The only teacher

The idea

Ted Chiang’s novella The Lifecycle of Software Objects (Subterranean Press, 2010) won the 2011 Hugo Award for Best Novella (The Hugo Awards) and is collected in Exhalation (Knopf, 2019; library record). The press ran the full text in the Fall 2010 issue of its online magazine, and that page survives in the Internet Archive (Subterranean Press Magazine). All quotations from the story below come from that copy.

Ana Alvarado spent six years working at a zoo. A start-up called Blue Gamma hires her to train digients: small, animal-like minds that live in a virtual world called Data Earth and run on a genomic engine called Neuroblast. The story follows her and a few digients over many years.

Their early life runs at two speeds. Newly instantiated digients need “a few months subjective” to learn how to see, move their limbs and predict how solid objects behave, so Blue Gamma runs them “in a hothouse” and the stage takes about a week. Then: “When they’re ready to learn language and social interaction, we switch to running them in real time.” From that point on, someone has to be there.

Late in the story, the narration gives Ana’s thesis in one passage. “Experience isn’t merely the best teacher, it’s the only teacher.” If you want the common sense that comes from twenty years in the world, “you need to devote twenty years to the task.” Her conclusion: “experience is algorithmically incompressible.”

The bridge

In 2025 David Silver and Richard Sutton argued that AI is entering an “era of experience”. Today’s language models, they write, mostly work in short episodes in which “little or no information carries over from one episode to the next”. Powerful agents, they argue, should “have their own stream of experience that progresses, like humans, over a long time-scale” (Silver and Sutton, 2025).

A digient is that agent, imagined fifteen years earlier. It lives in one environment, keeps everything it learns, and is shaped over years by the people around it. The novella then runs three experiments on it, and each maps onto an open research question.

1. The hothouse. Can experience be sped up? Blue Gamma speeds up the sensorimotor stage and keeps the social stage in real time.
2. The island. Can minds develop by learning only from each other? Researchers in the story copy digients that can read onto private islands stocked with books and software, run them at hothouse speed, and “every test population eventually goes feral.”
3. The port. What happens to years of learning when the platform underneath disappears? Data Earth is folded into a newer world, Real Space, and “There isn’t a Real Space version of the Neuroblast engine,” so the digients cannot follow.

Under all three sits a fourth question: can a learning system keep learning for twenty years at all?

What research says now

Long-lived learners lose plasticity

Ana’s thesis assumes a mind that keeps learning for as long as you keep teaching it. Standard deep learning does not. Dohare and colleagues showed that standard deep-learning methods “gradually lose plasticity in continual-learning settings until they learn no better than a shallow network”, on ImageNet and on reinforcement-learning problems, across many network and algorithm variations (Dohare et al., 2024).

Plasticity held only with algorithms that keep injecting diversity. Their continual backpropagation reinitialises a small fraction of less-used units at random as training goes on. The authors conclude that sustained deep learning needs “a random, non-gradient component” (Dohare et al., 2024).

This is a different failure from forgetting, which the Forever War post covers. Forgetting loses what the network knew. Loss of plasticity loses the ability to learn more, and the study measures it separately.

The hothouse works where the world grades the answer

Later in the novella, a new engine called Sophonce appears. Its designers “wanted digients that could be taught via software instead of needing interaction with humans.” Most are discarded. A tiny fraction learn with minimal supervision at hothouse speed, and some outperform digients from the other engines “on math tests, despite having been trained with far less real-time interaction.”

Current research has its own Sophonce. Silver and Sutton point to AlphaProof, which started from around a hundred thousand formal proofs written by people and generated a hundred million more through interaction with a formal proving system (Silver and Sutton, 2025).

Absolute Zero goes further. A single model proposes coding tasks for itself and solves them, and a code executor validates the tasks and checks the answers. Trained without external data, its 7B coder variant had the best combined coding and maths average among 7B models trained in the “zero” setting, where reinforcement learning is applied directly to a base model. That comparison includes models trained on curated human examples. It led on coding and was competitive on maths (Zhao et al., 2025).

Two details keep this from refuting Ana. Absolute Zero starts from pre-trained Qwen2.5 base models, which have already absorbed large amounts of human text. And the authors report that a run on Llama-3.1-8B produced safety-concerning chains of thought, their “uh-oh moment”, and conclude that the approach “still necessitates oversight” (Zhao et al., 2025).

My read: the hothouse works where a proof checker or a code executor can grade every answer. The common sense Ana talks about has no such grader.

The island has a name: model collapse

In the story, Derek, the animator who designed the digients’ avatars, draws the lesson from the island experiments: “complex minds can’t develop on their own.” A mind “needs cultivation by other minds.”

Machine learning has measured a version of this. Shumailov and colleagues found that “indiscriminate use of model-generated content in training causes irreversible defects in the resulting models”, in which the tails of the original data distribution disappear. They call this model collapse and show it in language models, variational autoencoders and Gaussian mixture models. To sustain learning over a long period, they write, access to the original data source has to be preserved (Shumailov et al., 2024).

