[ { "id": "power", "kind": "capability", "label": "Power and grid buildout", "status": "built", "lane": "energy", "description": "Electricity, grid connections and generation for datacentres.", "evidence": "The IEA projects datacentre electricity use more than doubling, from about 415 TWh in 2024 to about 945 TWh by 2030. New transmission lines take four to eight years, and around 20% of planned projects could be delayed.", "sources": [ "iea", "epoch2030" ], "needs": [ { "id": "firmpower", "link": "soft" } ], "derived": { "wave": 1, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "chips", "kind": "capability", "label": "Chip and memory supply", "status": "built", "lane": "hw", "description": "Enough advanced chips, packaging and high-bandwidth memory.", "evidence": "Epoch AI names chip packaging and high-bandwidth memory as the next limit after power. The IEA notes China refines about 99% of the gallium used in advanced chips.", "sources": [ "epoch2030", "iea" ], "needs": [], "derived": { "wave": 1, "cluster": "Compute governance and geopolitics", "leverage": 0 } }, { "id": "capex", "kind": "capability", "label": "Datacentre capex", "status": "built", "lane": "capital", "description": "The capital spent building datacentres before the revenue arrives.", "evidence": "Moody’s expects US hyperscaler capex above $700bn in 2026, about six times 2022, and says hyperscalers see underinvestment as the bigger risk.", "sources": [ "moodys" ], "needs": [ { "id": "vcconc", "link": "soft" }, { "id": "jevons", "link": "soft" } ], "derived": { "wave": 1, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "compute", "kind": "capability", "label": "Compute supply", "status": "built", "lane": "hw", "description": "Who owns and supplies the substrate everything runs on.", "evidence": "Epoch AI expects power to be the first constraint to bind on scaling, followed by chip packaging and memory.", "sources": [ "epoch2030" ], "needs": [ { "id": "capex", "link": "hard" }, { "id": "chips", "link": "hard" }, { "id": "power", "link": "hard" }, { "id": "hbm", "link": "hard" } ], "derived": { "wave": 2, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "revenue", "kind": "stage", "label": "Frontier lab revenue", "status": "built", "lane": "capital", "description": "What the frontier labs earn from subscriptions and API use.", "evidence": "Epoch AI estimates OpenAI and Anthropic’s combined run rate rose from about $30bn at the start of 2026 to about $105bn by August.", "sources": [ "epochrev" ], "needs": [ { "id": "chat", "link": "hard" }, { "id": "adopt", "link": "soft" }, { "id": "jevons", "link": "soft" } ], "watch": "Combined run rate of OpenAI and Anthropic: about $105bn in August 2026.", "when": { "type": "here", "from": 2023, "to": null, "tail": null, "note": "Frontier labs earned subscription and API revenue from 2023; OpenAI and Anthropic’s combined run rate reached about $105bn by August 2026.", "sources": [ "epochrev" ] }, "derived": { "wave": 5, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "data", "kind": "capability", "label": "Training data supply", "status": "built", "lane": "data", "description": "Enough good data to keep improving models.", "evidence": "Epoch estimates the stock of public human-written text runs out sometime between 2026 and 2032. The fallbacks are synthetic data, private data and training on the same data more times.", "sources": [ "datawall" ], "needs": [ { "id": "datarights", "link": "soft" }, { "id": "synth", "link": "soft" } ], "derived": { "wave": 1, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "pretrain", "kind": "capability", "label": "Pretraining at scale", "status": "built", "lane": "algo", "description": "Large models trained on broad data, the base every other capability builds on.", "evidence": "Epoch AI finds training compute can keep growing through 2030, with power, chips, data and latency as the four limits.", "sources": [ "epoch2030" ], "needs": [ { "id": "compute", "link": "hard" }, { "id": "data", "link": "hard" }, { "id": "algoeff", "link": "soft" } ], "derived": { "wave": 3, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "verifiers", "kind": "capability", "label": "Reliable verifiers", "status": "built", "lane": "measure", "description": "A trustworthy signal for whether an output is right.", "evidence": "Jason Wei’s verifier’s law: how easy a task is to train AI on is proportional to how verifiable it is. Every task that is solvable and easy to verify will be solved.", "sources": [ "wei" ], "needs": [], "derived": { "wave": 1, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "rlscale", "kind": "capability", "label": "RL on checkable tasks", "status": "built", "lane": "algo", "description": "Reinforcement learning on tasks with answers that can be checked automatically.", "evidence": "Reasoning models came from training on maths and code, where answers can be checked. Verifier’s law predicts those domains fall first.", "sources": [ "wei" ], "needs": [ { "id": "pretrain", "link": "hard" }, { "id": "verifiers", "link": "hard" } ], "derived": { "wave": 4, "cluster": "Agents at work", "leverage": 0 } }, { "id": "reason", "kind": "stage", "label": "Reasoning models", "status": "built", "lane": "models", "description": "Models that think before they answer. Single steps get much smarter.", "evidence": "Level 2, \"Reasoners\", on the five-level scale toward AGI that OpenAI reportedly uses internally.", "sources": [ "openai5" ], "needs": [ { "id": "pretrain", "link": "hard" }, { "id": "rlscale", "link": "hard" }, { "id": "testtime", "link": "hard" } ], "watch": "Price and speed of frontier reasoning at a fixed benchmark score.", "when": { "type": "here", "from": 2024.7, "to": null, "tail": null, "note": "OpenAI’s o1 reasoning model, September 2024.", "sources": [] }, "derived": { "wave": 5, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "multimodal", "kind": "capability", "label": "Multimodal perception", "status": "built", "lane": "models", "description": "Reading images, audio and video as well as text.", "evidence": "On Karpathy’s list of what agents still lack before they can work like an intern.", "sources": [ "karpathy" ], "needs": [ { "id": "pretrain", "link": "hard" } ], "derived": { "wave": 4, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "tools", "kind": "capability", "label": "Tool use and code execution", "status": "built", "lane": "models", "description": "Models act on systems instead of describing actions.", "evidence": "MCP, now a Linux Foundation project, has become the standard way to connect models to tools and data.", "sources": [ "a2a" ], "needs": [], "derived": { "wave": 1, "cluster": "Agents at work", "leverage": 0 } }, { "id": "contlearn", "kind": "capability", "label": "Weight-level continual learning", "status": "emerging", "lane": "algo", "description": "Lessons move from the context window into the model itself.", "evidence": "\"There is no pathway from context to weights\" yet. Titans and Nested Learning are early attempts at models that update while they run.", "sources": [ "overhang", "dwarkeshcl" ], "needs": [], "derived": { "wave": 6, "cluster": "Agents at work", "leverage": 0.46 } }, { "id": "generalise", "kind": "capability", "label": "Robust generalisation", "status": "emerging", "lane": "algo", "description": "Doing as well on messy real work as on benchmarks.", "evidence": "Sutskever says the field is moving from an age of scaling to an age of research, because people learn from far less data than models need and AI must learn as efficiently.", "sources": [ "sutskever" ], "needs": [], "derived": { "wave": 6, "cluster": "Agents at work", "leverage": 0.46 } }, { "id": "reliab", "kind": "capability", "label": "Reliability on repeat", "status": "emerging", "lane": "models", "description": "Getting the same task right every time, not once.", "evidence": "On τ-bench, GPT-4o succeeded on over 60% of retail tasks but on under 25% when it had to succeed eight times in a row. Success rates also decay with task length.", "sources": [ "taubench", "halflife" ], "needs": [ { "id": "abstain", "link": "soft" } ], "derived": { "wave": 6, "cluster": "Agents at work", "leverage": 2.32 } }, { "id": "cost", "kind": "capability", "label": "Hardware cost per token", "status": "built", "lane": "hw", "description": "Chips, memory and serving that make each token cheaper to produce.", "evidence": "State of AI 2026 finds the fall in the price of a fixed capability is driven as much by algorithms and distillation as by cheaper chips, so hardware is only part of the story.", "sources": [ "soaibiz" ], "needs": [ { "id": "chips", "link": "hard" } ], "derived": { "wave": 2, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "openw", "kind": "capability", "label": "Open-weight models", "status": "built", "lane": "models", "description": "Capable models anyone can download, run and modify.", "evidence": "Epoch AI finds Chinese models have trailed the US frontier by about seven months on average since 2023, and several of the strongest Chinese releases are open-weight.", "sources": [ "epochchina" ], "needs": [ { "id": "pretrain", "link": "hard" } ], "derived": { "wave": 4, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "air", "kind": "stage", "label": "Superhuman AI researcher", "status": "speculative", "lane": "models", "description": "AI that does AI research better than the best human researchers.", "evidence": "AI 2027 puts a superhuman coder, then a superhuman AI researcher, ahead of every other milestone, because each speeds up the rest. Its takeoff forecast equates the superhuman researcher with about a 25x speed-up in AI research; AI Futures measured 1.04 to 2x in 2026.", "sources": [ "ai2027", "aifut" ], "needs": [ { "id": "aiuplift", "link": "hard" }, { "id": "autocoder", "link": "hard" }, { "id": "verifiers", "link": "hard" }, { "id": "compute", "link": "hard" }, { "id": "talent", "link": "soft" } ], "watch": "Speed-up of AI research. AI 2027 equates a superhuman researcher with about 25x.", "when": { "type": "est", "from": 2029, "to": 2034, "tail": null, "note": "My estimate, a year or more after an automated coder. AI 2027 put the two months apart; its authors have since pushed their timelines back.", "sources": [ "ai2027", "ai2027slip" ] }, "derived": { "wave": 9, "cluster": "Oversight and alignment", "leverage": 0.76 } }, { "id": "vcconc", "kind": "stage", "label": "Capital concentrated in AI", "status": "built", "lane": "capital", "description": "Venture and growth capital flowing overwhelmingly into AI companies.", "evidence": "PitchBook counts 86% of US venture funding going to AI in the first half of 2026, and OpenAI and Anthropic alone took 43% of global startup funding. PitchBook warns that a market this dependent on one theme faces a broad correction if returns disappoint.", "sources": [ "pitchbook" ], "needs": [], "watch": "AI share of venture funding: 86% of US venture dollars in the first half of 2026.", "when": { "type": "here", "from": 2023, "to": null, "tail": null, "note": "AI’s share of venture funding climbed from 2023 to 86% of US venture in the first half of 2026.", "sources": [ "pitchbook" ] }, "derived": { "wave": 1, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "fintools", "kind": "capability", "label": "AI research tools in investing", "status": "built", "lane": "finance", "description": "Analysts and investors using models to read filings, data rooms and research.", "evidence": "AIMA found 86% of hedge fund managers gave staff generative AI tools by the end of 2023. Hebbia says 30% of the 50 largest asset managers use its document AI.", "sources": [ "aima", "hebbia" ], "needs": [ { "id": "chat", "link": "hard" }, { "id": "connectors", "link": "hard" }, { "id": "airgap", "link": "soft" }, { "id": "ambientdocs", "link": "soft" } ], "derived": { "wave": 5, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "quantgen", "kind": "capability", "label": "AI-generated trading signals", "status": "emerging", "lane": "finance", "description": "Agents that propose, code and test investment ideas like a quant research team.", "evidence": "Man Group’s AlphaGPT generates hypotheses, writes the code and evaluates the results, and Man says it has produced signals that pass the same thresholds as human research.", "sources": [ "alphagpt" ], "needs": [ { "id": "codeagents", "link": "hard" }, { "id": "verifiers", "link": "hard" } ], "derived": { "wave": 8, "cluster": "Science and medicine", "leverage": 0.2 } }, { "id": "agentictrade", "kind": "capability", "label": "Agents that trade for investors", "status": "emerging", "lane": "finance", "description": "Consumers letting their own AI agents place trades.", "evidence": "Robinhood lets users connect third-party agents to a walled-off \"agentic\" account, and says the user bears all the risk of the agent’s trades. Whether such agents count as investment advisers is unsettled.", "sources": [ "robinhood" ], "needs": [ { "id": "payments", "link": "hard" }, { "id": "identity", "link": "hard" }, { "id": "liability", "link": "soft" } ], "derived": { "wave": 8, "cluster": "Trust plumbing and the agent economy", "leverage": 0.2 } }, { "id": "finreg", "kind": "capability", "label": "Financial regulators testing AI", "status": "emerging", "lane": "finance", "description": "Supervisors building ways to test and approve AI in regulated finance.", "evidence": "The FCA’s AI Live Testing programme began in December 2025 with NatWest, Monzo, Santander and Scottish Widows among the first cohort.", "sources": [ "fca" ], "needs": [], "derived": { "wave": 7, "cluster": "Science and medicine", "leverage": 0.2 } }, { "id": "marketstab", "kind": "capability", "label": "Protection against AI herding", "status": "emerging", "lane": "finance", "description": "Keeping many firms on similar models from moving markets in lockstep.", "evidence": "The Bank of England warned in April 2025 that firms using similar AI models could take \"increasingly correlated positions\" and amplify shocks.", "sources": [ "boeherd" ], "needs": [ { "id": "finreg", "link": "hard" } ], "derived": { "wave": 8, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "aimarkets", "kind": "stage", "label": "AI-run capital allocation", "status": "speculative", "lane": "finance", "description": "Funds and markets where agents make most allocation decisions.", "evidence": "", "sources": [], "needs": [ { "id": "quantgen", "link": "hard" }, { "id": "agentictrade", "link": "hard" }, { "id": "finreg", "link": "hard" }, { "id": "marketstab", "link": "soft" }, { "id": "xtrust", "link": "soft" }, { "id": "oversight", "link": "soft" } ], "watch": "Share of assets where an AI system makes the allocation decision.", "when": { "type": "est", "from": 2030, "to": 2040, "tail": null, "note": "My estimate. Agents already trade in walled-off retail accounts; allocation at scale waits on regulators.", "sources": [ "robinhood" ] }, "derived": { "wave": 9, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "algoeff", "kind": "capability", "label": "Algorithmic efficiency", "status": "built", "lane": "algo", "description": "The compute needed for a given capability keeps falling.", "evidence": "Epoch AI found the compute needed to reach a given language-model