Clark Whitehead Research · the focus

Continual Learning

Can an AI keep learning new things without forgetting what it already knows? Does the lesson reach the weights?

Every page here is one line of that work, tested with games. The picture shows the idea; open it for the claim, the test and where it stands.

Why it matters

AI that could work for months on curing one disease, getting better at it the whole way instead of starting over each day. Today’s AI can’t.

The test: does an AI forget?

A game that tests whether an AI forgets

A battle game where each level can only be beaten with a trick you picked up playing an earlier level. If the AI forgets the trick, it gets stuck.

Fourteen battles. The enemy base sits behind a ring, and the only way in is a move an earlier battle made you learn: hold a hill, use a portal, trip a switch. Context wiped between battles. Who passes so far, and how to run your own agent.

Open the page 02

Does an AI that learns beat one that takes notes?

Two copies of the same AI play the same levels. One keeps a notebook of everything it saw and rereads it every move. The other trains on it. We add levels until one falls behind.

A proposed test with memory left on: lessons spread across hundreds of tries and timing skills, the kind a notebook carries badly. If the notebook player keeps up, notes are enough and the bet is wrong.

Open the page

Learning new things without losing the old

A third of forgetting is the arbitrary choice in the example

When a task has several right answers, training shows one of them. That pick turns out to be a third of what later training destroys. Teaching all the right answers at once removes it.

Measured in grid worlds with an exact solver, so every demonstration is provably optimal and only the tie-break differs: the opposite convention costs 0.10 retention in 24 of 24 cells; an all-answers skill is immune. 1,088 runs.

Open the page 04

Teaching an AI new things without losing the old

A plan: the AI learns from its own play, protects what it already knows while it updates, and replays earlier levels so they stay beatable.

Learn from own play, fence the weight updates, replay earlier levels. A recipe to be scored on the gated test.

Open the page 05

Protecting old knowledge is only half the job

A fence around old knowledge only decides where an update goes. Choosing what to protect, and when, has to come from memory and the learning rules.

A subspace fence only decides where an update goes. What to protect, when to lock in and what flows in come from elsewhere.

Open the page

Brain-computer interfaces

Teaching a computer to talk to someone’s brain

The computer sends a signal into the brain along with words on screen. The words slowly fade away, level by level, until the signal alone gets the message across.

A model learns to teach a person through a brain-computer interface: it sends a signal into the brain while the same message is on screen, and over a thousand levels the text fades until the signal alone is understood.

Open the page

Questions

Why do people learn so fast?

Big unanswered questions about how people learn so much from so little, and whether machines could too.

Why people learn so much from so little, whether the brain uses backpropagation, and what an agent should keep, revisit or forget.

Open the page

Earlier work

AI agents that have to survive in changing worlds

The original big idea: put AI agents in worlds where they must survive as things change, and see if they pick up skills that last.

The 2025 thesis: continual learning as a path toward superintelligence, survival pressure in changing multiagent worlds, and the research loop.

Open the page 09

Letting the AI decide how it learns

Ideas for an AI that chooses what to learn, what to pay attention to, and even how to change its own code.

Four earlier proposals: let the agent choose what to learn, a value function for attention, one agent across games, and editing its own model code.

Open the page 10

The AI players built so far, head to head

The players built before this test, and how they did when they played each other.

The combined learning loop, the MuZero and Lana architectures, the model tournament, the Lana benchmark and the 11-agent combat study.

Open the page

Papers Background & existing work Funding & contact