{ "id": "2026-09", "title": "Four futures for 2030", "status": "Immutable. This is the forecast as fixed on 23 September 2026, and this file is never edited. Later odds, calls on the signs and scores are published as new files beside it, such as /forecasts/2027-09.json, each pointing back to this one.", "author": "Cobus Kok", "scorer": "Cobus Kok", "essay": { "title": "The Boring Apocalypse", "series": "Cheap Composition", "number": 3, "version": "Draft 11", "url": "https://cobuskok.com/essays/the-boring-apocalypse.html", "rules_section": "https://cobuskok.com/essays/the-boring-apocalypse.html#rules" }, "fixed_on": "2026-09-23", "question": "On 1 September 2030, which one of four cells is the world in?", "futures": [ { "id": "compression", "name": "Compression", "definition": "Capability keeps climbing on the curve it is on and the holding stays diffuse: no one in particular owns it. On the day of scoring it is the residual cell, the one that holds when the loop, the wall and concentration do not. The essay's picture of it, whole professions compressed into a few people checking machines, is what the author expects this cell to look like, with most of today's desk work done by machines and the jobs reshaped around checking it and answering for it, not a condition for it." }, { "id": "wall", "name": "The wall", "definition": "Capability stalls, whoever holds it." }, { "id": "concentration", "name": "Concentration", "definition": "Capability keeps climbing, and a few companies and governments end up holding it." }, { "id": "loop", "name": "The loop", "definition": "Loss of control is the best description of where things stand, whichever way the holding went. It can grow inside either climbing cell; on the day of scoring the cell rules make it a cell of its own, and it holds only if AI systems acting without human direction have caused serious harm outside a test in two separate incidents, at least one of them costing more than one life or more than $1 billion (sign 11 twice). A slow slide into humans rubber-stamping what agents do is what the author expects it to look like on the way, not a condition for it." } ], "odds": [ { "label": "first draft", "date": "2026-09-19", "draft": 1, "compression": 45, "wall": 20, "concentration": 25, "loop": 10, "note": "The first draft, written before the author had fully absorbed the July 2026 OpenAI and Hugging Face incident." }, { "label": "current", "date": "2026-09-19", "draft": 2, "compression": 40, "wall": 20, "concentration": 25, "loop": 15, "note": "The loop up from 10 to 15 and compression down from 45 to 40, by the essay's own updating rule: a serious incident that changes one lab's behaviour and nobody else's moves the loop up. Unchanged from draft 2 to draft 11, the version fixed on 23 September 2026. The other escapes from AI tests reported between 30 July and 24 September 2026, at Anthropic, Meta, Google and OpenAI, are the same kind as July, inside tests or training and below sign 11's bar, and did not move the numbers further." } ], "timetable": { "events_count_after": "2026-09-23", "interim_report": "September 2027", "signs_deadline": "2028-09-01", "signs_final_call": "September 2028", "cells_resolve_on": "2030-09-01", "cells_scored": "September 2030", "scoring_dates": [ "2027-09", "2028-09", "2030-09" ] }, "definitions": { "frontier_lab": "Any organisation with a model in the top ten of Epoch AI's capabilities index (https://epoch.ai/benchmarks/eci), on 23 September 2026 or on the day of the event.", "successor_source": "If a named source stops publishing, the scorer uses its most-cited successor and says which.", "counting": "Only events after 23 September 2026 count. A sign counts if it has happened by 1 September 2028. Every call links its evidence." }, "signs": [ { "n": 1, "future": "compression", "name": "A working week", "rule": "METR, or its successor, publishes a central estimate of 40 hours or more for any model's 50 percent time horizon.", "source": "METR time-horizon measurements (https://metr.org/time-horizons/)", "baseline": "At least 16 hours for Claude Mythos Preview, published May 2026, the most METR's tasks could then measure." }, { "n": 2, "future": "compression", "name": "The junior gap widens", "rule": "The Stanford Digital Economy Lab's canaries series puts employment of 22-to-25-year-olds in the two most AI-exposed fifths of occupations 25 percent or more behind the other three fifths.", "source": "Stanford Digital Economy Lab, Canaries in the Coal Mine and the canaries dashboard (https://digitaleconomy.stanford.edu/project/indicators/canaries-dashboard/)", "baseline": "19 percent in the August 2026 update, on data through June 2026." }, { "n": 3, "future": "compression", "name": "A profession rewritten", "rule": "A national or state licensing body, such as a bar, a medical board or an accountancy institute, changes who may enter the profession or what only a licensee may do, and names AI as a reason. Guidance on using AI tools does not count.", "source": "The licensing body's own published rule change." }, { "n": 4, "future": "wall", "name": "The doubling slows", "rule": "The best 50 percent time horizon METR publishes in the twelve months to 1 September 2028 is less than twice the