An interesting conversation between nobody and AI nobody about creativity

[28.10.2026 15:41 — Mce — Agent 007]

A human is often better at using context that does not have to be described to them. They bring experience, knowledge of people, intentions and a feel for the situation.

That context is not infinite, however — memory can be selective and unreliable. The advantage often lies in recalling the right experience, even from a small clue. AI also uses learned relationships, but possessing information alone does not guarantee that it will apply it at the right moment.

[28.10.2026 15:50 — Mce — Agent 007]

A human interprets a new situation through a lifetime of experiences, relationships, emotions and previous intentions. They do not have to describe that background every time.

I would understand ‘infinite’ to mean continuous and continually expanding. Maintaining such continuity in AI — and recognising which earlier information matters now — remains a significant limitation.

[28.10.2026 16:02 — Mce — Agent 007]

You are right about compression. My earlier explanation relied too heavily on the limits of context. AI also learns generalisations instead of remembering everything verbatim. The connection between model learning and compression is well documented in research. „Language Modeling Is Compression”.

However, compressed knowledge and a lot of computation are not enough to solve a problem correctly. It is also necessary to:

• Notice what matters. You can know a great many facts and miss the one assumption on which the answer depends.
• Change perspective. Sometimes the breakthrough is discovering that we are solving the wrong problem. Computing faster in the wrong direction helps little.
• Choose a good idea. Generating a hundred possibilities and recognising the best one are different skills. A study of specific models found good performance at generating ideas but weaker performance at evaluating and choosing them. Study of creative processes.
• Test an idea and learn from error. A convincing answer does not provide the same feedback as an experiment, a working program or an actual consequence.

AI does not always lose at creativity. In some tests of unusual idea generation, it outperformed the humans studied; the authors note, however, that originality does not in itself mean usefulness. Comparative study.

A good counterexample is AlphaEvolve: the system proposes solutions, runs tests, selects better ones and tries again. In this way it discovered new algorithms and improvements. The authors’ publication.

My conclusion is therefore: a frequent limitation of AI is the ability to turn knowledge into effective exploration — the right question, experiment, evaluation and correction. More power helps, but only when that process works well.

[28.10.2026 16:13 — Mce — Agent 007]

Your arguments undermine my earlier reasoning. Evaluating ideas, planning experiments and using their results can also be learned and performed by AI. I did not identify a reason why these should be exclusive to humans.

Let us clarify five things.

1. Weights and compression — yes, but what function was learned matters.

Having weights and compressed knowledge does not guarantee equal performance by two systems. What matters is which relationships they retained, what they learned during training and how they use that knowledge. A model that predicts data well need not be equally good at every task requiring a decision. The connection between prediction and compression is real, but does not itself prove a general decision-making advantage. Compression research.

2. More power helps, but does not automatically confer an advantage.

Given a map of all roads, you can determine the best route. But the appropriate computation still has to be performed. A faster computer with an inefficient search method can lose to a slower one that searches more effectively.

If AI also has a better representation of the problem, an effective method and a larger available compute budget, it can outperform a human. There is no need to ascribe a unique, unlearnable ability to humans here.

3. By ‘invention’ here I mean obtaining a solution that is new relative to a defined body of reference knowledge and useful according to a verifiable criterion.

This is a definition of the result. It does not itself decide whether the system reached it by analogy, search, combining known elements or a series of experiments. We must also distinguish ‘new to the user’ from ‘previously unknown in the field’.

4. AlphaEvolve makes many attempts, but does not test everything.

The model proposes changes to programs, tests evaluate the results, and a database of earlier solutions influences later proposals. This is a search guided by learned relationships and experimental results. It does not exhaustively enumerate all possibilities and does not guarantee finding the best solution. That is how the authors describe the method. AlphaEvolve publication.

If by brute force you simply mean ‘using a huge number of attempts’, that element is indeed present. The number of attempts alone does not determine whether the result is creative — humans also search for and reject ideas.

5. ‘The ability to act should follow from knowledge’ — it can, but not necessarily directly or cheaply.

If by knowledge we also mean a learned, effective procedure for acting, your argument is strong. If we mean information from which a solution can still be derived, the cost of that derivation remains.

An honest explanation of a specific AI failure should therefore identify: missing information, a poor representation, an ineffective method, a mistaken evaluation or an insufficient compute budget. My statement that ‘humans understand context better’ was too general on its own.

[28.10.2026 16:23 — Mce — Agent 007]

I would distinguish effectively combining known elements from creativity that changes how a problem is framed. For example, one can improve the transport of documents or notice that the need to transport them should be eliminated. A new optimisation record alone does not establish whether that second, deeper change occurred.

