Build an AGIBeginnerLesson 617 min read

Three tiers of digital cognition

A model that talks, a model that checks itself, and a system that acts on its own. Most confusion about AGI comes from mixing these three up.

🧪 Lab · 15 min 📝 Lesson test

Lesson in motion

In 60 seconds

Three tiers of digital cognition

A model that talks, a model that checks itself, and a system that acts on its own. Most confusion about AGI comes from mixing these three up.

1/5

In simple words

One friend is great at finishing your sentences. A second friend checks whether what was said is actually true. A third friend goes out, does the job, comes back and tells you it is done. They are not the same friend.

Before any architecture makes sense, separate the three tiers. They are built on top of each other, and almost every argument about "is this AGI?" is really an argument about which tier someone is looking at.

Tap any box in the diagram

Tier 1 — the LLM

The memory and the vocabulary. It predicts the next token from statistical patterns absorbed in training. It knows which words sit well together; it has no mechanism at all for knowing whether they are true. Everything in Track A applies here. On its own it is a brilliant, confident, unreliable narrator.

Click each tier. The jump from 1 to 2 buys you reliability. The jump from 2 to 3 buys you autonomy — and autonomy is the one that has consequences.

A word about the vocabulary

Note

LLM is standard vocabulary everywhere. LLLM — "Large Language Logic Model" — is a useful framing for the verify-before-you-speak pattern, and it is not yet standard industry terminology; the same architecture appears in papers and product docs under names like tool-augmented reasoning, verifier-in-the-loop, program-aided language models and reasoning models (Module 51). We use LLLM here because it names the tier cleanly. Say "verifier-in-the-loop" when you are talking to someone outside this course, and you will be understood immediately.

What each tier can and cannot do

	LLM	LLLM	AGI
Produces fluent text	Yes	Yes	Yes
Knows if it is right	No	Sometimes — where a checker exists	Sometimes
Recovers from its own mistakes	No	Yes, within one task	Yes, across many tasks
Takes actions in the world	No	Only inside its checker	Yes — this is the point
Keeps a goal across many steps	No	No	Yes
Remembers yesterday	No	No	Only if you build the memory
Can hurt you	Only by being wrong	Only by being wrong	By doing the wrong thing

Danger

Read the last row slowly. Tiers 1 and 2 fail by producing a bad sentence. Tier 3 fails by taking a bad action. That is not a difference of degree — it is the entire reason Track B exists, and it is why the sandbox in Module 66 is not an optional extra.

Do this

The practical takeaway for a builder: most problems people try to solve with a bigger model are solved by moving up a tier instead. Wrapping a small model in a real verifier beats a huge model guessing alone, on nearly every task where correctness can be checked.

Watch and read more

**Deep dive into LLMs like ChatGPT**Andrej Karpathy · video

Tool-augmented and verifier-in-the-loop reasoningYouTube search

Lab

The same task at all three tiers, measured.

~15 min

The problem

Pick a task with a checkable answer. Solve it three ways: raw model (tier 1), model plus verifier loop (tier 2), autonomous agent with tools and memory (tier 3). Measure accuracy, cost and latency for each.

You are done when

Three implementations, same task, same eval
Tier 2 beat tier 1 on accuracy
You can say when tier 3 was NOT worth it

Hard questions

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

Q1Tier 2 is more accurate and cheaper than tier 3 on your task. When is tier 3 ever right?Reveal

When the task cannot be stated as a single verifiable step — when the sub-goals are discovered by attempting the work, when it spans many tools over a long horizon, or when the input arrives continuously rather than as one request. If you can write the whole task down in advance, tier 2 will usually win on every axis. Autonomy is a cost you pay to handle problems you could not fully specify.

Please sign in to continue.

Questions people ask

Is an LLLM a different kind of model?

No — and this is the most important thing to be clear about. It is the same model with machinery around it. Nobody trains an "LLLM". You build one, out of an ordinary model plus a checker plus a loop. The intelligence you gain lives in your architecture, not in the weights.

So is a reasoning model an LLLM?

Close. A reasoning model (Module 51) has the verify-and-retry behaviour trained into the weights, so it does more of this internally. An LLLM does it externally, in code you wrote and can inspect. External is slower and far more auditable — which is often the trade you want.

Do I need tier 3 for my product?

Usually not, and that is good news. Most valuable products are tier 2: a model that drafts and a checker that verifies, with a human pressing the button. Reach for tier 3 only when the task genuinely needs many unsupervised steps.

Where does today's frontier sit?

Solidly at tier 3 for short horizons — minutes to hours — with tier 2 reliability inside each step. What is missing is not the loop; it is memory that accumulates and reliability that holds over days. Modules 30 and 54.

Lesson test

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