Toward AGIAdvancedLesson 565 min read

Grounding: finding out you were wrong

A model trained on text learns what people say happens. Grounding is connecting a prediction to a consequence — and it is the loop most AI systems still do not close.

🧪 Lab · 20 min📝 Lesson test
Lesson in motion

In 60 seconds

Grounding: finding out you were wrong

A model trained on text learns what people say happens. Grounding is connecting a prediction to a consequence — and it is the loop most AI systems still do not close.

1/6
In simple words
You can read a hundred books about swimming. You only find out whether you understood when you get in the water.
Text describes outcomes; it does not deliver them. A model can produce a confident plan and never discover that the plan fails, because nothing in its training loop ever ran it.
UNGROUNDEDPredict what text comes nextCompare with the text that didlearns what people writenever learns what happensGROUNDEDPredict an outcome, then actObserve what actually happenedlearns what workserror signal comes from reality
The difference is the return arrow. Grounding means reality gets a vote, and that vote is the only error signal that cannot be gamed by writing more plausible text.

Four ways to ground a system, cheapest first

  1. 1

    Executable environments

    Code that runs, tests that pass or fail, simulators. Cheap, fast, unlimited, and completely honest within their scope. This is why coding is where agents got good first.
  2. 2

    Formal verification

    Proof assistants, type systems, constraint solvers. A narrow domain with an absolutely reliable signal.
  3. 3

    Real-world tools with outcomes

    Did the API call succeed? Did the query return rows? Did the user accept the draft? Weaker signal, and it is real.
  4. 4

    Embodiment

    Robotics, sensors, actuators. The richest grounding and by far the most expensive — data is slow, hardware breaks, and mistakes are physical.
Do this
Notice that the first one is available to you today, for free. An agent that writes code, runs it, reads the failure and fixes it is a grounded system — a small one, in a narrow domain, with a real feedback loop. That is not a toy; it is the template.

World models: predicting consequences before acting

A step beyond reacting to outcomes is predicting them. A world model takes a state and a proposed action and predicts the next state, so the system can evaluate plans internally before committing to any of them.
Plan against a predicted worldpython
def plan(world_model, state, goal, depth=3, branching=5):
    """Search over predicted futures instead of acting immediately."""
    best, best_score = None, -1e9
    for action in propose_actions(state, goal, n=branching):
        predicted = world_model.step(state, action)      # predicted next state
        if depth > 1:
            sub, score = plan(world_model, predicted, goal, depth - 1, branching)
            score = score * 0.9                          # discount the future
        else:
            score = value(predicted, goal)
        if score > best_score:
            best, best_score = action, score
    return best, best_score

# The catch: the plan is only as good as the world model.
# A confident, wrong world model produces confident, wrong plans --
# and unlike a human, it does not feel any doubt about them.
Watch out
That comment is the crux. Planning inside a flawed world model is how you get elaborate, internally consistent plans that fail immediately on contact with reality. Always execute the first step and re-observe rather than committing to a long predicted sequence.

What grounding gives you for safety

  • An honest error signal. Reality cannot be talked around, which makes it the one evaluation that Module 34's problem does not apply to.
  • Calibration. A system that has been wrong in measurable ways can learn how often it is wrong.
  • Verifiable rewards. Everything in Module 51 depends on having a grounded checker.
Danger
And what it costs you. A grounded system is a system that acts in the world to learn. Every experiment is a real action with real consequences. This is precisely the sandboxing argument from Module 18, and it stops being optional: a learning agent must do its experimenting somewhere cheap.

Watch and read more

World models in AIYouTube searchGrounding and embodimentYouTube search

Lab

The same agent, grounded and ungrounded, on the same task.

~20 min

The problem

Give an agent a task in two modes: one where it can run code and see results, one where it can only reason. Measure success rate on both. Then find a task where grounding does not help and explain why.

You are done when

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1Your grounded agent is worse on one task. How is that possible?Reveal
The feedback was misleading rather than absent. A test that passes for the wrong reason, an API that returns 200 on failure, or a simulator whose physics differ from reality all give confident wrong signal, and the agent optimises against it — arriving somewhere worse than honest uncertainty would have. Grounding is only as good as the ground: a bad verifier is worse than none, because it converts uncertainty into misplaced confidence.

Please sign in to continue.

Questions people ask

Is multimodal training the same as grounding?

Not quite. Images and video give a much richer representation of the world, which helps a lot. Grounding in the strict sense means a feedback loop where the system's own action changes what it observes next.

Do I need robots?

No. Code execution, test suites and simulators provide genuine grounding, and they are free. Robotics gives the richest signal and the worst data economics.

Why is coding the strongest agent domain?

Because it is the one domain with a perfect, free, instant, unlimited verifier: run it. That single property explains most of the capability gap between coding agents and everything else.

Can a world model be learned from text?

Partially, and unevenly. Text carries a great deal of causal information about the world, and it also carries confident nonsense. Learned world models tend to be excellent on typical cases and unreliable exactly where it matters — the unusual ones.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.

Sign in with your phone number to take the test and save your progress