Capabilities · dependencies · directions

From early neural networks to ASI

This map shows what AI systems have learned to do and how later capabilities grew from earlier ones. The main line connects major capabilities—from perception and knowledge transfer to reasoning and autonomous work. Large nodes mark the arrival of a new class of capability, branches show how it continued to develop, and the lower scale records specific models, research and demonstrations by date. AGI and ASI are shown as possible future frontiers, not scheduled releases.

wheel to zoom · drag to pan · click a branch to dive in
14%
AI capability line
ORIGINS · 194319861943·12The logical neuron1958·11The perceptron1986·10Backpropagation1998·11LeNet and convolutional networks2009ImageNet2012·12AlexNet2013·01Word2vec2014·06Adversarial networks2014·09Sequence to Sequence2015·12ResNet2016·01AlphaGo2017·06Attention Is All You Need2018·06GPT2018·10BERT2019·02GPT-22019·10T52020·01Scaling laws2020·05Retrieval (RAG)2020·05GPT-3, 175B2020·10Vision Transformer2021·01CLIP2021·01Mixture of experts2021·07AlphaFold 22022200 million protein structures2022·01Chain of thought2022·03InstructGPT2022·03Chinchilla2022·06Emergent abilities2022·09Whisper2022·11ChatGPT2022·12Constitutional AI20232.2M predicted crystals202336 materials in 17 days2023·03GPT-42023·04Segment Anything2023·09GPT-4V2023·12Gemini2023·12-11Mixtral 8×7B2024Llama 3.1 405B2024One and a half percent of world electricity2024AlphaFold 3: molecular interactions2024SWE-bench2024Gemini 1.5 Pro2024Reasoning as a separate mode2024·03Claude 32024·05GPT-4o2024·09o12024·09-20Three Mile Island restart2024·11Model Context Protocol2025DeepSeek-V3 and R12025Qwen 32025Repository-scale context2025·03An agent platform2025·03Autonomous task horizon2025·03Gemini 2.52025·04Agent2Agent2025·08Genie 32025·10Designing agent workflows2025·11The road to olympiad gold2025·12MCP handed to a foundation2026TPU v6e, Trainium3, Maia 2002026Repo-scale context as default2026Agents inside CI2026Some forecasts moved closer2026The spread stays enormous2026The academic objection2026Dangerous capability evals2026·01-06NVIDIA Vera Rubin2026·01-27Figure Helix 022026·01-29Project Genie2026·03Rubin in production2026·03-11Meta MTIA 3002026·03-12A2A 1.02026·04-26Sora app shut down2026·04-29Figure 03: 350 robots produced2026·059.8 GW of nuclear deals2026·05Gemini Omni2026·05-08Long tasks cross the threshold2026·05-31Cosmos 3: physical AI in one model2026·05-31NVIDIA Cosmos 32026·06565 TWh in a year2026·06-01MiniMax M32026·06-02US order on frontier-model security2026·06-09Claude Fable 52026·06-30Claude Sonnet 52026·07Longer work in a product2026·07AxiomProver: 42 out of 422026·07-09GPT-5.6 Sol2026·07-27Kimi K3: open weights published2026·08-02AI Act: the next deadline2026·12-02Labelling for pre-existing systems2027·12-02Annex III rulesPerception+1Transfer+2Context+2Instructions+1Multimodality+2Reasoning+1Autonomy+1ObstacleLearning from experienceToday a model does not remember yesterday's work: everything it knows is frozen at training time. We need a system that learns while working without forgetting what it already knew.ObstacleReliable autonomyBefore trusting AI with an operating theatre, a power plant or a company's books, you must be able to prove it will not fail silently. No such method exists today.ObstacleActing in the physical worldA model can describe an action, but the real world requires continuous perception, physical reasoning and correction before errors become irreversible.ObstacleCompute and power bottleneckAI capability is constrained by more than algorithms. Chips, memory, data centres, electrical grids and cooling determine which models can be trained and how many runs the economy can sustain.AGICONVERGENCEASIHYPOTHESISMemory and experienceWorld modelsDigital agents