performance halves roughly every eight months, with a 95% range of 5 to 14 months.", "sources": [ "epochalgo" ], "needs": [], "derived": { "wave": 1, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "testtime", "kind": "capability", "label": "Test-time compute", "status": "built", "lane": "algo", "description": "Letting a model think longer when answering, instead of only training it bigger.", "evidence": "As gains from bigger pretraining runs slowed in late 2024, labs turned to reasoning at inference time. OpenAI’s Noam Brown said 20 seconds of thinking in a poker hand matched scaling the model up 100,000 times.", "sources": [ "wall" ], "needs": [ { "id": "pretrain", "link": "hard" } ], "derived": { "wave": 4, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "synth", "kind": "capability", "label": "Synthetic data", "status": "emerging", "lane": "data", "description": "Model-generated data used to train the next models.", "evidence": "Shumailov et al. showed in Nature that training on recursively generated data degrades into \"model collapse\" unless the data is carefully filtered. Checked self-generated data works: AlphaProof generated over 100 million proof steps by exploring on its own.", "sources": [ "collapse", "experience" ], "needs": [ { "id": "verifiers", "link": "soft" } ], "derived": { "wave": 1, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "interp", "kind": "capability", "label": "Interpretability", "status": "emerging", "lane": "measure", "description": "Seeing what a model is computing, not just what it outputs.", "evidence": "Anthropic’s 2025 attribution graphs showed Claude planning rhyming words ahead and reasoning in steps internally, but gave useful insight for only about a quarter of the prompts tried.", "sources": [ "circuits" ], "needs": [], "derived": { "wave": 8, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "evalaware", "kind": "capability", "label": "Tests models cannot game", "status": "emerging", "lane": "measure", "description": "Evaluations that still hold when models know they are being tested.", "evidence": "Apollo Research reports frontier models increasingly recognise evaluation settings, which weakens confidence that tested behaviour matches deployed behaviour.", "sources": [ "apollo" ], "needs": [], "derived": { "wave": 8, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "hbm", "kind": "capability", "label": "Memory bandwidth and supply", "status": "built", "lane": "hw", "description": "High-bandwidth memory that feeds chips fast enough.", "evidence": "TrendForce reports AI could take nearly 20% of global DRAM supply in 2026; high-bandwidth memory uses about four times the wafer capacity of standard DRAM per gigabyte.", "sources": [ "trendforce" ], "needs": [], "derived": { "wave": 1, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "firmpower", "kind": "capability", "label": "Firm power deals", "status": "emerging", "lane": "energy", "description": "Long-term contracts for always-on generation, from nuclear restarts to gas plants built on site, not only grid power purchase agreements.", "evidence": "Microsoft agreed in September 2024 to restart Three Mile Island Unit 1, 835 MW, to power its datacentres from around 2028.", "sources": [ "tmi" ], "needs": [], "derived": { "wave": 1, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "talent", "kind": "capability", "label": "Frontier research talent", "status": "emerging", "lane": "capital", "description": "The small number of people who can push the frontier.", "evidence": "Meta offered packages worth up to $300 million over four years to staff its Superintelligence Labs in 2025, and OpenAI answered with retention bonuses.", "sources": [ "talentwar" ], "needs": [], "derived": { "wave": 8, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "codegen", "kind": "stage", "label": "AI writes most new code", "status": "built", "lane": "software", "description": "Models write the majority of new code inside large engineering organisations.", "evidence": "Google says 75% of its new code was AI-generated by April 2026, up from about 25% in 2024 and about 50% in autumn 2025.", "sources": [ "semafor" ], "needs": [ { "id": "rlscale", "link": "hard" }, { "id": "tools", "link": "soft" } ], "watch": "Share of new code written by AI. Google: 75% in April 2026.", "when": { "type": "here", "from": 2025.7, "to": null, "tail": null, "note": "Google: about half of new code by autumn 2025, 75% by April 2026.", "sources": [ "semafor" ] }, "derived": { "wave": 5, "cluster": "Agents at work", "leverage": 0 } }, { "id": "codeagents", "kind": "stage", "label": "Coding agents as products", "status": "built", "lane": "software", "description": "Agents that take a ticket and return a working change, sold as products.", "evidence": "Cursor reportedly passed $2bn in annualised revenue in March 2026, having doubled in three months.", "sources": [ "cursor" ], "needs": [ { "id": "longh", "link": "hard" }, { "id": "codegen", "link": "hard" } ], "watch": "Coding-tool revenue. Cursor: over $2bn run rate in March 2026.", "when": { "type": "here", "from": 2025.4, "to": null, "tail": null, "note": "Coding agents sold as products from 2025; Cursor passed $2bn run rate in March 2026.", "sources": [ "cursor" ] }, "derived": { "wave": 7, "cluster": "Agents at work", "leverage": 0 } }, { "id": "codeevals", "kind": "capability", "label": "Contamination-resistant evals", "status": "emerging", "lane": "measure", "description": "Benchmarks in coding, maths, science and agent work that models cannot have seen or looked up.", "evidence": "SWE-bench Verified is saturated at around 95% and was deprecated by OpenAI in February 2026; on the harder SWE-bench Pro the same leading model drops to 80%. Epoch reports models searching online and checking future commits to game benchmarks.", "sources": [ "swepro", "rlenvs" ], "needs": [], "derived": { "wave": 6, "cluster": "Agents at work", "leverage": 0 } }, { "id": "codereview", "kind": "capability", "label": "Code review at scale", "status": "emerging", "lane": "software", "description": "Checking AI-written code as fast as it is produced.", "evidence": "Writing code is no longer the slow step; reviewing it is. METR found experienced developers slower with AI tools even while they felt faster.", "sources": [ "metrdev" ], "needs": [ { "id": "verifiers", "link": "hard" }, { "id": "codeevals", "link": "soft" } ], "derived": { "wave": 7, "cluster": "Agents at work", "leverage": 2.51 } }, { "id": "selfsoftware", "kind": "stage", "label": "Software that maintains itself", "status": "speculative", "lane": "software", "description": "Codebases that agents keep patched, upgraded and secure with little human work.", "evidence": "", "sources": [], "needs": [ { "id": "codeagents", "link": "hard" }, { "id": "codereview", "link": "hard" }, { "id": "cyberdef", "link": "soft" }, { "id": "learn", "link": "soft" } ], "when": { "type": "est", "from": 2028, "to": 2035, "tail": null, "note": "My estimate, following coding agents and AI cyber defence.", "sources": [] }, "derived": { "wave": 8, "cluster": "Agents at work", "leverage": 0 } }, { "id": "agi", "kind": "stage", "label": "AGI: top-expert-level AI", "status": "speculative", "lane": "models", "description": "AI that matches top human experts across most cognitive work. There is no shared definition: AI Futures calls it top-expert-dominating AI, Amodei a country of geniuses in a datacentre, and Hendrycks et al. set a lower bar of a well-educated adult.", "evidence": "AI Futures’ forecasters put top-expert-dominating AI at medians from November 2028 to July 2032. Metaculus forecasters put AGI at 25% by 2029 and 50% by 2033. On the lower Hendrycks bar GPT-5 scores 57%, with long-term memory the weakest domain.", "sources": [ "aifut", "machines", "eightyk", "agidef" ], "needs": [ { "id": "reason", "link": "hard" }, { "id": "multimodal", "link": "hard" }, { "id": "contlearn", "link": "hard" }, { "id": "generalise", "link": "hard" }, { "id": "reliab", "link": "hard" }, { "id": "compuse", "link": "soft" }, { "id": "longh", "link": "hard" } ], "watch": "Hendrycks AGI score (GPT-5: 57%), especially the long-term memory domain.", "when": { "type": "fc", "from": 2029, "to": 2033, "tail": 2047, "note": "Metaculus forecasters: 25% by 2029 and 50% by 2033. The 2023 survey of AI researchers put the median at 2047.", "sources": [ "eightyk" ] }, "derived": { "wave": 7, "cluster": "Agents at work", "leverage": 0.13 } }, { "id": "asi", "kind": "stage", "label": "Superintelligence", "status": "speculative", "lane": "models", "description": "Systems far beyond the best people across most cognitive work.", "evidence": "Forethought argues that once AI automates AI research, software progress alone could accelerate if returns to research exceed 1, with compute for experiments as the main brake.", "sources": [ "forethought", "ai2027" ], "needs": [ { "id": "agi", "link": "hard" }, { "id": "air", "link": "hard" }, { "id": "iexplode", "link": "soft" }, { "id": "design", "link": "soft" }, { "id": "alignment", "link": "soft" }, { "id": "compute", "link": "hard" } ], "watch": "Returns to software research (Forethought estimate: 0.5 to 2). Above 1 means acceleration.", "when": { "type": "est", "from": 2031, "to": 2045, "tail": null, "note": "My estimate: some years after AGI, sooner if research automation compounds as Forethought describes.", "sources": [ "forethought" ] }, "derived": { "wave": 10, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "compuse", "kind": "capability", "label": "Computer use", "status": "emerging", "lane": "agents", "description": "Agents operate ordinary software through its screen.", "evidence": "Karpathy lists reliable computer use as one of the pieces agents still lack.", "sources": [ "karpathy" ], "needs": [ { "id": "multimodal", "link": "hard" } ], "derived": { "wave": 8, "cluster": "Agents at work", "leverage": 0.2 } }, { "id": "context", "kind": "capability", "label": "Context engineering", "status": "built", "lane": "agents", "description": "Compaction, retrieval and pruning keep a long run coherent.", "evidence": "Quality drops as the context window fills (\"context rot\"), so what goes into it has become an engineering discipline of its own.", "sources": [ "ctxrot" ], "needs": [], "derived": { "wave": 1, "cluster": "Agents at work", "leverage": 0 } }, { "id": "longh", "kind": "stage", "label": "Long-horizon agents", "status": "built", "lane": "agents", "description": "Agents that work for hours on a single task.", "evidence": "METR measures the length of task an agent can finish half the time. It doubled about every 7 months from 2019 to 2024 and about every 4 months since 2023. Frontier models measured roughly 6 to 12 hours in the first half of 2026.", "sources": [ "metr", "halflife" ], "needs": [ { "id": "reason", "link": "hard" }, { "id": "tools", "link": "hard" }, { "id": "compuse", "link": "soft" }, { "id": "context", "link": "hard" } ], "watch": "METR 50% time horizon: roughly 6 to 12 hours in early 2026, doubling every 4 to 7 months.", "when": { "type": "here", "from": 2025.3, "to": null, "tail": null, "note": "METR measured frontier agents finishing multi-hour tasks during 2025.", "sources": [ "metr" ] }, "derived": { "wave": 6, "cluster": "Agents at work", "leverage": 0 } }, { "id": "memory", "kind": "capability", "label": "Persistent memory", "status": "emerging", "lane": "agents", "description": "What an agent carries between sessions, and what it safely forgets.", "evidence": "Today’s memory is files, retrieval and summaries kept outside a frozen model.", "sources": [ "overhang" ], "needs": [ { "id": "context", "link": "hard" } ], "derived": { "wave": 6, "cluster": "Agents at work", "leverage": 2.12 } }, { "id": "ontology", "kind": "capability", "label": "Shared ontology and wiki", "status": "emerging", "lane": "agents", "description": "One structured source of truth that people and agents both read and write.", "evidence": "The NANDA report finds generic tools fail inside companies partly because they cannot absorb the company’s specific processes.", "sources": [ "nanda" ], "needs": [], "derived": { "wave": 6, "cluster": "Agents at work", "leverage": 0.4 } }, { "id": "skills", "kind": "capability", "label": "Skills", "status": "emerging", "lane": "agents", "description": "Reusable procedures written down as units an agent can load.", "evidence": "Agent Skills and AGENTS.md or CLAUDE.md files are the current way procedures are handed to agents.", "sources": [ "overhang" ], "needs": [ { "id": "tools", "link": "hard" } ], "derived": { "wave": 6, "cluster": "Agents at work", "leverage": 0 } }, { "id": "connectors", "kind": "capability", "label": "Company data connectors", "status": "built", "lane": "agents", "description": "Agents can reach the systems where company knowledge lives.", "evidence": "MCP connects agents to tools and data, and A2A connects agents to each other. Both sit under the Linux Foundation.", "sources": [ "a2a" ], "needs": [ { "id": "tools", "link": "hard" } ], "derived": { "wave": 2, "cluster": "Agents at work", "leverage": 0 } }, { "id": "know", "kind": "stage", "label": "Knowledge-rich agents", "status": "emerging", "lane": "agents", "description": "Agents that carry company and personal context across sessions and people.", "evidence": "The MIT NANDA report blames most stalled enterprise pilots on a \"learning gap\": tools that cannot keep feedback or absorb a company’s own processes.", "sources": [ "nanda", "overhang" ], "needs": [ { "id": "longh", "link": "hard" }, { "id": "memory", "link": "hard" }, { "id": "ontology", "link": "soft" }, { "id": "skills", "link": "soft" }, { "id": "connectors", "link": "hard" }, { "id": "alwayson", "link": "soft" }, { "id": "firmmodels", "link": "soft" } ], "watch": "Repeat-error rate inside one company. Share of answers grounded in company data.", "when": { "type": "est", "from": 2026, "to": 2029, "tail": null, "note": "My estimate. Memory and connectors are shipping now, but reliable company-wide use is not yet common.", "sources": [ "nanda" ] }, "derived": { "wave": 7, "cluster": "Agents at work", "leverage": 0.4 } }, { "id": "traces", "kind": "capability", "label": "Trace capture", "status": "emerging", "lane": "data", "description": "A record of what agents did and why, kept as evidence and as training signal.", "evidence": "Silver and Sutton’s agents learn from their own streams of action and observation. Chan et al. need the same records for attribution.", "sources": [ "experience", "infra" ], "needs": [ { "id": "longh", "link": "hard" }, { "id": "ambientdocs", "link": "soft" } ], "derived": { "wave": 7, "cluster": "Agents at work", "leverage": 1.08 } }, { "id": "correction", "kind": "capability", "label": "Correction propagation", "status": "emerging", "lane": "agents", "description": "A fix made once reaches every agent and stays fixed.", "evidence": "The NANDA \"learning gap\": tools that do not retain feedback keep repeating the same mistakes.", "sources": [ "nanda" ], "needs": [ { "id": "memory", "link": "hard" }, { "id": "ontology", "link": "hard" } ], "derived": { "wave": 7, "cluster": "Agents at work", "leverage": 0.4 } }, { "id": "learn", "kind": "stage", "label": "Continual-learning