Contents

The series, in order.

Six essays in five tiers — the lens first, then the bets it places, an existence proof, and the capstone. Read it front to back, or pull whatever thread you like.

IThe LensHow to bet on papers: definitions outlive methods.
  1. 1
    The Selection Lens: How to Bet on Papers

    Most reading lists optimize for recency. This one optimizes for survivorship — a lens you can backtest against the canon before you trust it with the future.

    EVFN· 18 min· Stanley & Miikkulainen (2002) · Finn et al. (2017) · Kirkpatrick et al. (2017) · Brown et al. (2020) · Kaplan et al. (2020) · Radford et al. (2021) · Jiang et al. (2022) · Kunin et al. (2026)
IIThe Ten-Year TierSix paradigm bets for 2036, each with its falsifier.
  1. 2
    Paradigm Bets: The Ten-Year Tier

    Six bets from the 2026 collection that look expensive today and inevitable in 2036 — each defended by the frame it sets, the scaling curve it opens, and the one observation that would prove it wrong.

    FNME· 22 min· Bruce et al. (2024) · Genie · Chang et al. (2026) · Neural Computers · Maes et al. (2026) · LeWorldModel · Behrouz et al. (2025) · Titans · Hafner et al. (2025) · Dreamer 4 · Jiang et al. (2022) · Exploration · RLM · ES (published collection)
IIIFive-Year ProgrammesOpen scaling curves still being climbed.
  1. 3
    Recursion: The Third Scaling Axis

    After depth and width gave us parameters, and the internet gave us data, a third axis arrived almost unannounced: recursion — running the same computation again over its own evolving output. In 2025–26 it showed up independently at four different scales within months of each other. That convergence is the signal.

    FNMA· 23 min
  2. 4
    On-Policy Distillation Quietly Ate Post-Training

    While RLVR took the headlines, eight papers in roughly six months made on-policy distillation the quiet workhorse of post-training — student-generated trajectories under dense teacher supervision, sitting between SFT's distribution shift and outcome-RL's sparse credit.

    ARFN· 27 min· Zhou et al. (2026) · Li et al. (2026) · Luo et al. (2026) · Yang et al. (2026) · Lu et al. (2026)
IVThe Existence ProofA capability that did not exist, demonstrated and verified.
  1. 5
    When AI Did Mathematics

    In 2026 an AI system produced a piece of mathematics that professional mathematicians checked, accepted, and had not already known. This is the existence proof — stripped of both the hype and the dismissal that surrounded it.

    AGFNEV· 22 min· Alon, Bloom, Gowers, Litt et al. (2026) · Tsoukalas et al. (2026) · Kung et al. (2026) · Wiemann et al. (2026)
VThe CapstoneWhere the bets land in a system you can run today.
  1. 6
    The Continual Agent

    Deployed agents are amnesic, and fine-tuning is the wrong first answer. The minimal continual substrate is episode capture, replay, and rule distillation in system space — and that is not a stopgap awaiting real learning. It is the auditable half of a two-substrate architecture production agents will keep.

    CLAG· 27 min· cl-agent · Goswami (2026) · OpenClaw-RL · Wang et al. (2026) · ACuRL · Xue et al. (2026) · CL-Bench · Asawa et al. (2026) · Continual Backprop · Dohare et al. (2021) · Tracking · Sutton, Koop & Silver (2007)