Toward AGIAdvancedLesson 545 min read

What a bigger LLM will not give you

You have now built one. So here is the engineering question honestly: which missing pieces are scale problems, and which are architecture problems?

🧪 Lab · 15 min📝 Lesson test
Lesson in motion

In 60 seconds

What a bigger LLM will not give you

You have now built one. So here is the engineering question honestly: which missing pieces are scale problems, and which are architecture problems?

1/7
In simple words
Making the brain bigger helps with some things and does nothing for others. A person with a perfect memory of every book still cannot ride a bicycle they have never touched.
Module 30 listed the gaps from the outside. Now you know what is inside, so we can be precise about where in the stack each gap lives — and that tells you whether more compute fixes it.
GapWhere it livesDoes scale fix it?
Continual learningWeights are frozen after trainingNo. It is a training-procedure problem, not a capacity one
GroundingTrained on descriptions, never on consequencesPartly — multimodal and embodied data help, slowly
CalibrationSoftmax over next tokens has no notion of "I do not know"Improves with scale, does not resolve
Long-horizon reliabilityErrors compound across a sampling loopNo. It is a systems problem — checkpoints and verification
Causal modelsLearned from correlation in textContested. Some argue it emerges; nobody has shown it robustly
Persistent goalsNo state between forward passesNot a capacity issue at all — it is an architecture choice
Do this
Read the right-hand column carefully. Four of the six are not compute problems. They are things you build around the model, which means they are engineering decisions you get to make — and which carry the safety consequences from Track B.

The two schools, stated fairly

Scale is close to enough
  • Every "fundamental" barrier so far has fallen to more compute and better data.
  • Reasoning training already produced strategies nobody demonstrated.
  • Tools and memory patch the remaining gaps well enough in practice.
  • The burden of proof is on those claiming a hard wall, given the track record.
Something structural is missing
  • Frozen weights cannot learn from experience, at any size.
  • A next-token sampler has no mechanism for persistent goals or planning over days.
  • Text is a shadow of the world, not the world.
  • Improvements on benchmarks have not translated into reliability on messy real tasks.
Nobody can currently settle this, and you should be suspicious of anyone who says otherwise. What you can do is notice that both camps agree on the immediate engineering agenda: memory, grounding, verification, and control.

The build view of the remaining track

  1. 1

    Module 55 · memory

    Make the system accumulate and reuse experience without retraining.
  2. 2

    Module 56 · grounding

    Connect predictions to consequences, so the system finds out when it was wrong.
  3. 3

    Module 57 · self-improvement

    Close the loop so the system improves its own performance — and understand exactly why that loop is dangerous.
  4. 4

    Module 58 · the reference stack

    Everything assembled, with the control points marked.
Danger
One thing to keep in view through all four. Every gap you close makes the system more capable and more autonomous. Memory makes attacks persist. Grounding gives it real-world effect. Self-improvement removes the human from the loop. The capability work and the risk work are the same work, viewed from two sides.

Watch and read more

What is missing from current AIYouTube searchRob Miles on AI safetyRobert Miles · channel

Lab

Each of the six gaps classified as scale or architecture, with evidence.

~15 min

The problem

For each gap in Module 54, find one published result or run one experiment that indicates whether more compute closes it. Mark each scale or architecture and say what would change your mind.

You are done when

Hard questions

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

Q1Someone says continual learning is a scale problem because a big enough context is equivalent to learning. Rebut precisely.Reveal
Context is re-read from scratch every forward pass and discarded; weights are unchanged, so nothing is learned — the system is re-reading notes, not getting better. Practically: cost grows with every accumulated experience, retrieval quality falls as the store grows, and nothing generalises — a lesson learned in one context does not transfer to a related one the way weight updates do. Big context is a very good prosthesis for memory. It is not learning.

Please sign in to continue.

Questions people ask

Is this just my opinion or the field's view?

The gap list is broadly agreed. Whether scale closes them is genuinely contested among serious researchers, which is why it is presented as two columns rather than an answer.

If four gaps are engineering, can a small team make progress?

On memory architectures, verification loops and evaluation, yes — meaningfully, on modest hardware. Those are systems problems, and systems problems reward careful engineering more than compute.

Would solving all six give us AGI?

Unknown. It would give something markedly more capable and considerably harder to oversee. Whether it would be "general" depends on which definition from Module 26 you chose.

Why does this track exist in a security guide?

Because the architectures in the next four modules are being built now, and each one changes the threat model. Understanding them is how you avoid securing last year's system.

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