--- title: "The technical lever the bull case keeps stepping over" publication: deslop.media byline: deslop.media editorial provenance: editorial_synthesis provenance_note: "Editorial synthesis by deslop.media editorial, assembled from the subject's public record. The subject did not sit for an interview with deslop.media and has not reviewed it." issue: 1 issue_title: The Scaling Bet grade: draft issue_version: 3 issue_version_date: 2026-10-11 piece_version: 1 first_published: 2026-10-11 reporting_current_through: 2026-10 canonical_url: https://deslop.media/feature/richard-sutton/loss-of-plasticity/ cite_as: "\"The technical lever the bull case keeps stepping over,\" deslop.media editorial, Issue 1 (Draft v1, October 11, 2026)" changes: https://deslop.media/issues/1-the-scaling-bet/changes/ flag: "MCP tool flag_content, or POST https://api.deslop.media/api/flags" --- On August 21, 2024, Nature published "Loss of plasticity in deep continual learning" by Shibhansh Dohare, J. Fernando Hernandez-Garcia, Qingfeng Lan, Parash Rahman, A. Rupam Mahmood, and Richard Sutton. The empirical claim was narrow and unsentimental: standard backpropagation-trained deep neural networks degrade under continual training, eventually performing no better than shallow networks on the same data. The paper proposed "continual backpropagation" as a remedy — a small algorithmic modification that periodically reinitializes a fraction of low-utility units. Scaling's advocates have treated the result as an engineering problem. Standard practice handles continual training with periodic re-training, replay buffers, and large-context in-context learning. The Anthropic, OpenAI, and Google DeepMind frontier programs have not adopted continual backpropagation as a standard component of their training pipelines. Dwarkesh Patel's follow-up post on his interview with Sutton argued the strongest version of the bull case: test-time fine-tuning could replicate continual learning without the architecture-level change Sutton's program requires. For Sutton, the "loss of plasticity" result is the technical lever the rest of his program rests on. The "OaK architecture" — "Options and Knowledge," the model-based reinforcement-learning agent Sutton has reprised at RLC 2025 in August and at NeurIPS 2025 on December 3 — depends on stable continual deep learning that does not yet exist. The "FC-STOMP" cycle — Feature Construction, SubTask, Option, Model, Plan — is Sutton's five-step abstraction-creation loop, and it requires a network that can keep learning indefinitely without silently shedding capacity. Without continual backpropagation or something like it, the architectural frame around continual experiential reinforcement learning is missing a piece. ## What standard deep nets do under continual training The paper's experiments cover image classification (Continual ImageNet, Permuted MNIST, Slowly-Changing Regression) and reinforcement learning (Continual Control). In each case, the experimenters train a network on a sequence of tasks rather than a single fixed dataset. Standard backpropagation, applied in the conventional way, produces networks whose performance on new tasks degrades over time — silently, without an obvious failure signal. Many of the network's units stop responding to inputs; the effective network gets shallower as training proceeds. The paper reports this happening across multiple network sizes and across both vision and control tasks. Continual backpropagation, the proposed remedy, periodically reinitializes a small fraction of the units identified as having low utility. The empirical claim is that this single change preserves plasticity across the training sequence. The mechanism is simple; the claim is that it generalizes. Khurram Javed, Sutton's former PhD student and his Oak Lab co-founder, has extended this thread with "SwiftTD" (with Sharifnassab and Sutton; Reinforcement Learning Conference 2024) — a fast, robust temporal-difference algorithm consistent with the streaming-RL frame. Mohamed Elsayed, Gautham Vasan, and A. Rupam Mahmood's "Streaming Deep RL" paper (arXiv 2410.14606, October 2024) is a sibling Alberta thread on the same agenda. The "Reward Centering" paper (Naik, Wan, Tomar, Sutton; RLC 2024) sits on the same line. The code that accompanies the Nature paper is open source [on GitHub](https://github.com/shibhansh/loss-of-plasticity). ## What OaK depends on; what the LLM stack substitutes The OaK architecture is a complete model-based reinforcement-learning agent built around continual learning. The components map roughly to FC-STOMP: Feature Construction creates new features from experience; SubTask defines a goal-directed problem — a goal in John McCarthy's sense — the agent will solve; Option packages a learned policy for that subtask; Model captures the transition dynamics; Plan composes options into a course of action. The cycle is open-ended — the agent keeps building new features, defining new subtasks, learning new options, refining its model, planning further. The LLM stack does not have these components in the same shape. Pretraining produces a language model. Reinforcement-learning fine-tuning shapes its outputs against a reward model. Inference-time compute extends the model's chain of thought. None of these steps are continual in Sutton's sense; the model is not learning from environmental consequences during deployment. The model that ships on Tuesday is the same model that ships on Friday, plus or minus a fine-tune. Yann LeCun, FAIR's founding director and Meta's chief AI scientist until the end of 2025, and since then executive chairman of AMI Labs, runs a different bear position through his Joint Embedding Predictive