![[AI learning.opus]]

My Apprehension About Learning in the Age of AI

AI is not inherently bad. My concern is that it makes learning feel easier than it actually is.

Learning Is Not Consumption

Learning involves more than receiving information:
Consumption → Processing → Retention

AI makes consumption extremely efficient, but that efficiency can create the illusion that processing and retention happened automatically.

• Understanding something in the moment is not same as learning it.
• Recognizing a concept is not same as being able to explain or retrieve it.
• Familiarity can make an idea feel true before it has been examined.

Why Reading Feels Different

Reading creates useful friction:

• Pause at a sentence.
• Question an assumption.
• Open another source.
• Follow a reference.
• Compare explanations.
• Take notes.
• Return to the idea later.

Video and audio often move at someone else's pace, making passive consumption easier.
That friction may be part of learning rather than an obstacle to it.

The Illusion of Progress

AI can instantly produce:

• Explanations
• Summaries
• Notes
• Diagrams
• Articles
• Presentations
• Code
• Learning paths

Because the output looks complete, it is easy to assume that progress has been made.
But:

The completeness of the output can hide the incompleteness of our understanding.

The same happens with AI-generated code. Thousands of lines can contain subtle bugs, incorrect assumptions, architectural problems, or missing edge cases.

Volume is not evidence of correctness.

AI removes friction from creation. It does not remove the need for judgment.

The Hardest Errors to Detect

Obvious hallucinations are relatively easy to catch.

More dangerous are small errors inside otherwise convincing explanations:

• A slightly incorrect definition

• A reversed causal relationship

• A fabricated historical detail

• An incorrect technical assumption

The problem becomes worse when learning something completely new.

The less I know about a subject, the more dependent I am on the information being correct—and the less capable I am of detecting when it isn't.

The problem is therefore not merely that AI makes mistakes.

It can make mistakes where the learner has no internal model capable of detecting them.

The Information Compression Problem

Knowledge is constantly compressed:

Reality → Experience → Understanding → Explanation → Summary → Consumer

Every transformation can remove information.

Compression itself is necessary. We cannot store reality exactly as it is.

The problem is deciding what can safely be removed.

Compression Depends on Source Quality

A rich understanding can often be compressed without destroying what matters.

But:

If the original understanding is weak, better compression cannot recover what was never there.

AI adds another layer of compression and transformation:

• Books → summaries

• Lectures → explanations

• Research → simplified answers

• Complex ideas → diagrams

• Multiple perspectives → one coherent narrative

Shorter does not mean equivalent.

Invisible Loss

A 500-page book compressed into 10 pages obviously signals that something was removed.

AI can make the compressed version feel complete.

We may not know:

• What was omitted.

• Which assumptions were made.

• Which perspectives disappeared.

• Whether the original source was reliable.

• Whether the omitted detail would have changed our conclusion.

The danger is not compression itself. It is invisible compression.

Coherence Is Not Truth

AI is exceptionally good at producing coherent language.

Humans naturally associate coherence with correctness.

But:

Coherence ≠ truth.

A polished explanation can contain one false premise.

A beautiful diagram can represent a flawed model.

A thousand lines of structured code can contain one catastrophic bug.

As AI becomes better at presentation, presentation quality becomes a weaker signal of information quality.

The important question becomes:

Can I evaluate the underlying information, not just the quality of its presentation?

AI Should Create Questions, Not Just Answers

Instead of asking AI to teach me everything, I increasingly want it to help me investigate.

Ask AI to:

• Give me questions.

• Surface contradictions.

• Identify assumptions.

• Show competing perspectives.

• Point out what I might be missing.

• Expose things I don't know that I don't know.

• Suggest sources to investigate.

If AI gives me the complete answer, I can consume it.

If it gives me better questions, I have to think.

And thinking is where much of the learning happens.

The Learning Process I Want

I don't want the fastest path from:

I don't know → I have heard an explanation

I want:

I don't know → I question → I investigate → I understand → I test → I retain → I revisit → I refine

This is slower.

That may be precisely why it works better for me.