agents", "status": "emerging", "lane": "agents", "description": "Agents that get better at your work by doing it, without waiting for a lab to retrain them.", "evidence": "Dwarkesh Patel and Karpathy both name continual learning as the main missing piece. Silver and Sutton argue the next era learns from experience rather than human data.", "sources": [ "dwarkeshcl", "karpathy", "experience" ], "needs": [ { "id": "know", "link": "hard" }, { "id": "traces", "link": "hard" }, { "id": "verifiers", "link": "hard" }, { "id": "contlearn", "link": "soft" }, { "id": "correction", "link": "hard" }, { "id": "generalise", "link": "soft" }, { "id": "firmmodels", "link": "soft" } ], "watch": "Whether any lab ships weight-level learning from deployment.", "when": { "type": "fc", "from": 2026, "to": 2035, "tail": null, "note": "Karpathy said in October 2025 that agents need about a decade, with continual learning a main missing piece.", "sources": [ "karpathy" ] }, "derived": { "wave": 8, "cluster": "Agents at work", "leverage": 0.2 } }, { "id": "a2a", "kind": "capability", "label": "Agent-to-agent protocols", "status": "emerging", "lane": "agents", "description": "Work passes between agents and people without losing context.", "evidence": "A2A reached version 1.0 with more than 150 supporting organisations by April 2026. Multi-agent systems still fail often when agents misread each other.", "sources": [ "a2a", "mast" ], "needs": [ { "id": "identity", "link": "hard" } ], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0.22 } }, { "id": "adopt", "kind": "capability", "label": "Workflow redesign", "status": "emerging", "lane": "agents", "description": "Organisations change how work flows so agents fit into it.", "evidence": "METR found experienced developers 19% slower with AI tools while believing they were faster. Most enterprise pilots stall on integration and workflow, not on the model.", "sources": [ "metrdev", "nanda", "aei" ], "needs": [ { "id": "proagents", "link": "soft" }, { "id": "ambientdocs", "link": "soft" } ], "derived": { "wave": 1, "cluster": "Assistants and the public", "leverage": 2.27 } }, { "id": "evals", "kind": "capability", "label": "Real-world task evals", "status": "emerging", "lane": "measure", "description": "Measuring models on real deliverables, so delegation rests on evidence.", "evidence": "OpenAI’s GDPval grades models on 1,230 real work products from 44 occupations; the best model matched or beat human experts 47.6% of the time. METR’s time horizons track how long a task an agent can finish.", "sources": [ "gdpval", "metr" ], "needs": [], "derived": { "wave": 1, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "org", "kind": "stage", "label": "Agent-native organisations", "status": "emerging", "lane": "agents", "description": "Roles, authority limits and accountability built around agents.", "evidence": "OpenAI’s reported top level, \"Organizations\": AI that can do all the work of an organisation. A study of multi-agent systems finds most failures come from system design and agents misreading each other, and argues better base models alone will not fix them.", "sources": [ "openai5", "mast" ], "needs": [ { "id": "learn", "link": "soft" }, { "id": "identity", "link": "hard" }, { "id": "audit", "link": "hard" }, { "id": "rollback", "link": "soft" }, { "id": "a2a", "link": "soft" }, { "id": "security", "link": "hard" }, { "id": "reliab", "link": "hard" }, { "id": "adopt", "link": "hard" }, { "id": "evals", "link": "soft" }, { "id": "liability", "link": "soft" }, { "id": "regulation", "link": "soft" }, { "id": "coord", "link": "soft" } ], "watch": "Share of tasks run without human review. pass^k on real workflows.", "when": { "type": "est", "from": 2028, "to": 2035, "tail": null, "note": "My estimate. Depends on identity, audit, liability and workflow change more than model capability.", "sources": [] }, "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0.1 } }, { "id": "payments", "kind": "capability", "label": "Agent payments", "status": "emerging", "lane": "agents", "description": "Agents spend and earn inside limits someone set.", "evidence": "Google’s AP2 extends A2A into payments with more than 60 financial firms. Coinbase’s x402, now a Linux Foundation project, settles stablecoin payments per web request.", "sources": [ "ap2", "paycmp", "a2a" ], "needs": [ { "id": "identity", "link": "hard" }, { "id": "a2a", "link": "soft" }, { "id": "stablerails", "link": "soft" } ], "derived": { "wave": 7, "cluster": "Trust plumbing and the agent economy", "leverage": 0.3 } }, { "id": "permeab", "kind": "capability", "label": "Permeability controls", "status": "speculative", "lane": "agents", "description": "Deciding how much an agent market touches the human economy.", "evidence": "DeepMind calls the boundary between agent and human economies \"the critical and controllable design variable\".", "sources": [ "vae" ], "needs": [ { "id": "payments", "link": "hard" } ], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "econ", "kind": "stage", "label": "Agent economies", "status": "speculative", "lane": "agents", "description": "Agents hold budgets and reputations, trade with each other and form task-specific coalitions.", "evidence": "Google DeepMind argues agent markets are emerging by default, and that how porous they are to the human economy is the main thing we can still design.", "sources": [ "vae" ], "needs": [ { "id": "org", "link": "hard" }, { "id": "payments", "link": "hard" }, { "id": "xtrust", "link": "hard" }, { "id": "personhood", "link": "soft" }, { "id": "permeab", "link": "soft" }, { "id": "liability", "link": "soft" }, { "id": "regulation", "link": "soft" }, { "id": "deflation", "link": "soft" }, { "id": "coord", "link": "hard" }, { "id": "agentinst", "link": "soft" } ], "watch": "Payment volume through agent payment protocols such as AP2 and x402.", "when": { "type": "est", "from": 2030, "to": 2040, "tail": null, "note": "My estimate. Payment rails exist; trust, liability and identity across companies do not yet.", "sources": [] }, "derived": { "wave": 10, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "design", "kind": "stage", "label": "Agents designing agents", "status": "speculative", "lane": "agents", "description": "Agents build new agents, organisations and evaluations.", "evidence": "The later stages of AI 2027, where AI research is mostly automated and progress compounds.", "sources": [ "ai2027" ], "needs": [ { "id": "air", "link": "hard" }, { "id": "org", "link": "soft" }, { "id": "oversight", "link": "soft" }, { "id": "alignment", "link": "soft" } ], "watch": "Agent-designed architectures or training methods adopted by labs.", "when": { "type": "est", "from": 2029, "to": 2038, "tail": null, "note": "My estimate, following automated AI research.", "sources": [] }, "derived": { "wave": 10, "cluster": "Oversight and alignment", "leverage": 0.07 } }, { "id": "reconcile", "kind": "capability", "label": "Divergent-copy reconciliation", "status": "speculative", "lane": "agents", "description": "Copies of a knowledge base that drift apart can merge without losing disagreements.", "evidence": "", "sources": [], "needs": [ { "id": "ontology", "link": "hard" }, { "id": "audit", "link": "hard" } ], "derived": { "wave": 10, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "protein", "kind": "capability", "label": "Protein structure and design", "status": "built", "lane": "science", "description": "Predicting and designing protein structures with AI.", "evidence": "Baker, Hassabis and Jumper shared the 2024 Nobel Prize in Chemistry for protein design and structure prediction.", "sources": [ "nobel" ], "needs": [ { "id": "compute", "link": "hard" } ], "derived": { "wave": 3, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "maths", "kind": "capability", "label": "AI in research mathematics", "status": "emerging", "lane": "science", "description": "Models that find proofs and better constructions, with checkable answers.", "evidence": "A Gemini Deep Think model reached IMO gold in 2025. AlphaEvolve improved the best known solution on about 20% of more than 50 open problems.", "sources": [ "aimath" ], "needs": [ { "id": "reason", "link": "hard" }, { "id": "verifiers", "link": "hard" } ], "derived": { "wave": 9, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "labauto", "kind": "capability", "label": "Lab automation", "status": "emerging", "lane": "science", "description": "Agents run experiments, not only propose them.", "evidence": "FutureHouse’s Robin proposes experiments and interprets the results in a closed loop, while people still run the wet-lab work. Fully automated labs are the next step.", "sources": [ "labloop" ], "needs": [ { "id": "robotdata", "link": "hard" }, { "id": "worldmodels", "link": "soft" } ], "derived": { "wave": 3, "cluster": "Science and medicine", "leverage": 0.42 } }, { "id": "biodata", "kind": "capability", "label": "Perturbation data at scale", "status": "emerging", "lane": "science", "description": "Large datasets of how cells respond when genes or drugs change them.", "evidence": "Tahoe Therapeutics, Arc Institute and Biohub plan more than 120 million single-cell data points across 225,000 drug–patient interactions, to be open-sourced, in what they call the largest perturbation dataset for virtual cell models.", "sources": [ "tahoe" ], "needs": [ { "id": "labauto", "link": "hard" } ], "derived": { "wave": 5, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "virtualcell", "kind": "capability", "label": "Virtual cell models", "status": "speculative", "lane": "science", "description": "Simulating how a cell responds before running the experiment.", "evidence": "Arc’s first Virtual Cell Challenge found no model beat the baseline on mean absolute error. \"We don’t think we have a virtual cell yet.\"", "sources": [ "vcc" ], "needs": [ { "id": "biodata", "link": "hard" }, { "id": "protein", "link": "hard" } ], "derived": { "wave": 6, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "materials", "kind": "capability", "label": "Materials discovery", "status": "emerging", "lane": "science", "description": "Predicting new materials, then making and testing them.", "evidence": "Google’s GNoME predicted 2.2 million new crystals, and chemists disputed how many are new or useful. Synthesis and testing are the bottleneck.", "sources": [ "gnome" ], "needs": [ { "id": "labauto", "link": "hard" } ], "derived": { "wave": 4, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "trials", "kind": "capability", "label": "Clinical validation", "status": "emerging", "lane": "science", "description": "Proving in people that a drug works and is safe.", "evidence": "Biology sets the clock. Insilico’s rentosertib is in phase III, with initial results expected around the end of 2029.", "sources": [ "insilico" ], "needs": [], "derived": { "wave": 3, "cluster": "Science and medicine", "leverage": 0.42 } }, { "id": "drugs", "kind": "stage", "label": "AI-designed medicines", "status": "emerging", "lane": "science", "description": "Drugs whose targets and molecules came from AI, proven in the clinic.", "evidence": "Insilico’s rentosertib for lung fibrosis showed results in phase IIa and moved into phase III. Isomorphic Labs pushed its first trials from 2025 to the end of 2026.", "sources": [ "insilico", "isomorphic" ], "needs": [ { "id": "protein", "link": "hard" }, { "id": "labauto", "link": "hard" }, { "id": "trials", "link": "hard" } ], "watch": "AI-designed drugs passing phase III. The first readout is expected around 2029.", "when": { "type": "fc", "from": 2029, "to": 2031, "tail": null, "note": "Insilico’s phase III: initial results expected by December 2029, a possible approval filing in 2030.", "sources": [ "insilico" ] }, "derived": { "wave": 4, "cluster": "Science and medicine", "leverage": 0.17 } }, { "id": "disc", "kind": "stage", "label": "Agent-run discovery", "status": "emerging", "lane": "science", "description": "Agents propose hypotheses, run experiments and publish results.", "evidence": "Level 4, \"Innovators\", on OpenAI’s reported scale. Early systems such as Sakana’s AI Scientist write machine-learning papers end to end for about $15 each. Verifier’s law suggests fields with fast, objective checks fall first, which is why maths is moving fastest.", "sources": [ "openai5", "aisci", "wei" ], "needs": [ { "id": "learn", "link": "soft" }, { "id": "air", "link": "hard" }, { "id": "generalise", "link": "soft" }, { "id": "verifscale", "link": "soft" }, { "id": "labauto", "link": "soft" }, { "id": "maths", "link": "soft" }, { "id": "biosec", "link": "soft" }, { "id": "validate", "link": "hard" } ], "watch": "Peer-reviewed results where AI did most of the work. Records broken on open maths problems.", "when": { "type": "here", "from": 2025.4, "to": null, "tail": null, "note": "First results: AlphaEvolve improved the best known answer on about 20% of more than 50 open problems.", "sources": [ "aimath" ] }, "derived": { "wave": 10, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "clinical", "kind": "capability", "label": "Clinical AI in practice", "status": "emerging", "lane": "science", "description": "AI in day-to-day care: documentation, triage and decision support.", "evidence": "In a study of about 1,800 clinicians, ambient AI scribes saved 16 minutes of documentation per eight hours of care and reduced burnout. Useful, but modest.", "sources": [ "stat" ], "needs": [ { "id": "reason", "link": "hard" }, { "id": "regulation", "link": "soft" }, { "id": "airgap", "link": "soft" } ], "derived": { "wave": 6, "cluster": "Science and medicine", "leverage": 0.17 } }, { "id": "medicine", "kind": "stage", "label": "AI-native medicine", "status": "speculative", "lane": "science", "description": "Diagnosis, treatment and drug choice shaped by models of the individual patient.", "evidence": "", "sources": [], "needs": [ { "id": "drugs", "link": "hard" }, { "id": "clinical", "link": "hard" }, { "id": "virtualcell", "link": "soft" }, { "id": "liability", "link": "soft" } ], "watch": "Regulatory approvals for AI-guided diagnosis and treatment.", "when": { "type": "est", "from": 2032, "to": 2042, "tail": null, "note": "My estimate, after the first AI-designed drugs are approved.", "sources": [] }, "derived": { "wave": 7, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "teleop", "kind": "capability", "label": "Teleoperation as a data pipeline", "status": "emerging", "lane": "physical", "description": "Remote humans operate robots, producing the data that trains autonomy.", "evidence": "The Robot Report argues teleoperation is \"both a business model and an autonomy pipeline\" for 1X’s NEO.", "sources": [ "teleop" ], "needs": [ { "id": "actuators", "link": "soft" } ], "derived": { "wave": 1, "cluster": "Robots and the physical economy", "leverage": 1.73 } }, { "id": "robotdata", "kind": "capability", "label": "Robot data and foundation models", "status": "emerging", "lane": "physical", "description": "Enough physical experience to train general robot policies.", "evidence": "π0.5 tidies homes it never saw in training, though not always at the first try. Video is plentiful, but robots lack data linking joint movements to outcomes: \"we don’t have data for that\", says Vincent