Architecture program: scalar reward is insufficient, and JEPA-style non-generative latent prediction is needed to escape compounding generation errors. LeCun's NIPS 2016 cake metaphor — revised at ISSCC 2019 — runs that supervised learning is the icing and reinforcement learning is the cherry on top of a much larger self-supervised cake. "Reward is Enough" — the 2021 paper Sutton co-authored with David Silver, Satinder Singh, and Doina Precup, published in Artificial Intelligence — is the doctrinal counter to LeCun's frame. Sutton does not need additional architectural machinery beyond OaK-style options and knowledge primitives plus continual learning. LeCun thinks more architecture is required, including non-reward objective functions and joint-embedding predictive layers. No paper on Sutton's publications list responds directly to JEPA. The "Era of Experience" preprint with Silver functions as the sharpest published contrast: a future in which agents learn from environmental experience rather than from human-generated text, with reward as the unifying signal. ## What the bull case has not addressed No frontier-lab system has publicly adopted continual backpropagation as a production component. The standard answer at frontier labs has been that periodic re-training and replay buffers handle the practical case; that the loss-of-plasticity finding, while real, applies to a regime the frontier labs do not run in. Dwarkesh Patel's follow-up framed test-time fine-tuning as an LLM-native substitute. Sutton's answer, on the public record, is that test-time fine-tuning addresses a narrow band of the problem. The continual-learning agent has to learn from streams of experience, not from a fixed batch of demonstration pairs at inference. What decides between these two readings is whether continual-learning agents in the OaK frame demonstrate capabilities the LLM stack cannot match — and whether the LLM stack, scaled further, hits a continual-learning wall the loss-of-plasticity result predicts. The ARC-AGI-3 frontier scores below 1% at the benchmark's March 2026 launch were one possible early signal; in September 2026, ARC Prize reported GPT-6 Astra at 62.7% on the Semi-Private set under its Standard harness at max reasoning effort. METR's time horizon, the length of software task, timed by how long human experts take, that a model completes with 50% probability, also pointed to fast gains: METR Time Horizon 1.1 (January 2026) estimated a doubling time of 130.8 days for models from 2023 onward, against a 7-month long-run trend. On September 8, 2026, METR noted that the public series was no longer actively updated. Neither benchmark measures loss of plasticity. Both score a model on tasks; neither tracks how well a system keeps learning as its weights change again and again. That long deployments will run into the wall the paper describes is a hypothesis, not a finding. The test is learning performance measured across successive weight updates. ## Why one side calls it decisive and the other an engineering problem The paper's empirical claim has not been retracted, refuted, or substantively revised in the two years since. No frontier lab has reported adopting continual backpropagation at scale. The bull-case response is that frontier scale does not need it because the regime in which the loss-of-plasticity result was measured is not the regime in which frontier models are trained. Sutton's counter is that AGI-grade continual learning is exactly that regime, and the bull-case stack will arrive there eventually. The "Big World Hypothesis" — Javed and Sutton's framing — runs that the world is, and always will be, vastly larger than any agent. All useful knowledge must be approximated at runtime, not packed in at design time. The loss-of-plasticity result is the empirical evidence that the standard runtime-learning machinery does not yet support that approximation indefinitely; standard backprop trades initial capacity for permanent loss of capacity. Without a fix at the algorithmic level, agents that try to keep learning eventually stop being able to learn. The Anthropic, OpenAI, and DeepMind frontier programs run training pipelines that are, in practice, not continual: a base model is trained, fine-tuned, deployed, retired, replaced. The replacement is a fresh model, not a continuation of the previous one. From inside that practice, loss of plasticity reads as an academic finding about a regime that does not need to be entered. From inside Sutton's frame, the practice itself is the symptom — a field that retrains rather than learns continually has not yet built the kind of agent the "Bitter Lesson" eventually demands. One side says the wall in the paper is real but not on the path. The other side says the path leads to the wall. Long-horizon agents and models left in production for weeks will not settle it on their own, because their weights do not change while they run. A system that keeps updating its weights in deployment, measured update after update on how well it still learns, would. ## Sources - [Dohare, Hernandez-Garcia, Lan, Rahman, Mahmood, and Sutton, "Loss of plasticity in deep continual learning," Nature 632 (August 2024)](https://www.nature.com/articles/s41586-024-07711-7) - [Loss-of-plasticity codebase](https://github.com/shibhansh/loss-of-plasticity) - [Streaming Deep RL, Elsayed et al., arXiv 2410.14606](https://arxiv.org/abs/2410.14606) - ["OaK" at NeurIPS 2025](https://neurips.cc/virtual/2025/loc/san-diego/invited-talk/109601) - [Patel, "Some thoughts on the Sutton interview"](https://www.dwarkesh.com/p/thoughts-on-sutton) - [METR Time Horizon 1.1, January 29, 2026](https://metr.org/blog/2026-1-29-time-horizon-1-1/)