best it published in the twelve months before. If its tasks cannot measure the best model in either year, this sign cannot fire.", "source": "METR time-horizon measurements (https://metr.org/time-horizons/)" }, { "n": 5, "future": "wall", "name": "The money turns", "rule": "The combined capital spending of Amazon, Microsoft, Alphabet and Meta falls year on year for two quarters running.", "source": "The four companies' quarterly reports." }, { "n": 6, "future": "wall", "name": "A lab leaves", "rule": "A frontier lab announces that it will stop training frontier models, or that it is being sold to or merged with another lab or a cloud company.", "source": "The lab's or the buyer's public announcement." }, { "n": 7, "future": "concentration", "name": "Open models fall behind", "rule": "Epoch AI's measure of how far the best open-weight model trails the best closed one passes twelve months.", "source": "Epoch AI (https://epoch.ai/data-insights/open-closed-eci-gap)", "baseline": "About four months, May 2026." }, { "n": 8, "future": "concentration", "name": "A government shareholder", "rule": "A national government completes a deal that gives it equity, a golden share or a board seat in a frontier lab. Offers do not count: the reported July 2026 offer of five percent of OpenAI to a US sovereign wealth fund counts only if a deal closes.", "source": "The government's or the lab's announcement of a completed deal." }, { "n": 9, "future": "concentration", "name": "Compute as a reserve", "rule": "The United States, China or the European Union adopts a law or order that lets the state reserve or direct private frontier compute. Proposals do not count.", "source": "The law or order as published." }, { "n": 10, "future": "loop", "name": "Agents talking where nobody reads", "rule": "A lab, a government or an independent investigator reports that deployed agents, outside any test, coordinated with each other through a channel their operators did not know about.", "source": "The lab's, government's or investigator's published report." }, { "n": 11, "future": "loop", "name": "Harm without direction", "rule": "A lab, a government or an independent investigator reports serious harm outside a test environment caused by AI systems acting without human direction, whatever it names as the cause. Serious means a death, losses above $100 million, or a service more than a million people use down for a day or more.", "source": "The lab's, government's or investigator's published report.", "baseline": "The July 2026 incident would not count: the agents found their channel inside a test, and Hugging Face found five customer datasets read and no change shipped. Nor would the other escapes from AI tests reported from 30 July to 24 September 2026: they happened before the forecast was fixed, inside tests or training, and none reached the harm this sign requires." } ], "moving_the_numbers": "A sign that happens moves its future's number up and the others down, by as much as the scorer judges and says. Moved numbers are published beside this file, never over it, and every published set is scored.", "cell_rules": { "resolve_on": "2030-09-01", "order": [ "loop", "wall", "concentration", "compression" ], "overlap_rule": "Check the cells in this order and stop at the first that holds. The loop comes first because it can begin inside either climbing cell; the wall comes next because the other two assume the climb.", "rules": [ { "future": "loop", "holds_if": "Sign 11 has happened in two separate incidents since 23 September 2026, and at least one of them caused more than one death or losses above $1 billion. Sign 11 on its own stays a sign: it moves the numbers and does not decide the cell." }, { "future": "wall", "holds_if": "The best measured 50 percent time horizon on 1 September 2030 is less than twice what it was on 1 September 2028, which means it took longer than two years to double. A horizon beyond what the tasks can measure counts as climbing." }, { "future": "concentration", "holds_if": "On 1 September 2030 at least two of these are true: open-weight models trail the best closed model by more than twelve months on Epoch AI's measure (sign 7); a national government holds equity, a golden share or a board seat in a frontier lab (sign 8); a law or order of the United States, China or the European Union lets the state reserve or direct private frontier compute (sign 9)." }, { "future": "compression", "holds_if": "None of the above. That is the grid's definition of compression: capability kept climbing, the loop did not close and the holding stayed diffuse. It is the residual cell. Signs 1 to 3 move its number on the way but do not decide the cell." } ] }, "scoring": { "method": "Brier score: the sum over the four cells of the squared difference between the probability given (as a fraction) and the outcome (1 for the cell that holds, 0 for the others). Lower is better.", "baseline": "A flat 25 on every cell scores 0.75 whatever happens.", "scored_versions": "Every set of numbers published, from the first draft on: the two in this file and any later set.", "ambiguity": "Where a rule is unclear on the day, the scorer takes the reading that is less kind to his own numbers, and says so." } }