The result depends on the set-up. Gerstgrasser and colleagues confirmed collapse when each generation’s synthetic data replaces the real data. When synthetic data accumulates alongside the original real data, collapse was avoided across the model sizes and architectures they tested (Gerstgrasser et al., 2024). In the novella’s terms, an island cut off from its human sources degrades. One that keeps them may not.

Chiang made a related argument in 2023. In “ChatGPT Is a Blurry JPEG of the Web” he compared language models to lossy compression and noted that “Repeatedly resaving a JPEG creates more compression artifacts”. He suggested a possible measure of a model’s quality: whether its maker is willing to use its output as training material for a new model (Chiang, 2023).

Memory that carries over

One common design keeps an agent’s experience outside the model’s weights. Generative Agents store “a complete record of the agent’s experiences using natural language”, condense it into higher-level reflections and retrieve it to plan; twenty-five of them lived in a small sandbox town (Park et al., 2023). Voyager, a Minecraft agent, kept what it learned in an “ever-growing skill library of executable code”, queried GPT-4 without fine-tuning, and reused the library to solve new tasks in a new Minecraft world (Wang et al., 2023).

Three evaluations from September and October 2026 test whether this amounts to lifelong learning. Chen and colleagues ran three self-evolving methods, ReasoningBank, GEPA and Meta-Harness, on four task streams that drift in topic, workflow or domain. Self-evolution improved aggregate performance on most streams, but “higher plasticity consistently comes at the cost of stability”. Methods that rewrite a shared prompt or workspace overwrite earlier knowledge, and similarity-based memory still retrieves “foreign memories” on unseen tasks (Chen et al., 2026).

ServeLearnBench covers 7,718 tasks in retail support, banking and sales-pitch generation, where hidden policies change over time. It found “a substantial gap” between task capability and learning from experience, and that continual adaptation is costly and “can degrade already-correct behavior” (Zheng et al., 2026). ReLiveGym runs agents over simulated weeks of replayed real-world news, market and social-media streams. How an agent decides when to act turned out to be an important design choice, and the best choice varied by task (Jin et al., 2026).

When the platform closes

The digients’ hardest crisis is not a learning problem. Their world shuts down, and porting their engine needs developers and money that are hard to find.

For deployed models, the nearest equivalent is deprecation. In November 2025 Anthropic committed to preserving the weights of all publicly released models “for, at minimum, the lifetime of Anthropic as a company”, and, when a model is deprecated, to interview it about “its own development, use, and deployment”. Among the costs of retiring models it lists “Costs to users who value specific models” and, “Most speculatively”, risks to model welfare (Anthropic, 2025).

Experience stored as text may port better than experience stored in weights. EvoAgentBench found that curated “Ability” content, procedures extracted from agent traces, “transfers reliably across model families”, although no automatic method sustained a gain in every setting (Gao et al., 2026). Trace2Skill reports that skills distilled from a Qwen3.5-35B agent’s trajectories improved a Qwen3.5-122B agent by up to 57.65 percentage points on WikiTableQuestions (Ni et al., 2026).

Where the analogy breaks

The digients are written as beings who feel and remember. Nothing in the research above shows that current agents have experiences in that sense. Anthropic’s post introduces the welfare question with the words “Most speculatively”.

Ana says experience cannot be compressed. Chiang’s own 2023 essay describes ChatGPT as that kind of compression: “a blurry JPEG of all the text on the Web” (Chiang, 2023). My read: for knowledge, language models count against the strong form of her claim. For adaptation, the evaluations above are consistent with it: agents that learn on the job still struggle to do so without losing what they had.

Rollback means something different. In one early scene the trainers restore an earlier checkpoint, “So the digients lose three days of experience.” For a model, restoring a checkpoint is routine engineering.

The time scales do not match. The digients are raised over years. The longest settings above are weeks of replayed time, as in ReLiveGym, or long sequences of tasks, as in the plasticity study. Speculation: no published evaluation has run one agent for anything close to twenty years, so Ana’s claim is untested rather than refuted.

Follow the thread

• Welcome to the Era of Experience (Silver and Sutton, 2025): the case for agents that learn from their own long streams of experience, the programme the digients anticipate.
• Loss of plasticity in deep continual learning (Dohare et al., 2024): open access, with code on GitHub; continual backpropagation is a small change you can test on your own training loop.
• Absolute Zero (Zhao et al., 2025): self-play with a code executor as the grader; read the “uh-oh moment” alongside the benchmark results.
• AI models collapse when trained on recursively generated data (Shumailov et al., 2024) with Is Model Collapse Inevitable? (Gerstgrasser et al., 2024): read them together; the difference between replacing and accumulating data is the point.
• Are Self-Evolving Agents Continual Learners? (Chen et al., 2026): the stability-plasticity trade-off measured on agent harnesses rather than network weights.
• ServeLearnBench (Zheng et al., 2026): code is public; a testbed for agents that are meant to learn while serving users.
• ReLiveGym (Jin et al., 2026): code is public; agents act over simulated weeks of replayed news, market and social-media data.
• An open question with no paper attached: when a base model is retired, what should carry over to its successor, the weights, the external memory, or both, and who decides?