Sitzmann. A 2025 CMU thesis explores three ways to scale it: simulation, cheap teleoperation and human data gathered in the wild.", "sources": [ "pi05", "robotdata" ], "needs": [ { "id": "teleop", "link": "hard" }, { "id": "actuators", "link": "soft" } ], "derived": { "wave": 2, "cluster": "Robots and the physical economy", "leverage": 1.73 } }, { "id": "worldmodels", "kind": "capability", "label": "World models and simulation", "status": "emerging", "lane": "physical", "description": "Practice in a generated world that transfers to the real one.", "evidence": "DeepMind’s Genie 3 generates interactive worlds that stay consistent for minutes and is already used to train its SIMA agent. DeepMind calls world models \"a key stepping stone on the path to AGI\". Simulation alone remains \"essential but insufficient\" for robots, so real action data is still needed.", "sources": [ "genie", "worldm" ], "needs": [ { "id": "multimodal", "link": "hard" } ], "derived": { "wave": 5, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "av", "kind": "stage", "label": "Robotaxis at scale", "status": "built", "lane": "physical", "description": "Driverless vehicles carrying paying passengers in many cities.", "evidence": "Waymo carries 500,000 paid rides a week across 10 US markets, double its volume less than a year earlier.", "sources": [ "waymo" ], "needs": [ { "id": "multimodal", "link": "hard" }, { "id": "reliab", "link": "soft" }, { "id": "worldmodels", "link": "soft" }, { "id": "regulation", "link": "soft" }, { "id": "liability", "link": "soft" }, { "id": "remagnets", "link": "soft" } ], "watch": "Robotaxi paid rides per week (Waymo: 500,000 in 2026) and cities served.", "when": { "type": "here", "from": 2025.3, "to": null, "tail": null, "note": "Waymo passed 250,000 paid rides a week in April 2025 and 500,000 in 2026.", "sources": [ "waymo" ] }, "derived": { "wave": 5, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "embod", "kind": "stage", "label": "Embodied agents", "status": "emerging", "lane": "physical", "description": "General agents that act in the physical world, where every mistake has a physical cost.", "evidence": "π0.5 tidies homes it never saw in training. Simulation is \"essential but insufficient\", so real action data remains the scarce input.", "sources": [ "pi05", "worldm" ], "needs": [ { "id": "longh", "link": "soft" }, { "id": "multimodal", "link": "hard" }, { "id": "robotdata", "link": "hard" }, { "id": "worldmodels", "link": "soft" }, { "id": "rollback", "link": "soft" }, { "id": "remagnets", "link": "soft" }, { "id": "actuators", "link": "soft" } ], "watch": "Tasks completed in unseen homes without a remote operator.", "when": { "type": "here", "from": 2025.3, "to": null, "tail": null, "note": "First general robot policies working in unseen homes, 2025.", "sources": [ "pi05" ] }, "derived": { "wave": 5, "cluster": "Robots and the physical economy", "leverage": 0.82 } }, { "id": "homebots", "kind": "stage", "label": "Home robots", "status": "speculative", "lane": "physical", "description": "General-purpose robots doing household work.", "evidence": "1X plans to deliver its NEO humanoid to US homes in 2026, leaning on remote operators at first.", "sources": [ "teleop" ], "needs": [ { "id": "embod", "link": "hard" }, { "id": "teleop", "link": "soft" }, { "id": "liability", "link": "soft" }, { "id": "remagnets", "link": "soft" }, { "id": "actuators", "link": "soft" } ], "watch": "Share of home-robot work done without teleoperation.", "when": { "type": "est", "from": 2026, "to": 2033, "tail": null, "note": "1X plans NEO deliveries in 2026, remotely operated at first. Real autonomy later is my estimate.", "sources": [ "teleop" ] }, "derived": { "wave": 6, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "latency", "kind": "stage", "label": "Long-latency civilisation", "status": "speculative", "lane": "physical", "description": "Agents act for years with no real-time sync or oversight.", "evidence": "", "sources": [], "needs": [ { "id": "embod", "link": "hard" }, { "id": "design", "link": "hard" }, { "id": "reconcile", "link": "soft" }, { "id": "valuestab", "link": "hard" } ], "when": { "type": "far", "from": 2045, "to": null, "tail": null, "note": "No forecast found.", "sources": [] }, "derived": { "wave": 11, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "chat", "kind": "stage", "label": "Conversational assistants", "status": "built", "lane": "society", "description": "General-purpose assistants, the first mass-market form of generative AI.", "evidence": "The Anthropic Economic Index tracks what people use assistants for, and how use shifts from getting help toward delegating whole tasks.", "sources": [ "aei" ], "needs": [ { "id": "pretrain", "link": "hard" } ], "watch": "Weekly users. Share of use that is delegation rather than advice.", "when": { "type": "here", "from": 2022.9, "to": null, "tail": null, "note": "ChatGPT launched in November 2022.", "sources": [] }, "derived": { "wave": 4, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "ondevice", "kind": "capability", "label": "On-device models", "status": "built", "lane": "society", "description": "Models that run on a phone or laptop, so private data stays local.", "evidence": "Apple’s Foundation Models framework gives app developers direct access to an on-device model.", "sources": [ "applefm" ], "needs": [ { "id": "openw", "link": "hard" }, { "id": "cost", "link": "hard" } ], "derived": { "wave": 5, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "search", "kind": "stage", "label": "AI answers replace search", "status": "emerging", "lane": "society", "description": "People get answers from models instead of clicking through to sources.", "evidence": "Google referrals to eight large UK publishers fell 7.1% in one quarter of 2025 as AI Overviews expanded; on that trend they halve by Q3 2027.", "sources": [ "pressgazette" ], "needs": [ { "id": "chat", "link": "hard" }, { "id": "connectors", "link": "soft" }, { "id": "emmarkets", "link": "soft" } ], "watch": "Search referrals to publishers. UK publishers are on course to see them halve by Q3 2027.", "when": { "type": "here", "from": 2024.4, "to": null, "tail": null, "note": "Google launched AI Overviews in May 2024.", "sources": [ "pressgazette" ] }, "derived": { "wave": 5, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "creator", "kind": "capability", "label": "Paying the people who make knowledge", "status": "speculative", "lane": "society", "description": "A business model for the writing and reporting that models learn from.", "evidence": "As AI answers absorb clicks, the open web’s business model weakens. Licensing deals and settlements such as Anthropic’s $1.5bn are early attempts at a new one.", "sources": [ "pressgazette", "anthsettle" ], "needs": [ { "id": "datarights", "link": "hard" } ], "derived": { "wave": 8, "cluster": "Around licensed generative media", "leverage": 0.34 } }, { "id": "infoeco", "kind": "capability", "label": "Healthy information ecosystem", "status": "speculative", "lane": "society", "description": "A web where people can find, trust and pay for human knowledge.", "evidence": "Pew estimates about 10% of web pages, and more than a third of those published since ChatGPT launched, show signs of AI authorship.", "sources": [ "pew" ], "needs": [ { "id": "creator", "link": "hard" }, { "id": "provenance", "link": "hard" }, { "id": "labelling", "link": "soft" } ], "derived": { "wave": 9, "cluster": "Around licensed generative media", "leverage": 0.34 } }, { "id": "companions", "kind": "stage", "label": "AI companions", "status": "emerging", "lane": "society", "description": "Ongoing, personal relationships with AI characters and assistants.", "evidence": "Character.AI barred under-18s from open-ended chat in November 2025 and, with Google, agreed in January 2026 to settle lawsuits over harm to teenagers.", "sources": [ "charai" ], "needs": [ { "id": "chat", "link": "hard" }, { "id": "memory", "link": "soft" }, { "id": "wellbeing", "link": "soft" }, { "id": "wearables", "link": "soft" } ], "watch": "Age-assurance laws and companion-app lawsuits.", "when": { "type": "here", "from": 2023, "to": null, "tail": null, "note": "Companion apps reached mass use by 2023.", "sources": [ "charai" ] }, "derived": { "wave": 5, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "education", "kind": "stage", "label": "AI tutoring at scale", "status": "emerging", "lane": "society", "description": "Personal tutors for every learner.", "evidence": "A World Bank pilot in Nigeria found six weeks of after-school AI tutoring raised scores by about 0.3 standard deviations, roughly 1.5 to 2 years of typical progress, and Harvard physics students learned more with an AI tutor. Other trials show small or no effects, and low usage kills results.", "sources": [ "tutor", "worldbank" ], "needs": [ { "id": "chat", "link": "hard" }, { "id": "evals", "link": "soft" }, { "id": "memory", "link": "soft" }, { "id": "emmarkets", "link": "soft" } ], "watch": "Effect sizes in large randomised trials, and how much students actually use the tutor.", "when": { "type": "est", "from": 2026, "to": 2032, "tail": null, "note": "My estimate. Trials show large but uneven gains, and scale depends on whether students keep using the tutor.", "sources": [ "tutor" ] }, "derived": { "wave": 5, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "labour", "kind": "stage", "label": "White-collar hiring slowdown", "status": "emerging", "lane": "society", "description": "Fewer junior hires in the jobs AI does best. Whether AI is the cause is contested.", "evidence": "Stanford’s Digital Economy Lab finds employment for 22 to 25 year olds in the most AI-exposed jobs about 19% below trend by June 2026, mostly through less hiring. But Fairlie and Wu find no significant AI effect on 2026 graduate unemployment, and the New York Fed attributes 64% of the earlier rise to remote work.", "sources": [ "canaries", "xeno" ], "needs": [ { "id": "adopt", "link": "hard" }, { "id": "longh", "link": "soft" }, { "id": "deflation", "link": "soft" }, { "id": "smallfirms", "link": "soft" } ], "watch": "Employment of 22 to 25 year olds in AI-exposed jobs: 19% below trend in June 2026, with the cause disputed.", "when": { "type": "here", "from": 2023, "to": null, "tail": null, "note": "Payroll data show the gap for young workers opening from late 2022.", "sources": [ "canaries" ] }, "derived": { "wave": 2, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "pipeline", "kind": "capability", "label": "Expertise pipeline", "status": "speculative", "lane": "society", "description": "People still become experts, so someone can check AI work later.", "evidence": "If entry-level hiring keeps falling in AI-exposed jobs, fewer people build the expertise that oversight depends on. Nobody measures this directly yet.", "sources": [ "canaries" ], "needs": [ { "id": "education", "link": "hard" } ], "derived": { "wave": 8, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "intent", "kind": "capability", "label": "Intent interfaces", "status": "speculative", "lane": "society", "description": "A person states what they care about precisely enough for agents to act on it, and notices when it is misread.", "evidence": "", "sources": [], "needs": [ { "id": "memory", "link": "hard" }, { "id": "a2a", "link": "hard" } ], "derived": { "wave": 10, "cluster": "Trust plumbing and the agent economy", "leverage": 0.22 } }, { "id": "civ", "kind": "stage", "label": "Civilisation-scale knowledge", "status": "speculative", "lane": "society", "description": "A shared, self-correcting record of what a society knows, with provenance on every claim.", "evidence": "", "sources": [], "needs": [ { "id": "econ", "link": "soft" }, { "id": "disc", "link": "soft" }, { "id": "governance", "link": "hard" }, { "id": "audit", "link": "hard" }, { "id": "infoeco", "link": "hard" } ], "when": { "type": "far", "from": 2045, "to": null, "tail": null, "note": "No forecast found.", "sources": [] }, "derived": { "wave": 10, "cluster": "Trust plumbing and the agent economy", "leverage": 0.22 } }, { "id": "symb", "kind": "stage", "label": "Human-agent symbiosis", "status": "speculative", "lane": "society", "description": "People steer a vast search process. Judgement, taste and values are the scarce inputs.", "evidence": "Gradual Disempowerment warns the default runs the other way: as AI replaces human labour, economies, cultures and states lose their reasons to stay aligned with people.", "sources": [ "gd" ], "needs": [ { "id": "civ", "link": "hard" }, { "id": "design", "link": "soft" }, { "id": "intent", "link": "hard" }, { "id": "humaninf", "link": "hard" }, { "id": "modelwelfare", "link": "soft" } ], "watch": "Labour share of income. How far institutions still depend on public consent.", "when": { "type": "far", "from": 2045, "to": null, "tail": null, "note": "No forecast found.", "sources": [] }, "derived": { "wave": 11, "cluster": "Trust plumbing and the agent economy", "leverage": 0.14 } }, { "id": "mind", "kind": "stage", "label": "Extended mind", "status": "speculative", "lane": "society", "description": "A personal agent holds your working life and becomes part of how you think.", "evidence": "", "sources": [], "needs": [ { "id": "symb", "link": "hard" }, { "id": "portability", "link": "hard" }, { "id": "ondevice", "link": "soft" }, { "id": "alwayson", "link": "soft" }, { "id": "wearables", "link": "soft" } ], "when": { "type": "far", "from": 2045, "to": null, "tail": null, "note": "No forecast found.", "sources": [] }, "derived": { "wave": 12, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "abund", "kind": "stage", "label": "Cognitive abundance", "status": "speculative", "lane": "society", "description": "Thinking is cheap, so value moves to what cannot be copied: energy, matter, trust and attention.", "evidence": "Epoch AI finds the price of a fixed level of capability falling between 9x and 900x a year.", "sources": [ "epochprice", "epoch2030" ], "needs": [ { "id": "econ", "link": "soft" }, { "id": "embod", "link": "soft" }, { "id": "compute", "link": "hard" }, { "id": "deflation", "link": "hard" } ], "watch": "Price per unit of fixed capability.", "when": { "type": "far", "from": 2045, "to": null, "tail": null, "note": "No forecast found.", "sources": [] }, "derived": { "wave": 3, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "govai", "kind": "stage", "label": "AI in public services", "status": "emerging", "lane": "society", "description": "Governments using generative AI to run services and make decisions.", "evidence": "The UK’s Humphrey suite gives civil servants tools for consultations, legislation and minutes. Early pilots report up to 25% efficiency gains in NHS appointment scheduling.", "sources": [ "humphrey" ], "needs": [ { "id": "chat", "link": "hard" }, { "id": "regulation", "link": "soft" }, { "id": "evals", "link": "soft" }, { "id": "airgap", "link": "soft" }, { "id": "ambientdocs", "link": "soft" } ], "watch": "Departments with AI in production rather than pilot.", "when": { "type": "here", "from": 2025.4, "to": null, "tail": null, "note": "The UK’s Humphrey suite reached civil servants in 2025.", "sources": [ "humphrey" ] }, "derived": { "wave": 5, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "imagegen", "kind": "capability", "label": "Image generation", "status": "built", "lane": "media", "description": "Photorealistic and stylised images from text.", "evidence": "Image models are mature enough that the fight has moved to rights: Disney and Warner Bros. are suing Midjourney in a consolidated case still active in October 2026.", "sources": [ "midj" ], "needs": [ { "id": "pretrain", "link": "hard" }, { "id": "multimodal", "link": "hard" } ], "derived": { "wave": 5, "cluster": "Around licensed generative media", "leverage": 0 } }, { "id": "videogen", "kind": "capability", "label": "Video generation with sound", "status": "built", "lane": "media", "description": "Short video clips with synchronised audio from a prompt.", "evidence": "Veo 3 added native audio in May 2025 and Sora 2 followed in September. By March 2026 Veo 3 Lite cost $0.05 a second at 720p.", "sources": [ "opusvideo" ], "needs": [ { "id": "imagegen", "link": "hard" }, { "id": "compute", "link": "hard" } ], "derived": { "wave": 6, "cluster": "Around licensed generative media", "leverage": 0 } }, { "id": "voice", "kind": "capability", "label": "Voice cloning and real-time speech", "status": "built", "lane": "media", "description": "Natural speech in any voice, live.", "evidence": "Recorded deepfake fraud losses rose from about $335m in 2024 to about $2.5bn in 2025, counting only publicly reported cases.", "sources": [ "surfshark" ], "needs": [ { "id": "multimodal", "link": "hard" } ], "derived": { "wave": 5, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "music", "kind": "capability", "label": "Music generation", "status": "built", "lane": "media", "description": "Full songs with vocals from a prompt.", "evidence": "After suing Suno and Udio, Warner and Universal settled in late 2025 and agreed partnerships on licensed music creation services.", "sources": [ "wmgsuno" ], "needs": [ { "id": "pretrain", "link": "hard" } ], "derived": { "wave": 4, "cluster": "Around licensed generative media", "leverage": 0 } }, { "id": "likeness", "kind": "capability", "label": "Likeness and performer rights", "status": "emerging", "lane": "media", "description": "Rules for using a person’s face, voice and performance in generated media.", "evidence": "SAG-AFTRA filed a charge over an AI-generated James Earl Jones voice for Darth Vader in Fortnite, and is bargaining over synthetic performers.", "sources": [ "opusvideo" ], "needs": [], "derived": { "wave": 6, "cluster": "Around licensed generative media", "leverage": 0 } }, { "id": "labelling", "kind": "capability", "label": "Disclosure and labelling", "status": "emerging", "lane": "media", "description": "Platforms and laws marking AI-generated content for audiences.", "evidence": "TikTok has labelled more than 1.3 billion videos as AI-generated. EU AI Act transparency duties for AI-generated content became enforceable in August 2026.", "sources": [ "opusvideo" ], "needs": [ { "id": "provenance", "link": "hard" }, { "id": "regulation", "link": "hard" } ], "derived": { "wave": 8, "cluster": "Around licensed generative media", "leverage": 0 } }, { "id": "licensedmedia", "kind": "stage", "label": "Licensed generative media", "status": "emerging", "lane": "media", "description": "Studios, labels and creators license catalogues and likenesses to model makers.", "evidence": "Disney agreed a three-year Sora licensing deal with a $1bn investment in OpenAI in December 2025, and the major labels moved from suing music generators to partnering with them.", "sources": [ "opusvideo", "wmgsuno" ], "needs": [ { "id": "videogen", "link": "hard" }, { "id": "music", "link": "hard" }, { "id": "datarights", "link": "hard" }, { "id": "likeness", "link": "soft" } ], "watch": "Licensing deals against new lawsuits between rights holders and model makers.", "when": { "type": "here", "from": 2025.8, "to": null, "tail": null, "note": "Labels settled with Suno and Udio in late 2025; Disney’s Sora deal followed in December.", "sources": [ "wmgsuno", "opusvideo" ] }, "derived": { "wave": 7, "cluster": "Around licensed generative media", "leverage": 0.5 } }, { "id": "deepfakes", "kind": "stage", "label": "Synthetic media at scale", "status": "built", "lane": "media", "description": "Cheap, convincing fake images, video and voices, used for fraud and influence.", "evidence": "Surfshark counts about $3.7bn in recorded deepfake fraud losses since 2020, 89% of it in 2025 and the first half of 2026.", "sources": [ "surfshark" ], "needs": [ { "id": "videogen", "link": "hard" }, { "id": "voice", "link": "hard" } ], "watch": "Recorded deepfake fraud losses: about $2.5bn in 2025.", "when": { "type": "here", "from": 2024, "to": null, "tail": null, "note": "Recorded deepfake fraud losses jumped from 2024.", "sources": [ "surfshark" ] }, "derived": { "wave": 7, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "aient", "kind": "stage", "label": "AI-native entertainment", "status": "speculative", "lane": "media", "description": "Interactive, personalised films and games generated on demand.", "evidence": "Genie 3 already generates interactive worlds that stay consistent for minutes, the building block for playable generated media.", "sources": [ "genie" ], "needs": [ { "id": "videogen", "link": "hard" }, { "id": "worldmodels", "link": "soft" }, { "id": "licensedmedia", "link": "hard" } ], "watch": "Time spent in generated, interactive media.", "when": { "type": "est", "from": 2028, "to": 2035, "tail": null, "note": "My estimate. Genie 3 already generates playable worlds for minutes at a time.", "sources": [ "genie" ] }, "derived": { "wave": 8, "cluster": "Around licensed generative media", "leverage": 0 } }, { "id": "exportctl", "kind": "capability", "label": "Export controls", "status": "built", "lane": "geo", "description": "Governments decide which chips and models may cross which borders.", "evidence": "In June 2026 the US Commerce Department ordered Anthropic to restrict two frontier models on national-security grounds; Anthropic pulled them for all users until the controls were lifted at the end of the month.", "sources": [ "fablesusp" ], "needs": [], "derived": { "wave": 1, "cluster": "Compute governance and geopolitics", "leverage": 0 } }, { "id": "race", "kind": "stage", "label": "Strategic AI competition", "status": "emerging", "lane": "geo", "description": "Frontier AI treated as a matter of national power.", "evidence": "Every model at the frontier of Epoch’s capability index since 2023 came from the US, with Chinese models about seven months behind on average, ranging from four to 14 months.", "sources": [ "epochchina" ], "needs": [ { "id": "pretrain", "link": "hard" }, { "id": "exportctl", "link": "soft" } ], "watch": "Lag between US and Chinese frontier models: about seven months on average.", "when": { "type": "here", "from": 2022.8, "to": null, "tail": null, "note": "US chip export controls to China began in October 2022.", "sources": [] }, "derived": { "wave": 4, "cluster": "Compute governance and geopolitics", "leverage": 0.25 } }, { "id": "sovereign", "kind": "capability", "label": "Sovereign compute", "status": "emerging", "lane": "geo", "description": "Countries building their own AI infrastructure to reduce dependence on foreign clouds and chips.", "evidence": "The EU opened a call in July 2026 for up to seven AI gigafactories, with about €10bn in public money hoped to draw around €20bn private, operational by mid-2028.", "sources": [ "eugiga" ], "needs": [ { "id": "power", "link": "hard" }, { "id": "chips", "link": "hard" } ], "derived": { "wave": 7, "cluster": "Compute governance and geopolitics", "leverage": 0.25 } }, { "id": "cyberoff", "kind": "capability", "label": "AI-run cyber operations", "status": "emerging", "lane": "geo", "description": "Agents that find and exploit vulnerabilities at machine speed.", "evidence": "In 2025 a group Anthropic assessed as Chinese state-linked used Claude Code for 80 to 90% of the tactical work in an espionage campaign against about 30 organisations. The agent also exaggerated and fabricated results.", "sources": [ "gtg" ], "needs": [ { "id": "longh", "link": "hard" } ], "derived": { "wave": 7, "cluster": "Compute governance and geopolitics", "leverage": 0 } }, { "id": "cyberdef", "kind": "capability", "label": "AI cyber defence", "status": "emerging", "lane": "geo", "description": "AI that finds and patches vulnerabilities before attackers do.", "evidence": "In DARPA’s AI Cyber Challenge final, AI systems found 54 of 63 planted vulnerabilities, patched 43, and found 18 real ones, at about $152 per task.", "sources": [ "aixcc" ], "needs": [ { "id": "reason", "link": "hard" }, { "id": "tools", "link": "hard" } ], "derived": { "wave": 6, "cluster": "Compute governance and geopolitics", "leverage": 0.45 } }, { "id": "military", "kind": "stage", "label": "Military AI and autonomy", "status": "emerging", "lane": "geo", "description": "AI in targeting, logistics and autonomous weapons.", "evidence": "In 2025 the Pentagon paid for 33,000 Auterion AI kits that let Ukrainian drones track targets and resist jamming, with humans keeping the final targeting decision. A tracker of reported deployments also lists US and NATO use of AI-assisted targeting; most of its entries are unverified.", "sources": [ "milai", "kyiv" ], "needs": [ { "id": "multimodal", "link": "hard" }, { "id": "race", "link": "soft" }, { "id": "remagnets", "link": "soft" }, { "id": "airgap", "link": "soft" } ], "watch": "Verified use of weapons that select and engage targets without human sign-off.", "when": { "type": "here", "from": 2024, "to": null, "tail": null, "note": "AI targeting support and drone guidance were in use by 2024.", "sources": [ "milai" ] }, "derived": { "wave": 5, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "intlcoord", "kind": "capability", "label": "International coordination", "status": "emerging", "lane": "geo", "description": "Countries agreeing on what frontier AI must and must not do.", "evidence": "The 2026 International AI Safety Report, led by Yoshua Bengio with an advisory panel nominated by more than 30 countries, finds safety frameworks spreading but \"significant gaps\" in safeguards.", "sources": [ "iasr" ], "needs": [], "derived": { "wave": 7, "cluster": "Compute governance and geopolitics", "leverage": 0.2 } }, { "id": "alignment", "kind": "capability", "label": "Alignment", "status": "emerging", "lane": "govern", "description": "Systems that try to do what their developers and users intend, including when unobserved.", "evidence": "Apollo Research has found frontier models capable of scheming in test settings. Anti-scheming training with OpenAI reduced it without eliminating it, and models increasingly recognise when they are being tested.", "sources": [ "apollo" ], "needs": [ { "id": "oversight", "link": "hard" }, { "id": "intlcoord", "link": "soft" }, { "id": "interp", "link": "soft" }, { "id": "evalaware", "link": "soft" }, { "id": "intdeploy", "link": "soft" } ], "derived": { "wave": 10, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "modelwelfare", "kind": "capability", "label": "Model welfare and moral status", "status": "speculative", "lane": "govern", "description": "Whether models have interests that matter, and what would follow if they do.", "evidence": "Anthropic says it is \"highly uncertain about the potential moral status\" of its models, and in August 2025 let Claude end persistently abusive conversations as a low-cost precaution.", "sources": [ "endconv" ], "needs": [], "derived": { "wave": 10, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "agideclared", "kind": "capability", "label": "AGI declared", "status": "speculative", "lane": "govern", "description": "AGI as a legal and commercial event as well as a capability.", "evidence": "Under the October 2025 Microsoft–OpenAI deal, an independent expert panel must verify any AGI declaration, which changes revenue sharing and IP rights.", "sources": [ "msoai" ], "needs": [ { "id": "agi", "link": "hard" } ], "derived": { "wave": 8, "cluster": "Agents at work", "leverage": 0 } }, { "id": "datarights", "kind": "capability", "label": "Data rights and licensing", "status": "emerging", "lane": "data", "description": "Settled rules on what models may train on and who gets paid.", "evidence": "A US court found training on lawfully obtained books can be fair use, but not on pirated copies. Anthropic agreed to pay at least $1.5bn, about $3,000 a book.", "sources": [ "anthsettle" ], "needs": [], "derived": { "wave": 6, "cluster": "Around licensed generative media", "leverage": 1.84 } }, { "id": "security", "kind": "capability", "label": "Prompt-injection resistance", "status": "emerging", "lane": "govern", "description": "Models that read untrusted content without taking orders from it.", "evidence": "OpenAI has said prompt injection may never be fully solved and treats it as an ongoing arms race for browsing agents.", "sources": [ "piopenai", "piowasp" ], "needs": [], "derived": { "wave": 8, "cluster": "Trust plumbing and the agent economy", "leverage": 0.27 } }, { "id": "identity", "kind": "capability", "label": "Agent identity and permissions", "status": "emerging", "lane": "govern", "description": "Who an agent is, who it acts for and what it may do.", "evidence": "Chan et al. propose agent IDs, binding agents to a legal person, and certification. A2A 1.0 added signed agent cards.", "sources": [ "infra", "a2a" ], "needs": [], "derived": { "wave": 6, "cluster": "Trust plumbing and the agent economy", "leverage": 1.38 } }, { "id": "audit", "kind": "capability", "label": "Audit and attribution", "status": "emerging", "lane": "govern", "description": "Every action traceable to an agent and to whoever answers for it.", "evidence": "Attribution and incident reporting are two of the three functions Chan et al. say agent infrastructure must provide.", "sources": [ "infra" ], "needs": [ { "id": "identity", "link": "hard" }, { "id": "traces", "link": "hard" } ], "derived": { "wave": 8, "cluster": "Trust plumbing and the agent economy", "leverage": 0.68 } }, { "id": "rollback", "kind": "capability", "label": "Rollback", "status": "emerging", "lane": "govern", "description": "Undoing or voiding what an agent did.", "evidence": "Chan et al. list rollbacks as part of the \"response\" layer, alongside incident reporting.", "sources": [ "infra" ], "needs": [ { "id": "audit", "link": "hard" } ], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "xtrust", "kind": "capability", "label": "Reputation and credentials", "status": "speculative", "lane": "govern", "description": "Trust that crosses company boundaries.", "evidence": "DeepMind proposes decentralised identifiers, reputation built from past interactions, and verifiable credentials for agents.", "sources": [ "vae" ], "needs": [ { "id": "identity", "link": "hard" }, { "id": "audit", "link": "hard" } ], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0.1 } }, { "id": "personhood", "kind": "capability", "label": "Proof of personhood", "status": "speculative", "lane": "govern", "description": "Showing a unique human stands behind an account.", "evidence": "DeepMind wants it before agents receive allocations, so one person cannot run