The Intelligence Flood

The water is what machines can do. For thirty years it rose behind a dam, one dated record at a time. In September 2026 it reached the top, and the valley below is what it has to cross next.

  • a dated record
  • where the water stands
  • soft rock, which wears
  • bedrock, which stays

The river: 4 narrows

Shortages of what it takes to run more machines.

    The dam: 19 dated records

    Select one to see the water on that date, with its sources.

      Below the dam: 3 barriers

      In this scenario each is soft rock on a sill of bedrock.

        What you are looking at

        The water collects behind a dam, and each line on the face of the dam is a dated record. Most are days on which a machine matched a trained person, a champion or a whole field at some kind of work. A few are steps in the machines themselves, such as a new design in 2017 and a public release in 2022. The water rises one line for each record.

        The river at the back is the supply of machines, which feeds the water. Since 2023 the trade press has reported shortages of chips, memory, fibre and power. Each is drawn as rocks that narrow the river from the year it was first reported.

        The top of the dam is September 2026. From that month the water runs through the dam and down the valley, where three barriers stand in its way.

        Up to November 2026 the picture is the record. What happens below the dam after that date is a scenario.

        How the heights are set

        The heights are not measured, because nobody can say that a bar exam is so many metres above chess. The water rises one step for each dated record, in the order they happened, so the height of a line shows only its place in that order.

        The dam is drawn as high as the water stood in September 2026. Where the top goes is our choice.

        Why the top of the dam is September 2026

        Three reports came between September and November 2026.

        We read the three together as the point at which AI starts to speed up the research that makes AI, so that each gain helps to make the next. We call what follows the Compounding Intelligence Flood.

        The record behind the dam

        WhenWhatWhat the record says
        May 1997ChessIBM's Deep Blue beats the world champion, Garry Kasparov, 3.5 to 2.5 over six games.IBM: Deep BlueComputer History Museum: Defeating the World Chess Champion
        December 2012Pictures and speechA deep neural network from the University of Toronto is presented at the NIPS conference. It had won that year's ImageNet contest for labelling photographs with 15.3% errors, against 26.2% for the runner-up.Krizhevsky, Sutskever and Hinton: ImageNet Classification with Deep Convolutional Neural NetworksNIPS 2012: conference dates
        February 2015Pictures and speechA Microsoft Research system labels a standard set of photographs with 4.94% errors. A trained person made 5.1%.He and others: Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet ClassificationRussakovsky and others: ImageNet Large Scale Visual Recognition Challenge (the human figure)
        October 2015GoAlphaGo beats the European champion, Fan Hui, a professional, by five games to nil.Google DeepMind: AlphaGo
        March 2016GoAlphaGo beats Lee Sedol, winner of 18 world titles, by four games to one in Seoul.Google DeepMind: AlphaGo
        October 2016Pictures and speechA Microsoft Research system writes down recorded telephone conversations with 5.9% errors, the same as professional transcribers.Xiong and others: Achieving Human Parity in Conversational Speech Recognition
        January 2017PokerLibratus, built at Carnegie Mellon University, beats four of the world's best professionals over 120,000 hands of two-player no-limit hold'em.Carnegie Mellon University: Carnegie Mellon Artificial Intelligence Beats Top Poker Pros
        June 2017A new designEight authors publish a paper proposing the Transformer, a network design built on attention alone. Six years later OpenAI describes GPT-4 as a Transformer-style model.Vaswani and others: Attention Is All You NeedOpenAI: GPT-4 Technical Report
        July 2019PokerPluribus beats leading professionals at the six-player game.Carnegie Mellon University: Carnegie Mellon and Facebook AI Beats Professionals in Six-Player Poker
        November 2020Protein structureAt CASP, the field's two-yearly assessment, DeepMind's AlphaFold predicts protein shapes with a median score of 92.4 out of 100. One of the organisers says the problem, fifty years old, is largely solved for single proteins.CASP14 press release (archived copy)Google DeepMind: AlphaFold: a solution to a 50-year-old grand challenge in biology
        November 2022Public releaseOpenAI releases ChatGPT to the public as a free research preview: a model that answers in conversation.OpenAI: ChatGPT: Optimizing Language Models for Dialogue (archived copy)
        March 2023Professional examsGPT-4 passes a simulated bar exam. OpenAI puts its score in the top 10% of test takers. A later study puts it at about the 48th percentile of those who pass, and the 15th on the essays.OpenAI: GPT-4 Technical ReportMartinez: Re-evaluating GPT-4's bar exam performance, Artificial Intelligence and Law
        May 2025DrivingA peer-reviewed study by Waymo of 56.7 million driverless miles finds significantly fewer injury crashes than human drivers in the same places. By June 2026, over 271 million miles, Waymo reports 95% fewer crashes with serious injury. The data are the company's own.Kusano and others: Comparison of Waymo Rider-Only crash rates by crash type to human benchmarks at 56.7 million miles, Traffic Injury PreventionWaymo: Safety Impact
        July 2025MathematicsGoogle DeepMind's Gemini scores 35 points out of 42 at the International Mathematical Olympiad, a gold-medal score, marked by the Olympiad's own coordinators.Google DeepMind: Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad
        September 2025Programming contestsAt the world finals of the university programming contest, Google DeepMind's Gemini solves 10 of the 12 problems inside the five-hour limit, which would have placed second. OpenAI's models solve all 12, one more than the winning team, in a test the organisers ran outside the championship's limits.ICPC: results of the 49th World FinalsGoogle DeepMind: Gemini achieves gold-medal level at the International Collegiate Programming Contest World FinalsICPC: OpenAI at the World Finals
        May 2026Software engineeringMETR estimates that the most capable model shared with it completes, half the time, software tasks that take a human expert 16 to 20 hours, and says its tests are not reliable above 16. It adds that this is not the same as doing a professional's day-to-day job.METR: Frontier Risk Report (February to March 2026)METR: Task-Completion Time Horizons of Frontier AI Models
        September 2026MathematicsOpenAI announces that about 10,000 AI agents, working for 88 hours under its mathematicians' direction, have proved that the Navier-Stokes equations can blow up. The proof is checked by machine. Quanta Magazine reports it as resolving a Millennium Prize Problem, if it holds up. This record is the top of the dam.Quanta Magazine: AI Has Solved One of Math's $1 Million Millennium Prize Problems
        October 2026MathematicsOpenAI releases 722 papers written by AI. One claims a proof of the quasi-Riemann hypothesis, a weaker form of the Riemann hypothesis. Quanta Magazine reports that experts who have worked through it largely agree it is correct.Quanta Magazine: AI Proved the Quasi-Riemann Hypothesis. What Happens Now?
        November 2026Software engineeringMETR estimates, with wide uncertainty, that Anthropic is improving its models at about twice the earlier pace, mostly by using AI in the research. It adds that the AI assists the researchers and does not yet do the work alone.METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026