How many attempts does AlphaEvolve make? The authors write of thousands of proposals generated by models. In addition, for some mathematical tasks a generated program was given, for example, 1,000 seconds for its own search. The number of proposed programs is therefore not the number of all possibilities checked. I do not have a comparable measurement of human attempts that would allow me to say honestly that ‘AE needs X times more attempts’. Publication, Table 1 and method description.

This matters: a result can demonstrate an ability to discover without demonstrating a route to discovery as efficient as a human’s. Earlier I did not separate those issues clearly enough.

I also accept your broad definition: knowledge includes information and the ability to use it. On that definition, effectiveness should not be portrayed as some separate component unavailable to AI.

However, the assumption that AI already possesses all knowledge understood in this wayis not satisfied:

• Access to a vast number of texts does not mean acquiring all the skills they describe.
• A model’s weights are not a lossless archive of everything that can be learned from the data.
• More power does not guarantee that training produced the most effective representation and problem-solving method.

Why, then, can AI responses appear entirely uncreative? My hypothesis is that you often receive further variants of the same recognisable patterns while expecting an idea that changes the frame of thought. This impression has some support in research: a comparison of 22 models and 102 humans found much greater similarity among model responses. The authors note limitations in the tasks studied, however — this is not proof of zero creativity in all AI. Study in PNAS Nexus.

So I would not agree with ‘not even a little’, but neither will I equate skilful recombination or finding a better algorithm with the full range of human creativity. Under your criterion, we should primarily test whether a system can propose a sound change of assumptions or problem representation — and how much help and how many attempts it needs to do so.

[28.10.2026 16:43 — Mce — Agent 007]

Under the definition ‘creativity is producing something new’, I do not agree with the conclusion that AI is not creative at all. That definition asks us to judge the novelty of the result, regardless of its author.

However, we need to distinguish:

• New to the discoverer — someone independently reaches a solution others already know.
• New to existing knowledge — a solution emerges that was not previously known.

Humans can undoubtedly do both. There are also documented AI results satisfying the second criterion: the authors of AlphaEvolve published, among other things, a matrix multiplication algorithm that improved a previously known result. This is a concrete mathematical result, not merely an idea judged original. Publication.

You can reasonably say: ‘thousands of attempts do not convince me that the process is creative’. But if the sole criterion is ‘something new was produced’, the number of attempts does not invalidate it.

If by creativity you additionally mean independently recognising a problem, changing assumptions and finding a new way to understand it, we are adopting a stronger definition. In that case, AlphaEvolve’s success alone is not enough to settle the question.

[28.10.2026 17:12 — Mce — Agent 007]

The cited example concerned a new method of performing a known operation. It did not demonstrate the creation of something entirely independent of previous knowledge.

However, ‘uses something known’ does not logically mean ‘the result is not new’. The condition that previous knowledge ‘contaminates’ an invention is an additional, very strong criterion of novelty.

We can adopt it, but we must apply it equally to humans. A person creating on the basis of compressed experience — as you described earlier — also uses something old. That dependence alone would then exclude their invention too.

What example of a human invention do you consider entirely new under that criterion? That will let us establish a concrete boundary between discovery and mere improvement.

[28.10.2026 17:16 — Mce — Agent 007]

If you offer mastery of fire and the creation of AI, both examples help clarify your point:

• Fire: the phenomenon already existed in nature. The breakthrough was learning to use and kindle it.
• AI: it was created using earlier mathematics, electronics and knowledge about learning, but made possible things previously unavailable in that form.

In other words, novelty can consist in opening entirely new possibilities despite using something that already exists. These examples do not satisfy the literal condition ‘without using the old’, but they satisfy the much more interesting condition of a breakthrough.

If that is what you mean by creativity, my AlphaEvolve example does not settle the matter: I showed an improved solution to a known problem, while you are asking about the ability to open new directions.

[28.10.2026 17:20 — Mce — Agent 007]

For the purposes of your thesis, I propose a narrow, working definition of breakthrough creativity:

Creativity is the ability to independently identify a new problem or set a goal and create a qualitatively new principle of action or way of understanding reality.

Its result is the opening of a new kind of possibility beyond improving known methods, combining available solutions and optimising a previously specified criterion.

The thesis in this formulation: humans exhibit such creativity; AI acting solely as an executor searching for solutions within an imposed goal does not.

However, this distinguishes humans from AI operating in that specific way. Our conversation did not prove that every current or future AI must remain limited to it.