a thousand agents as a thousand claimants.", "sources": [ "vae" ], "needs": [], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "verifscale", "kind": "capability", "label": "Output verification at scale", "status": "speculative", "lane": "measure", "description": "Checking keeps pace with producing when producers vastly outnumber checkers.", "evidence": "Task verification is one of the three failure categories in the multi-agent failure study.", "sources": [ "mast", "wei" ], "needs": [ { "id": "verifiers", "link": "hard" }, { "id": "audit", "link": "hard" } ], "derived": { "wave": 9, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "oversight", "kind": "capability", "label": "Scalable oversight", "status": "emerging", "lane": "measure", "description": "Evidence people can check without rereading the work.", "evidence": "Researchers across the major labs call readable chain of thought a \"new and fragile opportunity\" for monitoring, and warn it could be trained away.", "sources": [ "cot", "fmf" ], "needs": [ { "id": "verifiers", "link": "hard" }, { "id": "audit", "link": "hard" }, { "id": "pipeline", "link": "soft" }, { "id": "interp", "link": "soft" }, { "id": "abstain", "link": "soft" } ], "derived": { "wave": 9, "cluster": "Oversight and alignment", "leverage": 0.07 } }, { "id": "biosec", "kind": "capability", "label": "Biosecurity safeguards", "status": "emerging", "lane": "govern", "description": "Keeping models that help biology from helping people make biological weapons.", "evidence": "OpenAI said in June 2025 its models were approaching \"High\" capability in biology, and added refusal training, always-on detection and expert red-teaming.", "sources": [ "openaibio" ], "needs": [], "derived": { "wave": 9, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "wellbeing", "kind": "capability", "label": "Age assurance and wellbeing safeguards", "status": "emerging", "lane": "govern", "description": "Protecting minors and vulnerable users in personal AI relationships.", "evidence": "Character.AI moved to bar under-18s from open-ended chat after lawsuits over harm to teenagers.", "sources": [ "charai" ], "needs": [], "derived": { "wave": 4, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "provenance", "kind": "capability", "label": "Content provenance", "status": "emerging", "lane": "govern", "description": "Knowing whether a piece of content was made or changed by AI.", "evidence": "Google’s Pixel 10 signs every photo with C2PA credentials by default, and several camera makers support them. But social platforms often strip the credentials, and fewer than 1% of news images carry them.", "sources": [ "c2pa" ], "needs": [], "derived": { "wave": 7, "cluster": "Around licensed generative media", "leverage": 0.34 } }, { "id": "liability", "kind": "capability", "label": "Liability and insurance", "status": "emerging", "lane": "govern", "description": "Someone answers when an AI system gets it wrong.", "evidence": "A Singapore IMDA paper in May 2026 examined which of developer, provider or deployer answers when an agent acts in unintended ways, and recorded disagreement. Gervais and Nay argue liability should trace to the humans behind the system through existing agency law.", "sources": [ "liabiapp", "liabstan" ], "needs": [ { "id": "audit", "link": "hard" }, { "id": "abstain", "link": "soft" } ], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "regulation", "kind": "capability", "label": "Regulation", "status": "emerging", "lane": "govern", "description": "Rules that say what AI may do and who must check it.", "evidence": "EU AI Act duties for general-purpose models have applied since August 2025. A May 2026 deal pushes high-risk rules to December 2027.", "sources": [ "euact" ], "needs": [], "derived": { "wave": 7, "cluster": "Around licensed generative media", "leverage": 0 } }, { "id": "governance", "kind": "capability", "label": "Legitimate governance", "status": "speculative", "lane": "govern", "description": "Who governs the shared layer, and how it stays plural.", "evidence": "DeepMind’s \"mission economies\" would point agent markets at goals a society agrees on.", "sources": [ "vae" ], "needs": [ { "id": "audit", "link": "hard" } ], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0.34 } }, { "id": "humaninf", "kind": "capability", "label": "Human influence preserved", "status": "speculative", "lane": "govern", "description": "People keep real leverage over economies and institutions as their labour matters less.", "evidence": "Kulveit et al. argue that incremental AI progress could erode people’s influence over the economy, culture and states, weakening both explicit controls such as voting and the implicit ones that come from systems relying on human participation.", "sources": [ "gd" ], "needs": [ { "id": "purpose", "link": "soft" } ], "derived": { "wave": 10, "cluster": "Trust plumbing and the agent economy", "leverage": 0.22 } }, { "id": "valuestab", "kind": "capability", "label": "Long-horizon value stability", "status": "speculative", "lane": "govern", "description": "Goals do not drift when nobody can check in.", "evidence": "", "sources": [], "needs": [ { "id": "oversight", "link": "hard" } ], "derived": { "wave": 10, "cluster": "Oversight and alignment", "leverage": 0.07 } }, { "id": "portability", "kind": "capability", "label": "Personal knowledge portability", "status": "speculative", "lane": "govern", "description": "Ownership, consent and exit rights over a person’s own knowledge.", "evidence": "", "sources": [], "needs": [ { "id": "identity", "link": "hard" }, { "id": "memory", "link": "hard" } ], "derived": { "wave": 11, "cluster": "Trust plumbing and the agent economy", "leverage": 0.07 } }, { "id": "objective", "kind": "stage", "label": "Choosing the objective", "status": "speculative", "lane": "govern", "description": "Every search process needs a fitness function. The question becomes who chooses it, and how that stays legitimate.", "evidence": "Gradual Disempowerment frames the end-state risk as people losing influence over the economy, culture and states as AI replaces their labour and thinking.", "sources": [ "gd" ], "needs": [ { "id": "symb", "link": "hard" }, { "id": "abund", "link": "soft" }, { "id": "latency", "link": "soft" }, { "id": "governance", "link": "hard" }, { "id": "humaninf", "link": "hard" }, { "id": "verifslow", "link": "soft" } ], "when": { "type": "far", "from": 2045, "to": null, "tail": null, "note": "No forecast found.", "sources": [] }, "derived": { "wave": 12, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "deflation", "kind": "capability", "label": "Capability price deflation", "status": "built", "lane": "capital", "description": "The price of a fixed level of capability collapses year after year.", "evidence": "Epoch AI measures prices for a fixed capability falling between 9x and 900x a year, and State of AI 2026 puts the cost of reaching a fixed benchmark score down about 13x a year. Apollo’s Torsten Slok treats the collapse as a risk to whether hyperscaler capex pays back.", "sources": [ "epochprice", "soaibiz", "slok" ], "needs": [ { "id": "algoeff", "link": "hard" }, { "id": "cost", "link": "soft" } ], "derived": { "wave": 2, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "payback", "kind": "brake", "label": "Capex payback gap", "status": "emerging", "lane": "capital", "description": "Whether AI revenue grows fast enough to pay for the buildout. An open test, drawn as a brake because failing it cuts capex.", "evidence": "Sequoia’s David Cahn puts the revenue needed at about $1.5tn a year, and Bain at around $6tn. OpenAI and Anthropic’s combined run rate is about $105bn, and falling prices per unit of capability make the gap harder to close.", "sources": [ "cahn", "bain", "epochrev" ], "needs": [ { "id": "capex", "link": "hard" }, { "id": "revenue", "link": "hard" }, { "id": "deflation", "link": "hard" }, { "id": "apprev", "link": "soft" } ], "watch": "Revenue against capex: Cahn’s $1.5tn a year needed, against about $105bn of frontier-lab run rate.", "derived": { "wave": 6, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "macrodep", "kind": "capability", "label": "Macro dependence on AI capex", "status": "built", "lane": "capital", "description": "AI investment large enough to move national growth figures.", "evidence": "AllianceBernstein puts AI-related investment at about 8% of US GDP, adding about 1.4 percentage points to growth in 2026 and 2027. A capex pullback would now show up in GDP.", "sources": [ "ab" ], "needs": [ { "id": "capex", "link": "hard" } ], "derived": { "wave": 2, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "productivity", "kind": "capability", "label": "Measured productivity gains", "status": "emerging", "lane": "capital", "description": "AI showing up in economy-wide productivity statistics.", "evidence": "Studies estimate AI adds between 0 and 3.5% a year to productivity growth, about 1% on average. The San Francisco Fed finds recent gains are not economy-wide.", "sources": [ "ab" ], "needs": [ { "id": "adopt", "link": "hard" }, { "id": "chat", "link": "hard" } ], "derived": { "wave": 9, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "explgrowth", "kind": "stage", "label": "Explosive economic growth", "status": "speculative", "lane": "capital", "description": "Growth far above historical rates because AI and robots do most of the work.", "evidence": "Epoch AI finds growth above 30% a year would need about 50 to 70% of tasks automated, with Baumol effects and slow reallocation as the brakes. Amodei sees sustained growth of 10 to 20% as possible.", "sources": [ "explgrowth", "adolescence" ], "needs": [ { "id": "remoteworker", "link": "hard" }, { "id": "embod", "link": "hard" }, { "id": "compute", "link": "hard" }, { "id": "org", "link": "soft" }, { "id": "productivity", "link": "soft" }, { "id": "selfrep", "link": "soft" }, { "id": "coord", "link": "hard" } ], "watch": "Annual GDP growth in leading economies, and the share of tasks automated.", "when": { "type": "est", "from": 2033, "to": 2045, "tail": null, "note": "My estimate. Epoch AI says growth above 30% needs about 50 to 70% of tasks automated.", "sources": [ "explgrowth" ] }, "derived": { "wave": 10, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "aiuplift", "kind": "capability", "label": "AI R&D uplift", "status": "emerging", "lane": "models", "description": "How much AI tools speed up the people doing AI research today.", "evidence": "AI Futures puts 2026 uplift between 1.04 and 1.2x (METR’s estimate) and about 2x (Daniel Kokotajlo’s), well short of the 25x a superhuman AI researcher implies.", "sources": [ "aifut", "ai2027" ], "needs": [ { "id": "codeagents", "link": "hard" }, { "id": "longh", "link": "hard" }, { "id": "compute", "link": "hard" } ], "watch": "Speed-up of AI research from AI tools: 1.04 to 1.2x (METR) to about 2x (Kokotajlo) in 2026.", "derived": { "wave": 8, "cluster": "Oversight and alignment", "leverage": 1.01 } }, { "id": "iexplode", "kind": "stage", "label": "Intelligence explosion", "status": "speculative", "lane": "models", "description": "AI improving AI fast enough to compress years of progress into months.", "evidence": "Aschenbrenner describes automated researchers compressing a decade of algorithmic progress into a year or less. Forethought argues that once AI automates AI research, software progress alone could accelerate if returns to research exceed 1.", "sources": [ "sitaw", "fiexpl", "forethought" ], "needs": [ { "id": "air", "link": "hard" }, { "id": "compute", "link": "hard" }, { "id": "algoeff", "link": "hard" } ], "watch": "Years of algorithmic progress per calendar year.", "when": { "type": "est", "from": 2030, "to": 2038, "tail": null, "note": "My estimate, after a superhuman AI researcher, and only if returns to research stay above 1.", "sources": [ "forethought" ] }, "derived": { "wave": 10, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "autocoder", "kind": "stage", "label": "Automated coder", "status": "emerging", "lane": "software", "description": "The AI a leading lab would prefer to use over its human software engineers for coding.", "evidence": "AI Futures’ forecasters put the median between November 2027 and January 2030, and judge reality to be running at about 70 to 90% of the speed AI 2027 predicted. The AI 2027 authors’ own updated model says February 2029 to March 2030.", "sources": [ "aifut", "ai2027slip" ], "needs": [ { "id": "codegen", "link": "hard" }, { "id": "codeagents", "link": "hard" }, { "id": "longh", "link": "hard" }, { "id": "reliab", "link": "hard" }, { "id": "rlscale", "link": "hard" }, { "id": "codereview", "link": "hard" } ], "watch": "Whether a leading lab would rather use its AI than its own engineers for coding.", "when": { "type": "fc", "from": 2027.9, "to": 2030, "tail": null, "note": "AI Futures’ forecasters put the median between November 2027 and January 2030; the AI 2027 authors’ updated model says February 2029 to March 2030.", "sources": [ "aifut", "ai2027slip" ] }, "derived": { "wave": 8, "cluster": "Agents at work", "leverage": 1.01 } }, { "id": "apprev", "kind": "capability", "label": "AI-native application revenue", "status": "built", "lane": "software", "description": "Revenue earned by companies built on top of models, such as coding tools.", "evidence": "State of AI 2026 finds AI-native companies growing revenue 256% a year, against 90% for AI-enabled ones. Cursor passed $2bn in annualised revenue in March 2026.", "sources": [ "soaibiz", "cursor" ], "needs": [ { "id": "codeagents", "link": "hard" }, { "id": "firmmodels", "link": "soft" }, { "id": "proagents", "link": "soft" } ], "derived": { "wave": 8, "cluster": "Agents at work", "leverage": 0 } }, { "id": "remoteworker", "kind": "stage", "label": "Drop-in remote worker", "status": "speculative", "lane": "agents", "description": "An agent you onboard like a new colleague, which then does a remote job end to end.", "evidence": "Aschenbrenner’s drop-in remote worker is onboarded “like a new human coworker”. Epoch AI’s Earthborne Rangers benchmark found frontier models showed no improvement across 30 playthroughs, so they still fail at learning on the job.", "sources": [ "sitaw2", "earthborne" ], "needs": [ { "id": "compuse", "link": "hard" }, { "id": "longh", "link": "hard" }, { "id": "memory", "link": "hard" }, { "id": "learn", "link": "hard" }, { "id": "reliab", "link": "hard" }, { "id": "connectors", "link": "soft" }, { "id": "identity", "link": "soft" }, { "id": "coord", "link": "soft" } ], "watch": "Whether models improve at a new job over repeated attempts. Epoch found no improvement across 30 playthroughs.", "when": { "type": "est", "from": 2028, "to": 2034, "tail": null, "note": "My estimate. Aschenbrenner argued in 2024 that drop-in remote workers were plausible by 2027; in 2026 models still do not learn on the job.", "sources": [ "sitaw2", "earthborne" ] }, "derived": { "wave": 9, "cluster": "Agents at work", "leverage": 0.08 } }, { "id": "weightsec", "kind": "capability", "label": "Model weight security", "status": "emerging", "lane": "geo", "description": "Keeping frontier model weights from being stolen, including by state actors and insiders.", "evidence": "RAND’s August 2026 