Why We Need Continual Learning: From Amnesic Models to Real Accumulation
This piece draws on a16z's *Why We Need Continual Learning* and argues that today's large language models behave like Leonard Shelby in *Memento*: capable and resourceful, yet unable to truly accumulate experience because their weights are frozen after training. Every new fact has to live outside the model, as retrieval, prompt, or external note.
Key points
- The Filing Cabinet Fallacy. Bigger context windows, smarter retrieval, and tool layers produce impressive results, but they are not learning. Training compresses the internet into parameters, forcing the model to discard noise and discover structure. After deployment, we stop compressing and start storing. As Rich Sutton's *Bitter Lesson* suggests, scaling compute and generic learning tends to beat hand-crafted scaffolds.
- Ilya Sutskever's framing. Pretraining, on its own, has missed the point. Humans are not AGI; they have general skills but limited memorized knowledge, and they rely on continual learning. Deployment should include a process of trial, error, and adaptation, not a single frozen deliverable.
- What retrieval still cannot do. Some knowledge is tacit and resists verbalization: the texture intuition behind a radiologist's diagnosis, the subtle prosody of a specific speaker, the conceptual leap that lets Andrew Wiles prove Fermat's Last Theorem after seven years of isolated work. Context learning also fails on genuinely novel problems, adversarial adaptation, and any insight that requires inventing new connections rather than recombining existing text.
- A spectrum from context to weights. Continual learning is not a single method but a layered system. Context handles first-pass adaptation. Modular memory, such as compressed KV caches, adapters, or external stores, personalizes without touching the base model. Direct weight updates are the hardest and most essential layer, addressing discovery, adversarial adaptation, and tacit knowledge.
- Five directions for weight-level learning. Regularization and weight-space methods (EWC, weight interpolation) lock important parameters or blend old and new configurations. Test-time training (TTT layers, TTT-E2E, TTT-Discover) runs gradient descent during inference to compress fresh input into weights. Meta-learning (MAML, Nested Learning) trains the model to learn how to learn. Self-distillation (LoRD, SDFT) lets the model act as both teacher and student to constrain drift. Recursive self-improvement (STaR, AlphaEvolve) bootstraps reasoning from self-generated traces. The next generation of systems will combine these, not choose one.
- Google's Nested Learning. This work dissolves the boundary between architecture and optimizer. The model becomes a continuum of nested optimization problems with multiple timescales: fast weights reshape during inference, slow weights consolidate long-term memory, and higher-order modules monitor and tune the learning dynamics of lower-order modules. Some observers call it 'Attention Is All You Need — Part II.'
- Why not just fine-tune? Four engineering obstacles stand in the way. Catastrophic forgetting creates a stability–plasticity dilemma. Time-decoupling means constant rules and changing facts are entangled in the same weights, so updating one risks corrupting the other. Logical integration fails because new facts do not propagate their semantic consequences through the model. There is no differentiable 'subtraction' operation, so harmful or incorrect knowledge cannot be cleanly excised. Safety amplifies the problem: even narrow fine-tuning can degrade alignment unpredictably, weight-level updates create a persistent data-poisoning surface, and continual updates break static versioning, regression testing, and audit.
- 2025–2026 breakthroughs. TTT-E2E (Astera Institute, arXiv 2512.23675) showed 2.7x faster inference than full-attention Transformers on 128K contexts while loss continued to drop, demonstrating that compression beats brute storage. Nested Learning and HOPE introduced multi-timescale memory and self-referential learning-to-learn. SDFT (MIT, ETH Zurich) reached 89% strict accuracy on new knowledge acquisition and 98% out-of-distribution generalization while suppressing forgetting. Silver and Sutton's *Welcome to the Era of Experience* argued that models can now learn from their own actions, with AlphaProof generating hundreds of millions of self-play proofs after bootstrapping from 100,000 human examples.
- The Memento metaphor. Leonard's tragedy is not that he cannot function; it is that he cannot compound. Every experience stays external, written on a Polaroid, tattooed on skin, or scribbled by someone else. Today's AI faces the same ceiling. The path forward is layered: context for surface adaptation, modules for personalization, weights for true discovery. The end-state may require redefining 'model' itself, from a fixed bundle of weights to an evolving system that includes memory, update algorithms, and the ability to abstract from its own experience. That day has not arrived, but as of 2026 the engine has started.
Source
*Why We Need Continual Learning* — Malika Aubakirova, Matt Bornstein (a16z, 2026.04.22).