AGI Is a Learning Problem, Not a Storage Problem

A transcript that reframed the question

Recently a transcript of a four-hour investor call began circulating, reportedly featuring DeepSeek founder Liang Wenfeng. In it he names what today's AI lacks. It isn't creativity and it isn't reasoning. According to the transcript, the real shortfall is that AI can't keep learning the way people do, and a model shouldn't count as next generation until that changes.

The roadmap he reportedly laid out starts with chain-of-thought reasoning and moves to agents, then continual learning, then a "singularity point" and finally embodied intelligence.

That idea has an independent argument behind it, coming from a different direction.

Intelligence isn't a stockpile

Each new language model release triggers the same conversation: benchmark scores, reasoning ability, context length, parameter counts and how much smarter it feels than its predecessor. Intelligence gets measured almost entirely by what a model already knows. That may be the wrong yardstick.

People aren't intelligent because of how much their brains store. If storage were the measure, a medical textbook would beat a doctor.

Consider a person asked about a field they know nothing about. The answer doesn't materialize. First comes the realization that they don't know enough. Then they search, compare sources, read, question what they find, build understanding slowly and finally reach an answer. Weeks later, with no revisiting, many details have faded. Needing that knowledge again means repeating the process. The intelligence didn't vanish with the forgotten facts, because the ability to learn stayed intact.

A person is no AGI, and even human general intelligence forgets constantly. Skills unused for long enough fade and have to be recovered by retraining on the topic. That isn't a defect in intelligence. It's how learning has worked in every intelligent system observed, ourselves included.

How language models differ

Today's models work differently. A giant pretraining run comes first, followed by carefully planned fine-tuning stages, and from there the parameters stay essentially fixed until the next training cycle begins. If something new happens in the world, the model has no way to choose to study it. Engineers, researchers and data pipelines make that choice while the model sits idle. In effect, intelligence has been handled as a storage question when it's really a learning one.

Catastrophic forgetting

Machine learning has a stubborn obstacle called catastrophic forgetting. Fine-tune a model on fresh data and it frequently loses part of its earlier knowledge, because the very parameters supporting old abilities get modified to take in new ones. The cause is built into the design. A transformer keeps what it knows in the same fixed weight matrices it uses to reason, so touching one affects the other. Years of research have gone into dampening the effect, yet it remains open whether rewriting one shared representation over and over is a sensible strategy at all.

Not one founder's hunch

The concern isn't limited to one founder. Demis Hassabis, who runs DeepMind, said much the same in public this year. He listed consistency, long-horizon reliability and human-like adaptability, with continual learning inside that group, as what still blocks progress, and he didn't blame a shortage of compute. His position contrasts with Anthropic's, where Dario Amodei has spoken of something like a country full of brilliant researchers living in a data center, reached largely by extending present scaling methods.

Both claims can be right together. Scaling keeps paying off, and continual learning remains the one problem nobody has fully cracked. Together they show the idea isn't fringe. It runs along a genuine fault line in how the largest labs describe the work still ahead.

Borrowing from the brain

The human brain doesn't seem to rewrite itself every time it picks up one new fact. It has several interacting systems: working memory for immediate reasoning, long-term memories built over years, skills that become automatic through repetition and memories that fade when they stop being useful. Learning isn't one operation. It's an ecosystem of processes.

A handy analogy says the brain contains something like RAM and something like ROM, which are separate systems. Working memory plays the RAM role: fast, short-lived and overwritten all the time. Long-term knowledge plays the ROM role, slower to write but steady afterward. For AGI to resemble human intelligence, that split probably has to be designed in from the beginning and not left to one enormous set of weights covering both jobs. The brain needn't be the only template, but dividing reasoning from memory also addresses practical engineering issues.

A model that notices its own gaps

Imagine a reasoning model that continually notices where it's weak. Rather than wait months for a training run, it finds what it's unsure about, collects trustworthy information, verifies it, builds a tiny amount of fresh training data and proposes a small change to itself. Next it checks whether that change really helps without harming current abilities, then accepts or rejects it. Any accepted change stays small enough to roll back if later evidence exposes an error.

A mechanical version could go like this:

1. The model assembles a tiny dataset on its own, by researching a subject on the web or reflecting on where it already stumbles.
2. It fine-tunes a copy of itself on that small set. That's a small nudge to the weights and not a full retrain. The nudge will probably cost a bit of earlier performance somewhere, so a little forgetting is likely unavoidable, and the aim is to keep that cost low enough that overall performance doesn't decline.
3. The tuned copy is compared with the untouched original on held-out cases that the update never saw.
4. If the copy truly improves, it's kept and another small step follows, with a learning rate controlling the stride just as it does in training. If not, it's discarded and the cycle restarts.

The obvious danger

The danger is garbage in and garbage out. Training on self-generated data can lead a model to strengthen its own mistakes, and tiny errors can accumulate across many steps although each one looked harmless. One partial safeguard is letting the model audit its generated dataset before training, grading its own examples with a few-shot prompt and keeping only the good ones, rather than taking everything it produced. That falls short of a solution. It simply restates the self-evaluation problem on which the whole idea depends, in a smaller and more candid form.

What's new here and what isn't

The claim that continual learning is the missing piece is far from new. Liang stated it days ago, Hassabis has voiced versions of it in public and researchers have studied catastrophic forgetting for years. What's scarce in public is a concrete how. Liang's roadmap stops at one sentence saying a continuously learning model can build its own next version, which names the destination and leaves the mechanism blank.

The loop above, with a small dataset, a bounded fine-tune, validation against the frozen original and keep-or-discard, is one attempt at the gap. It may not be correct, and it may not even be original at the implementation level, since someone in a lab is almost certainly trying a version already. It's concrete enough to be wrong in a specific, checkable way, and an idea that can be proven wrong beats one too vague to test.

The result wouldn't be a model constantly rewriting its identity. It'd be a model gradually evolving through thousands of validated corrections.

Hard questions that follow

This approach immediately raises difficult questions:

• How does the system verify that its own learning is right?
• How does it avoid reinforcing its own misconceptions?
• Is forgetting a binary event, with a fact either present or gone, or is it a gradient that fades a little more with each unrelated update?
• How much knowledge should become permanent, and how much should stay external?
• When should information go into memory, and when should it just be retrieved from trusted sources on demand?

These are tougher than the original proposal, and there are no ready answers here. They're exactly the questions worth researching.

The pieces already exist

Plenty of ingredients already exist, including retrieval systems, continual learning methods, synthetic data generation, autonomous agents, parameter-efficient fine-tuning and increasingly strong reasoning models. What appears to be missing isn't another scaling law, but a framework that lets those ingredients work together inside a system that keeps learning over its whole lifespan.

Perhaps AGI won't arrive because a single model grows to an unimaginable size, and will instead arrive once AI stops being static. Intelligence could depend less on collecting parameters than on building systems with good judgment about when to doubt themselves, search for new information, revise beliefs and admit they aren't sure.

Whether this is the road to AGI is unknown, and history may well prove the intuition wrong. One idea still deserves to be taken seriously: the most intelligent systems people know, the human mind itself, aren't remarkable because they never forget. They're remarkable because they never stop learning.