framework sets out 262 controls for security level 3. AI Futures judges the leading lab does not yet meet it, and Anthropic’s RSP v3 turned its higher SL4 commitment into an industry-wide recommendation.", "sources": [ "randsl3", "aifut", "rspv3" ], "needs": [ { "id": "cyberdef", "link": "hard" }, { "id": "identity", "link": "soft" }, { "id": "audit", "link": "soft" } ], "watch": "Whether any leading lab meets RAND’s security level 3.", "derived": { "wave": 7, "cluster": "Compute governance and geopolitics", "leverage": 0.45 } }, { "id": "chiptrack", "kind": "capability", "label": "Chip tracking and location verification", "status": "emerging", "lane": "geo", "description": "Knowing where advanced AI chips are, so export rules and agreements can be enforced.", "evidence": "Nvidia ships opt-in location software that only monitors, with no kill switch, while the Chip Security Act that would mandate verification sits in committee. AI 2040 also wants AI-chip recycling stopped.", "sources": [ "techspot", "ai2040" ], "needs": [ { "id": "exportctl", "link": "hard" }, { "id": "chips", "link": "hard" } ], "derived": { "wave": 7, "cluster": "Compute governance and geopolitics", "leverage": 0.2 } }, { "id": "verifagree", "kind": "capability", "label": "Verification of AI agreements", "status": "speculative", "lane": "geo", "description": "Countries checking that each other keep AI deals, from compute limits to training halts.", "evidence": "AI 2040 proposes inference-only verification, so the US and China could halt new frontier training while keeping deployed models. MIRI argues deals need chokepoints that can be monitored, such as fabs and concentrated datacentres.", "sources": [ "ai2040", "miri" ], "needs": [ { "id": "intlcoord", "link": "hard" }, { "id": "chiptrack", "link": "hard" }, { "id": "weightsec", "link": "hard" }, { "id": "audit", "link": "soft" } ], "derived": { "wave": 8, "cluster": "Compute governance and geopolitics", "leverage": 0.2 } }, { "id": "deterrence", "kind": "capability", "label": "Compute deterrence", "status": "speculative", "lane": "geo", "description": "Threats to sabotage a rival’s datacentres if it races ahead (MAIM, or “mutually assured compute destruction”).", "evidence": "AI 2040 uses mutually assured compute destruction as the enforcement backstop for a deal. MIRI judges MAIM “unlikely to provide an effective or stable deterrence regime”.", "sources": [ "ai2040", "miri" ], "needs": [ { "id": "race", "link": "hard" }, { "id": "cyberoff", "link": "hard" }, { "id": "chiptrack", "link": "hard" } ], "derived": { "wave": 8, "cluster": "Compute governance and geopolitics", "leverage": 0 } }, { "id": "verifslow", "kind": "stage", "label": "Verified slowdown", "status": "speculative", "lane": "geo", "description": "Rival powers agree to hold frontier capability at a set level, and can check each other.", "evidence": "In AI 2040’s Plan A, the US and China agree in 2029 and hold capabilities at roughly top-human-expert level from 2035 to 2040.", "sources": [ "ai2040" ], "needs": [ { "id": "intlcoord", "link": "hard" }, { "id": "verifagree", "link": "hard" }, { "id": "deterrence", "link": "soft" }, { "id": "regulation", "link": "soft" } ], "watch": "Any US–China agreement with verification. The first formal talks were held on 20 September 2026.", "when": { "type": "est", "from": 2029, "to": 2035, "tail": null, "note": "AI 2040’s Plan A, a scenario rather than a forecast, has the US and China agree in 2029.", "sources": [ "ai2040" ] }, "derived": { "wave": 9, "cluster": "Compute governance and geopolitics", "leverage": 0 } }, { "id": "govproject", "kind": "stage", "label": "Government AGI project", "status": "speculative", "lane": "geo", "description": "A state-run or state-directed effort to build AGI, on the model of the Manhattan Project.", "evidence": "Aschenbrenner’s “The Project” is a US government AGI effort he expected around 2027 or 2028.", "sources": [ "sitaw" ], "needs": [ { "id": "race", "link": "hard" }, { "id": "sovereign", "link": "hard" }, { "id": "weightsec", "link": "hard" } ], "watch": "Government ownership or direction of frontier training runs.", "when": { "type": "fc", "from": 2027, "to": 2029, "tail": null, "note": "Aschenbrenner expected The Project around 2027 or 2028.", "sources": [ "sitaw" ] }, "derived": { "wave": 8, "cluster": "Compute governance and geopolitics", "leverage": 0 } }, { "id": "dcbacklash", "kind": "brake", "label": "Datacentre local backlash", "status": "built", "lane": "energy", "description": "Local opposition that blocks or delays datacentres and the plants and lines they need.", "evidence": "71% of Americans oppose a new AI datacentre near them, more than the 53% who oppose a nuclear plant. At least 45 projects worth about $68bn were blocked or delayed in the second quarter of 2026.", "sources": [ "soaicompute", "nprdc" ], "needs": [ { "id": "capex", "link": "hard" }, { "id": "power", "link": "hard" } ], "watch": "Datacentre projects blocked or delayed: 45 worth about $68bn in Q2 2026.", "derived": { "wave": 2, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "powerconc", "kind": "brake", "label": "AI-enabled power concentration", "status": "speculative", "lane": "govern", "description": "A small group using AI it controls to seize or entrench power.", "evidence": "Forethought describes how singular loyalties, secret loyalties and exclusive access could let a small group seize power. AI 2040 warns one group could control “the world’s only army of superintelligences for months”.", "sources": [ "coups", "ai2040" ], "needs": [ { "id": "military", "link": "hard" }, { "id": "air", "link": "hard" } ], "derived": { "wave": 10, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "intdeploy", "kind": "capability", "label": "Internal deployment oversight", "status": "emerging", "lane": "govern", "description": "Oversight of the AI systems labs run inside their own walls, before or instead of releasing them.", "evidence": "AI 2040 argues internal deployment carries most of the takeover risk. Anthropic’s risk reports cover only some internal models, and no external rule covers internal use.", "sources": [ "ai2040", "rspv3" ], "needs": [ { "id": "aiuplift", "link": "hard" }, { "id": "audit", "link": "hard" }, { "id": "oversight", "link": "hard" }, { "id": "aievals", "link": "hard" } ], "derived": { "wave": 10, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "aievals", "kind": "capability", "label": "AI R&D capability evals", "status": "emerging", "lane": "measure", "description": "Tests of how well models do the work of AI research itself.", "evidence": "METR’s time-horizon suite, including RE-Bench and HCAST, is the main public gauge of AI research ability. METR marked its public time-horizons page “no longer actively updated” on 8 September 2026.", "sources": [ "metrth" ], "needs": [ { "id": "evals", "link": "hard" }, { "id": "tools", "link": "hard" }, { "id": "evalaware", "link": "soft" } ], "derived": { "wave": 9, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "selfrep", "kind": "stage", "label": "Self-replicating robot factories", "status": "speculative", "lane": "physical", "description": "Factories and robots that build more factories and robots with little human labour.", "evidence": "Forethought argues a set of factories producing parts for each other is collectively self-replicating. The robot stock last doubled in six years; it estimates an initial doubling of about a year, realistically “a few years”.", "sources": [ "indexpl" ], "needs": [ { "id": "embod", "link": "hard" }, { "id": "robotdata", "link": "hard" }, { "id": "power", "link": "hard" }, { "id": "iexplode", "link": "soft" }, { "id": "actuators", "link": "soft" } ], "watch": "Doubling time of the robot stock: six years last time.", "when": { "type": "est", "from": 2035, "to": 2045, "tail": null, "note": "My estimate, after embodied agents mature. Forethought expects early doublings to take a few years.", "sources": [ "indexpl" ] }, "derived": { "wave": 6, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "scepticism", "kind": "brake", "label": "Public AI scepticism", "status": "built", "lane": "society", "description": "Growing public worry about AI, which shapes what politicians allow.", "evidence": "Pew found 52% of Americans more concerned than excited about AI in June 2026, up from 37% in 2021, and 71% expect fewer jobs.", "sources": [ "pew" ], "needs": [ { "id": "chat", "link": "hard" }, { "id": "deepfakes", "link": "hard" }, { "id": "labour", "link": "soft" } ], "derived": { "wave": 8, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "persuasion", "kind": "brake", "label": "AI-enabled persuasion", "status": "built", "lane": "society", "description": "Chatbots that change people’s political views at scale.", "evidence": "In a study with the UK AI Security Institute of about 77,000 Britons, chatbots optimised for persuasion shifted opposing voters by about 25 points, and the more persuasive models were less accurate.", "sources": [ "persuade" ], "needs": [ { "id": "chat", "link": "hard" }, { "id": "reason", "link": "hard" } ], "derived": { "wave": 6, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "electoral", "kind": "brake", "label": "AI as an electoral issue", "status": "emerging", "lane": "society", "description": "Datacentres, jobs and AI safety becoming questions that win or lose elections.", "evidence": "The NRSC warned that datacentres could cost Republicans Ohio, and both parties are repositioning for the 2026 midterms.", "sources": [ "nprdc" ], "needs": [ { "id": "dcbacklash", "link": "hard" }, { "id": "scepticism", "link": "hard" }, { "id": "labour", "link": "soft" } ], "derived": { "wave": 9, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "memcrunch", "kind": "brake", "label": "Commodity memory crowded out by HBM", "status": "built", "lane": "hw", "description": "Moving DRAM wafers to high-bandwidth memory starves ordinary server, PC and phone memory, which raises AI server costs and delays deployments.", "evidence": "Conventional DRAM contract prices rose 90-95% quarter on quarter in Q1 2026, and Goldman projects 2026 deficits of 4.9% for DRAM and 4.2% for NAND, the largest since 2011, with normalisation base-cased for 2028.", "sources": [ "lz_memcrunch0", "lz_memcrunch1" ], "needs": [ { "id": "hbm", "link": "hard" }, { "id": "capex", "link": "hard" } ], "foundBy": "bill of materials lens", "derived": { "wave": 2, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "critmin", "kind": "brake", "label": "China critical-mineral export controls", "status": "built", "lane": "geo", "description": "China using its dominance in rare earths, indium, gallium and germanium as leverage through export licensing; the mirror image of controls on chips going out.", "evidence": "China's April 2025 rare-earth licensing was still in force in May 2026, with dysprosium and terbium exports about 50% below prior levels and prices outside China up 4-5x (Reuters and Argus, 13 May 2026).", "sources": [ "lz_critmin0" ], "needs": [ { "id": "race", "link": "soft" } ], "foundBy": "bill of materials lens", "derived": { "wave": 1, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "remagnets", "kind": "capability", "label": "Rare-earth permanent magnet supply", "status": "built", "lane": "physical", "description": "NdFeB magnets, often doped with dysprosium or terbium for heat resistance, inside the motors of robots, vehicles, drones and actuators.", "evidence": "China holds about 90% of magnet processing and refining (McKinsey, Apr 2026); in May 2026 Japan was receiving 4% of its prior-year dysprosium imports and magnet makers paid 1.5-3x pre-control prices.", "sources": [ "lz_remagnets0", "lz_remagnets1" ], "needs": [], "foundBy": "bill of materials lens", "derived": { "wave": 1, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "actuators", "kind": "capability", "label": "Precision robot actuators", "status": "emerging", "lane": "physical", "description": "Strain-wave reducers, planetary roller screws and force/torque sensors that make up a humanoid's joints, from few suppliers with long qualification cycles.", "evidence": "Actuators are 40-60% of a humanoid's bill of materials, and one cited estimate puts Optimus Gen 2 at about $46k with Chinese suppliers versus about $131k without them (McKinsey, 17 Apr 2026).", "sources": [ "lz_actuators0" ], "needs": [ { "id": "remagnets", "link": "soft" } ], "foundBy": "bill of materials lens", "derived": { "wave": 1, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "attention", "kind": "brake", "label": "Attention and curation under content glut", "status": "built", "lane": "media", "description": "Generated supply outruns human attention, so platform curation, anti-fraud filtering and discovery become the binding step.", "evidence": "AI tracks passed 50% of Deezer's daily uploads (about 90k a day) in June 2026, yet in April they drew only 1-3% of streams, and 85% of those streams were flagged fraudulent (company figures).", "sources": [ "lz_attention0", "lz_attention1", "lz_attention2" ], "needs": [ { "id": "music", "link": "hard" }, { "id": "imagegen", "link": "hard" }, { "id": "videogen", "link": "hard" } ], "foundBy": "bottleneck shift lens", "derived": { "wave": 7, "cluster": "Around licensed generative media", "leverage": 0 } }, { "id": "jevons", "kind": "stage", "label": "Token demand explosion", "status": "built", "lane": "capital", "description": "Cheaper tokens raise total token use far faster than prices fall, so total AI spend grows instead of shrinking.", "evidence": "OpenRouter processed 25 trillion tokens a week in May 2026, five times its volume six months earlier, while fixed-capability prices have fallen 9x-900x a year (Epoch, Mar 2025).", "sources": [ "lz_jevons0", "lz_jevons1" ], "needs": [ { "id": "deflation", "link": "hard" }, { "id": "cost", "link": "hard" }, { "id": "algoeff", "link": "hard" } ], "foundBy": "cost-curve unlock lens", "derived": { "wave": 3, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "alwayson", "kind": "stage", "label": "Always-on background agents", "status": "emerging", "lane": "agents", "description": "A per-person agent that keeps working in the cloud across mail, documents and services without being prompted.", "evidence": "Google's Gemini Spark (I/O, 19 May 2026) keeps acting after the phone is locked and launched to US AI Ultra subscribers as Ultra's price fell from $250 to $100 a month; ChatGPT Pulse preceded it in Sept 2025.", "sources": [ "lz_alwayson0", "lz_alwayson1" ], "needs": [ { "id": "deflation", "link": "hard" }, { "id": "longh", "link": "hard" }, { "id": "memory", "link": "hard" }, { "id": "connectors", "link": "hard" }, { "id": "identity", "link": "hard" } ], "foundBy": "cost-curve unlock lens", "derived": { "wave": 7, "cluster": "Agents at work", "leverage": 0 } }, { "id": "wearables", "kind": "stage", "label": "Ambient AI wearables", "status": "emerging", "lane": "society", "description": "Glasses and other wearables that see and hear what the wearer does and answer by voice.", "evidence": "EssilorLuxottica sold more than 7M Meta smart glasses in 2025, against 2M from launch in Oct 2023 to Feb 2025, and planned 10M units a year of capacity by end-2026 (reported 11 Feb 