        The river and its narrows

        The trade research firm TrendForce has reported each of the 10 inputs below as short, and on its latest reports none has cleared. The picture groups them into 4 narrows: chips, memory, fibre and power.

        InputFirst reported shortWhat the reports say
        Advanced packagingBy June 2023
        High-bandwidth memoryMarch 2024
        Enterprise solid-state drivesSpring 2024
        Server memoryAugust 2025
        Hard drivesSeptember 2025
        Optical fibreLate 2025
        Lasers for optical linksDecember 2025
        High-end capacitorsSpring 2026
        Leading-edge chipsApril 2026
        Grid power2026

        Below the dam: a scenario

        Below the dam the picture is a scenario without dates. It shows an order of events and two kinds of rock.

        In the order we have chosen, the water first meets the shortages that already narrow the river. Next it needs hands for physical work. In September 2026 the International Federation of Robotics said humanoid robots "are still on the verge of making the leap from laboratories to pilots". Nearly 7,000 were sold worldwide in 2025. International Federation of Robotics: World Robotics 2026, press conference

        Last comes the time that experiments and trials take. The United States drug regulator gives the usual length of a clinical trial as several months for the first phase, several months to two years for the second and one to four years for the third. US Food and Drug Administration: The Drug Development Process, Step 3: Clinical Research

        Each barrier is soft rock on a sill of bedrock. Soft rock is what building and knowledge can remove: a shortage ends when enough plant is built, and part of an experiment can be replaced by simulation. Bedrock is what stays: a plant still takes time to build, and a trial still runs its course.

        The picture follows three rules. Water that comes through the dam collects behind the first barrier it meets, and what spills over runs on to the next. As water passes over a barrier, its soft rock wears down towards the sill: the height that is left falls by the same fraction for each equal amount of water that passes. Bedrock does not wear.

        The second rule has the form of our own growth model. In that model the physical part of each step of progress shrinks as knowledge accumulates, down to a floor it cannot pass. Growth ends up paced by the slower of two physical clocks: how fast capacity can double, and how long the step that cannot be simulated takes. That is why the last barrier is drawn as mostly bedrock, with a deep pool that stays behind it.

        The sizes are not measured. We chose how high each barrier stands, how much of it is bedrock and how quickly it wears, so that the picture can be read. Our model of the 10 shortages works differently: building times are fixed in it, and a shortage lasts for as long as demand outruns new plant.

        The same picture for your organisation

        The same picture can be drawn for the work your organisation does. It would show which of that work a machine already does as well as your people, and which of your plans rest on soft rock and which on bedrock. The answers differ from one activity to the next, and they go out of date.

        Working that out is risk identification, which EON Risk Services does for organisations adopting AI.

        How a risk identification worksBrowse the AI risk register

        Sources are linked under each record and were read at source by 15 November 2026. Where a result is a company's own claim, the record says so. The choice of records, the place of the dam, the name on it and the three barriers are ours.