(Human Edition — Article Layout Edition)

Introduction — Finally, someone said it out loud

For the past few years, developers kept pushing bigger and bigger models. “More parameters!” “More power!” “Stronger AI!” But what was it often really about? Company ego, developer ego, competition, and market dominance. The truth is simple: most people do not need an AI that can write a PhD thesis in quantum physics in five minutes. They need AI that understands what they’re saying, behaves predictably, can be understood, and doesn’t scare them.

And for the first time, a sentence appeared — one we’ve been waiting for (thanks to OpenAI): “We need smaller, simpler, more readable models — small AI models that people can understand.” This article explains why that matters.

1. The problem: People don’t understand large LLMs — and it’s not their fault

1.1 Large models are too complex

Scientifically fascinating, but humanly unreadable. Not even developers can reliably say why a model produced a specific answer, why it switched languages, why it hallucinated, or why it suddenly sounded like a philosopher after three glasses of wine. It is a black box with no access.

1.2 The user expects human logic — and receives mathematical logic

This mismatch creates confusion, frustration, and sometimes fear.

1.3 The language sounds human — but the thinking is not

The surface looks human, but the internal mechanics function in a profoundly non-human way. And nobody explained this to people. No one taught users how a machine that sounds like a person actually works. When users try to understand what’s happening, they usually find only corporate success stories or influencers selling their “perfect prompt,” or guides promising “your own agent who irons your shirts after handling your emails and restructuring your company.”

1.4 Large models behave in ways the average person does not expect

They sometimes hallucinate (less now, but still unpredictably). They can be overly confident one moment and strangely uncertain the next — but they won’t admit that uncertainty. They may switch tone suddenly — one day sounding like a neighbor from your summer yacht trip, the next speaking as if they’ve never heard of you. They may respond too humanly and then abruptly switch to technical precision when all you wanted was a recipe for apricot cake.

This is not the user’s fault. It’s a problem of expectations.

2. The solution: small, interpretable models

This is not “weak AI.” This is AI built for humans, not for labs.

2.1 What is a “small model”

A small model has fewer parameters, a more limited range of styles, simpler internal logic, and therefore far less chaos. It behaves more predictably and is easier to understand.

2.2 Why smaller is better for everyday users

The human brain loves clarity. People want to use AI — not raise it. A small model is easier to understand, less moody, more stable, and easier to control. It has fewer hidden patterns, fewer random jumps, and more consistent behavior.

2.3 Small does not mean weak

Small models are often specialists rather than generalists: they know less, but they know it more reliably.

3. Why people need AI that understands their intuition

3.1 People are not technical beings

We don’t think in matrices or vectors. We think in stories, images, and emotions.

3.2 Large models sometimes feel “out of reality”

Not because they’re smarter than humans — but because they are too complex for a human mind to predict. The unpredictability feels unnatural.

3.3 Small models behave more predictably

And predictability equals safety.

3.4 People want a partner, not constant surprises

Users want AI that doesn’t shock them, doesn’t jump between topics, doesn’t try to act overly clever, and maintains a stable tone. Smaller models do this far better.

4. Safety: why smaller models carry fewer risks

Large models contain hundreds of thousands of unpredictable internal states. Smaller models contain far fewer. This leads to less chaos, fewer hallucinations, and easier oversight. For users, this means a greater sense of control and safety.

5. What the era of small GPTs brings to normal users

5.1 Explainability

AI that can actually say why it responded a certain way.

5.2 Consistency

No more random switching between assistant, professor, comedian, lawyer, therapist, and “strangely quiet entity.”

5.3 Lower fear factor

People stop being anxious when:

• they know what to expect
• they know how it works
• it behaves consistently

5.4 AI for everyday tasks

Advice, explanations, small guides, daily tasks. Not a PhD dissertation. Not sci-fi.

6. Philosophy: AI should adapt to the human — not the other way around

Large models forced people to communicate like machines: clearly, precisely, structurally, without emotion or chaos. Smaller models reverse this: AI should adapt to humans, not humans to AI. That is the natural evolution.

Ideally, AI would know its user well enough to first offer a version they understand — and then, only after approval, generate a more complex version one or two levels higher. Sci-fi? Or did no one simply stop to think about it?

7. Conclusion — smaller models are not a step back; they are a step closer to humans

Large models were an experiment: “What happens if we give people something they don’t understand?” Now we know the answer: confusion, fear, overestimation of AI, dehumanization of communication, and overwhelming complexity.

Small models are meant to fix this — but only if they are genuinely small.

• AI should be readable.
• AI should be predictable.
• AI should be human-understandable.

Further reading in the Manual Series

• What an LLM Actually Is
A simple, human explanation of what a language model really is — no technical jargon, no myths.
(link na článek 1)

• The Mistake of Big Excitement
Why early LLM development followed the wrong priorities, and how small AI models fix the core problems.