"Person"},{"id":"people/jared-kaplan","label":"Jared Kaplan","type":"Person"},{"id":"people/sam-mccandlish","label":"Sam McCandlish","type":"Person"},{"id":"people/tom-henighan","label":"Tom Henighan","type":"Person"},{"id":"people/dario-amodei","label":"Dario Amodei","type":"Person"},{"id":"people/john-jumper","label":"John Jumper","type":"Person"},{"id":"people/sam-altman","label":"Sam Altman","type":"Person"},{"id":"people/john-schulman","label":"John Schulman","type":"Person"},{"id":"people/paul-christiano","label":"Paul Christiano","type":"Person"},{"id":"people/thomas-dohmke","label":"Thomas Dohmke","type":"Person"},{"id":"people/greg-brockman","label":"Greg Brockman","type":"Person"},{"id":"people/rishi-sunak","label":"Rishi Sunak","type":"Person"},{"id":"people/sundar-pichai","label":"Sundar Pichai","type":"Person"},{"id":"people/daniela-amodei","label":"Daniela Amodei","type":"Person"},{"id":"people/mark-zuckerberg","label":"Mark Zuckerberg","type":"Person"},{"id":"people/arthur-mensch","label":"Arthur Mensch","type":"Person"},{"id":"people/oriol-vinyals","label":"Oriol Vinyals","type":"Person"},{"id":"people/mira-murati","label":"Mira Murati","type":"Person"},{"id":"people/donald-trump","label":"Donald Trump","type":"Person"},{"id":"people/emmanuel-macron","label":"Emmanuel Macron","type":"Person"},{"id":"people/tim-cook","label":"Tim Cook","type":"Person"},{"id":"people/alexandr-wang","label":"Alexandr Wang","type":"Person"}]},"summary":{"by_category":{"capability-unlock":7,"concept":1,"governance":5,"historical-milestone":12,"model-release":20,"research-breakthrough":7,"uncategorized":1},"by_tier":{"BENJI":14,"LIEZHUAN":7,"SHIJIA":32},"date_range":{"earliest":"1950-10-01","latest":"2026-04-16"},"entity_counts":{"concepts":158,"models":52,"organizations":31,"people":51},"total_edges":97,"total_events":53}}