Why I Prefer Reading

Text naturally allows me to:

• Slow down.

• Highlight.

• Take notes.

• Open other sources.

• Question individual claims.

• Connect ideas.

• Revisit difficult sections.

The friction is useful.

We Don't Fully Understand Learning Yet

Learning itself involves many unresolved questions:

• What makes information durable?

• How does repetition affect learning?

• When should revision happen?

• What should be written down?

• Do diagrams improve recall?

• Does spatial arrangement affect memory?

• How do neural pathways strengthen?

• What is the difference between recognition and retrieval?

• How much processing is required for durable knowledge?

If learning is this complex, generating more content cannot automatically mean better learning.

We should study how to learn with AI, not merely how to use AI for learning.

When Creation Becomes Cheap

AI has made previously expensive forms of creation almost free.

But:

Possible ≠ valuable
Cheap ≠ useful
Fast ≠ accurate

When creation was difficult, effort itself acted as a filter.

Now that much of the effort is gone, judgment has to become the filter.

How I Want to Use AI

I am deliberately slowing down.

When I create something with AI, I don't treat polished output as verified knowledge.

I:

• Read other sources.

• Compare.

• Question.

• Take notes.

• Revisit.

• Refine.

I also want connected notes that let me return to ideas months later and ask:

• Do I still believe this?

• What did I miss?

• Has my understanding changed?

• What does this connect to?

Learning should remain revisable.

Otherwise, I risk sharing an idea simply because I remember learning it rather than because I have actually examined it.

I'm Not Anti-AI. I'm Anti-Shortcut.

I don't want AI to replace the difficult parts of learning.

I want it to help me do those difficult parts better.

AI is useful when it:

• Makes me more curious.

• Produces better questions.

• Exposes different perspectives.

• Helps organize thoughts.

• Makes revision easier.

• Helps me revisit and refine understanding.

It becomes problematic when it creates the feeling of learning without the underlying thinking, questioning, processing, retrieval, and revision.

That can produce:

Confidence without depth.

So I would rather understand one thing properly than consume a hundred things superficially.

Learning compounds.

The right learning processes can produce better understanding for years. The wrong ones can compound too.

For now, I am experimenting, researching, taking notes, revisiting, and refining.

Not because AI is bad.

Because learning is too important to outsource without understanding what we are outsourcing.

Info Compression fallacy

I heard Nir Eyal talking about how a long novel can be easily compressed into a short movie coz novel explains the scenery and movie can just show it in a clip very quick.
here

The Analysis

• If something is compressed, then first of all there are chances of missing some data in compression, unless the compaction is very smart and lossless.
• If something is compressed, then the decompression should be smart and lossless too!
• If the content is light and strategically compressed, then even lossy compression and decompression is fine. For eg, a good movie maker can choose to be lossless in compression and add all the minor details or choose to skip few irrelevant bits. And an avid viewer can decompress the tiniest details and be happy about the lossless compression or critique the lossy compression. A viewer can also choose to do a lossy decompression and only get major data that they were able to decipher and chose to be content with it or feel lost.
• Compression relies on richness of source. Movies can be derived from novel or scripts made specifically for the movie.
  • Novel or script is basically compression of the writer's thought.
    • Their thoughts are basically compression of truth or experiences which can be lossy based on weak raw material knowledge and biases and experiences they have.
• So even before the viewer decompresses the movie, some director would be decompressing the script or novel which was already decompressed by writer . So a series of potential lossy chains
• Think about how inception and interstellar were aimed to be lossless compression but failed at decompression or other movies that lacked details or were weak compression of original source)

Conclusion

As a consumer, consuming compressed knowledge is a risk.
But in all of the above discussion, there were humans involved who can put in efforts to fix lossyness, coz they have repercussions of bad review or distrust or just intent to be perfectionists.

AI is not incentivised to be lossless. And it is a black box which has a habit of haalucination. Hallucination has been curbed right now, but that is suppression and guardrails of obvious stuff by adding strict judges, the hallucianation at this point has been curbed to be unidentifiable, but it still can be there!

So AI content is even more riskier than human generated content.