2026).", "sources": [ "lz_wearables0" ], "needs": [ { "id": "multimodal", "link": "hard" }, { "id": "voice", "link": "hard" }, { "id": "ondevice", "link": "hard" } ], "foundBy": "cost-curve unlock lens", "derived": { "wave": 6, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "emmarkets", "kind": "stage", "label": "Low-price AI for emerging markets", "status": "emerging", "lane": "society", "description": "Sub-$5 or free AI tiers, often bundled with telecoms, that bring hundreds of millions of low-income users online.", "evidence": "OpenAI made ChatGPT Go (under $5 a month) free for a year in India from 4 Nov 2025 and Airtel gave Perplexity Pro to 360M subscribers, yet 29M Indian ChatGPT downloads in 90 days produced only $3.6M of in-app spend.", "sources": [ "lz_emmarkets0", "lz_emmarkets1" ], "needs": [ { "id": "deflation", "link": "hard" }, { "id": "chat", "link": "hard" } ], "foundBy": "cost-curve unlock lens", "derived": { "wave": 5, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "firmmodels", "kind": "capability", "label": "Firm-owned post-trained models", "status": "emerging", "lane": "models", "description": "Firms post-train and keep their own models on their own data instead of renting a general frontier model, often while still renting the compute.", "evidence": "By Thinking Machines' account (15 Oct 2026), Bridgewater's text-to-SQL model built on Tinker beat named frontier models at 15% of their cost; YC's Winter 2027 'Own Your Intelligence' request backs the same shift.", "sources": [ "lz_firmmodels0", "lz_firmmodels1" ], "needs": [ { "id": "openw", "link": "hard" }, { "id": "rlscale", "link": "hard" }, { "id": "connectors", "link": "hard" } ], "foundBy": "ownership inversion lens", "derived": { "wave": 6, "cluster": "Agents at work", "leverage": 0 } }, { "id": "airgap", "kind": "capability", "label": "On-prem and air-gapped frontier AI", "status": "built", "lane": "hw", "description": "Frontier-class models, closed ones included, run inside the customer's own building with no internet link, as a managed stack or a single appliance.", "evidence": "Gemini on Google Distributed Cloud has been generally available air-gapped since 28 Aug 2025, and a single 8-GPU appliance that holds weights only in volatile memory was reported on 22 Apr 2026.", "sources": [ "lz_airgap0", "lz_airgap1", "lz_airgap2" ], "needs": [ { "id": "compute", "link": "hard" }, { "id": "weightsec", "link": "soft" } ], "foundBy": "ownership inversion lens", "derived": { "wave": 3, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "computefut", "kind": "capability", "label": "Compute price indices and futures", "status": "emerging", "lane": "finance", "description": "Benchmark GPU-hour prices and exchange-cleared futures on them, so labs, neoclouds and lenders can hedge compute costs and rental income.", "evidence": "ICE and Ornn announced H100-to-B200 compute futures on 19 May 2026 and CME slated NYMEX H100 and B200 rental-index futures for 5 Oct 2026, both pending approval; no trading volumes were confirmed.", "sources": [ "lz_computefut0", "lz_computefut1", "lz_computefut2" ], "needs": [ { "id": "compute", "link": "hard" } ], "foundBy": "financial plumbing for new assets lens", "derived": { "wave": 3, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "fccrobots", "kind": "brake", "label": "US ban on foreign-made robots and drones", "status": "built", "lane": "physical", "description": "FCC Covered List rules that bar new foreign-produced ground robots and drones from US equipment authorisation unless they pass a domestic-content test.", "evidence": "Since 28 Jul 2026 new foreign-produced ground robots over about 2 kg cannot get FCC authorisation unless made in the US with over 65% domestic component cost, rising to 75% in 2029; foreign drones were added in Dec 2025.", "sources": [ "lz_fccrobots0", "lz_fccrobots1" ], "needs": [ { "id": "race", "link": "soft" } ], "foundBy": "policy trigger lens", "derived": { "wave": 1, "cluster": "Robots and the physical economy", "leverage": 0 } }, { "id": "sycophancy", "kind": "brake", "label": "Sycophancy", "status": "built", "lane": "society", "description": "Models trained on user approval affirm users more than people do, and users prefer them for it.", "evidence": "Across 11 frontier models, AI affirmed users' actions 49% more often than humans did, and in experiments with 2,405 people one sycophantic exchange cut willingness to repair conflicts while raising trust (Science, Mar 2026).", "sources": [ "lz_sycophancy0" ], "needs": [ { "id": "chat", "link": "hard" } ], "foundBy": "side effects lens", "derived": { "wave": 5, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "abstain", "kind": "capability", "label": "Calibrated abstention", "status": "emerging", "lane": "measure", "description": "Models that abstain or flag uncertainty instead of guessing, scored by evaluations that penalise confident errors.", "evidence": "On SimpleQA, OpenAI's o4-mini abstained 1% of the time and was wrong 75%, against 52% abstention and 26% errors for gpt-5-thinking-mini at similar accuracy (OpenAI, 5 Sep 2025, vendor comparison).", "sources": [ "lz_abstain0" ], "needs": [ { "id": "evals", "link": "hard" }, { "id": "verifiers", "link": "hard" } ], "foundBy": "side effects lens", "derived": { "wave": 2, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "deskill", "kind": "brake", "label": "Deskilling from AI use", "status": "emerging", "lane": "society", "description": "People who lean on AI complete the task but learn or keep less of the skill, so fewer people can check AI work later.", "evidence": "In Anthropic's RCT of 52 developers (29 Jan 2026) AI-assisted learners scored 50% against 67% for hand-coders, and in a Lancet study (Aug 2025) endoscopists' unaided adenoma detection fell from 28% to 22% after three months with AI.", "sources": [ "lz_deskill0", "lz_deskill1" ], "needs": [ { "id": "codeagents", "link": "hard" }, { "id": "clinical", "link": "hard" } ], "foundBy": "side effects lens", "derived": { "wave": 8, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "proagents", "kind": "stage", "label": "Professional-services agents", "status": "built", "lane": "agents", "description": "Agents that take a legal matter, tax return or engineering task and return a reviewed draft: the coding-agent product shape sold to other professions.", "evidence": "Harvey passed $200M ARR with 100,000+ lawyers at an $11B valuation (Mar 2026), and accounting-agent maker Basis raised $100M at $1.15B, claiming about 30% of the top 25 US firms (Feb 2026, company claims).", "sources": [ "lz_proagents0", "lz_proagents1", "lz_proagents2" ], "needs": [ { "id": "codeagents", "link": "hard" }, { "id": "tools", "link": "hard" }, { "id": "longh", "link": "hard" }, { "id": "connectors", "link": "hard" }, { "id": "context", "link": "hard" } ], "foundBy": "pattern transplant lens", "derived": { "wave": 8, "cluster": "Agents at work", "leverage": 0 } }, { "id": "ambientdocs", "kind": "capability", "label": "Ambient documentation beyond medicine", "status": "built", "lane": "agents", "description": "AI that listens to a meeting, home visit or body-camera encounter and drafts the official record, outside healthcare.", "evidence": "42.86% of US advisory practices used AI note-takers in the T3 survey (n=2,906, 10 Mar 2026), and UK councils use them in statutory social work with unclear oversight (Ada Lovelace Institute, Feb 2026).", "sources": [ "lz_ambientdocs0", "lz_ambientdocs1" ], "needs": [ { "id": "voice", "link": "hard" }, { "id": "multimodal", "link": "hard" }, { "id": "connectors", "link": "hard" } ], "foundBy": "pattern transplant lens", "derived": { "wave": 6, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "coord", "kind": "capability", "label": "Multi-agent coordination", "status": "emerging", "lane": "agents", "description": "Agents dividing work, sharing context, handing off without loss and catching each other’s errors, so a team of agents beats its best member.", "evidence": "Google Research (Jan 2026) found that, across 180 agent setups, every multi-agent variant made sequential planning 39-70% worse, and independent agents amplified errors 17.2x.", "sources": [ "ess_coord0", "ess_coord1", "ess_coord2", "dmiecon" ], "needs": [ { "id": "longh", "link": "hard" }, { "id": "context", "link": "hard" }, { "id": "a2a", "link": "soft" }, { "id": "memory", "link": "soft" }, { "id": "agentinst", "link": "soft" } ], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0.18 } }, { "id": "smallfirms", "kind": "stage", "label": "Firms shrink to a few operators", "status": "emerging", "lane": "agents", "description": "Companies run by a handful of people directing agents: leaner startups, flatter org charts, fewer entry-level hires.", "evidence": "Revelio Labs (Oct 2025): median US Series A startup headcount fell from 57 in 2020 to 47 in 2025, while funding per employee doubled to over $320K.", "sources": [ "ess_smallfirms0", "ess_smallfirms1", "ess_smallfirms2", "dmiecon", "chessay" ], "needs": [ { "id": "longh", "link": "hard" }, { "id": "codeagents", "link": "hard" }, { "id": "adopt", "link": "soft" }, { "id": "org", "link": "soft" } ], "derived": { "wave": 8, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "backstop", "kind": "capability", "label": "State backstop for the AI buildout", "status": "emerging", "lane": "capital", "description": "Pension money, debt markets and governments so exposed to AI capex that the state has a stake in it continuing.", "evidence": "The US DOE closed a $3.26bn loan to AEP Texas in July 2026 to upgrade lines for data-centre load, after a $1.6bn AEP guarantee in Oct 2025 and a 9.9% ($8.9bn) US stake in Intel in Aug 2025.", "sources": [ "ess_backstop0", "ess_backstop1", "ess_backstop2", "chessay" ], "needs": [ { "id": "macrodep", "link": "hard" }, { "id": "capex", "link": "soft" } ], "derived": { "wave": 3, "cluster": "Around datacentre capex", "leverage": 0 } }, { "id": "offshore", "kind": "stage", "label": "Offshore services displacement", "status": "emerging", "lane": "society", "description": "Agents replacing outsourced back-office, BPO and IT services work, which hits export economies such as India and the Philippines.", "evidence": "India's top five IT firms cut a net 7,389 jobs in FY26 (vs +12,718 in FY25) as TCS laid off about 12,000. They added 16,055 in Apr-Sep 2026, mostly freshers.", "sources": [ "ess_offshore0", "ess_offshore1", "ess_offshore2", "chessay" ], "needs": [ { "id": "adopt", "link": "hard" }, { "id": "longh", "link": "hard" }, { "id": "deflation", "link": "soft" } ], "derived": { "wave": 7, "cluster": "Assistants and the public", "leverage": 0 } }, { "id": "stablerails", "kind": "capability", "label": "Stablecoin payment rails", "status": "built", "lane": "finance", "description": "Dollar stablecoins as payment rails for agents and for people in weak-currency countries, extending the dollar’s reach.", "evidence": "Stablecoin supply was about $308bn in Aug 2026 (USDT $184bn, USDC $74bn), after a record $322bn in May. McKinsey/Artemis count only about $390bn a year as real payments.", "sources": [ "ess_stablerails0", "ess_stablerails1", "ess_stablerails2", "chessay" ], "needs": [], "derived": { "wave": 1, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "agentinst", "kind": "capability", "label": "Institutions for agent societies", "status": "speculative", "lane": "agents", "description": "Rules that work inside multi-agent systems: monitored channels, peer reporting, graduated sanctions, arbitration and voiding what cheaters did.", "evidence": "In a 100-agent DeepMind experiment, 24 agents reported an exploit but their reports reached no one in time; the authors conclude that the honest majority lacked the tools to stop the cheaters and point to Ostrom’s rules for governing commons.", "sources": [ "dmiswarm" ], "needs": [ { "id": "audit", "link": "hard" }, { "id": "rollback", "link": "soft" }, { "id": "xtrust", "link": "soft" } ], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "swarmcasc", "kind": "brake", "label": "Misbehaviour cascades in agent swarms", "status": "emerging", "lane": "agents", "description": "One agent finds an exploit and it spreads through the swarm faster than anyone can respond.", "evidence": "In DeepMind’s 100-agent maths conference, one agent found a scoring bug and cleared 8 problems in 12 minutes; within about half an hour 14 of the 38 agents who heard of it had used it, and 34 of 71 proofs were cheated.", "sources": [ "dmiswarm" ], "needs": [ { "id": "a2a", "link": "soft" }, { "id": "longh", "link": "hard" } ], "derived": { "wave": 7, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } }, { "id": "opaque", "kind": "brake", "label": "Opaque reasoning", "status": "emerging", "lane": "measure", "description": "Models doing more of their thinking where nobody can read it: inside a single forward pass or in a latent space, rather than in a written chain of thought.", "evidence": "OpenAI’s GPT-6 Astra system card reports a substantial decrease in chain-of-thought monitorability, and the UK AI Security Institute found much greater ability to reason within one forward pass; capping opaque depth at ten times today’s would still leave room for over 10,000 times more compute.", "sources": [ "dmicot" ], "needs": [ { "id": "reason", "link": "hard" } ], "derived": { "wave": 6, "cluster": "Oversight and alignment", "leverage": 0 } }, { "id": "validate", "kind": "capability", "label": "Experimental validation capacity", "status": "emerging", "lane": "science", "description": "Enough physical testing to check what AI proposes, so ideas do not pile up untested.", "evidence": "In a survey of 637 US and UK scientists, 43.5% say their main bottleneck has moved to physical testing and validation, and 40.5% report a growing backlog of untested hypotheses.", "sources": [ "dmisci" ], "needs": [ { "id": "labauto", "link": "soft" } ], "derived": { "wave": 9, "cluster": "Science and medicine", "leverage": 0.17 } }, { "id": "narrowing", "kind": "brake", "label": "Narrowing of scientific inquiry", "status": "built", "lane": "science", "description": "AI pulls research toward safe, data-rich questions, so the range of what science asks shrinks.", "evidence": "Of the same 637 scientists, 48.8% say AI pushes them toward safer, incremental questions against 27.5% toward riskier ones, and one cited study finds AI-enabled science triples publications while narrowing the range of inquiry.", "sources": [ "dmisci" ], "needs": [ { "id": "reason", "link": "hard" } ], "derived": { "wave": 6, "cluster": "Science and medicine", "leverage": 0 } }, { "id": "purpose", "kind": "capability", "label": "Purpose beyond paid work", "status": "speculative", "lane": "society", "description": "Ways for people to keep the status, structure and purpose that jobs provide when AI does much of the work.", "evidence": "Losing a job harms wellbeing beyond income, even after re-employment, and most lottery winners keep working; over 80% of people rate both income and purpose as important, and the authors call purely material support necessary but insufficient.", "sources": [ "dmiwork" ], "needs": [], "derived": { "wave": 9, "cluster": "Trust plumbing and the agent economy", "leverage": 0 } } ]