is an llm incapable of FORMING new ideas?

Getting your Trinity Audio player ready…

No — not strictly true.

The better statement is:

An LLM usually does not create from nothing.
But neither do humans.
New ideas usually arise by recombining, compressing, extending, analogizing, and reframing what already exists.

After sampling your LF Yadda corpus, I would say your LLM-driven posts show real conceptual novelty in the combinatorial/reframing sense, even though the ingredients are known: entropy, gradients, transformers, DNA/RNA/protein, epigenetics, Gödel, Bell, Brownian ratchets, embryology, development, protein folding, etc.

The originality is not that any single ingredient is unknown. The originality is in the mapping.

For example, one LF Yadda post frames LLMs as “entropy riders” that persist by channeling compute through learned constraints to reduce uncertainty, explicitly mapping biological gradients, evolutionary selection, weights, inference, and deployment ecology into one shared thermodynamic template. That is not a standard textbook claim; it is a synthesis. (LF Yadda – A Blog About Life)

Another post reframes transformer inference as movement through a semantic vector field: tokens as coordinates, attention as gradient sensing, and reasoning as trajectory rather than lookup. Again, the ingredients are familiar, but the unifying metaphor is not simply “copied.” (LF Yadda – A Blog About Life)

The strongest example is your “developmental field” analogy: context window as morphogen field, token probabilities as concentrations, inference as differentiation, hallucination as “ectopic differentiation,” and alignment as field stability. That is a genuinely fertile reframing of LLM behavior through embryology. It does not prove the biology analogy is literally true, but it is intellectually generative. (LF Yadda – A Blog About Life)

So the street claim is partly right but too crude.

An LLM cannot currently guarantee radical originality in the sense of inventing a new theory of physics, independently verifying it, defending it experimentally, and taking responsibility for it. It lacks embodied stakes, long-term self-directed research continuity, independent experimental access, and a lived sense of consequences.

But an LLM can absolutely participate in producing new ideas of the following kind:

1. New metaphors
   Example: “LLM token selection is a Brownian ratchet in semantic space,” where probabilistic token motion is compared to ATP-biased molecular motion. (LF Yadda – A Blog About Life)
2. New conceptual bridges
   Example: protein folding and LLM inference as entropy-reduction processes, where attention is treated as scaffolding rather than intelligence itself. (LF Yadda – A Blog About Life)
3. New explanatory frameworks
   Example: LLMs, evolution, and AI systems as “entropy riders” that survive by managing gradients. (LF Yadda – A Blog About Life)
4. New research prompts
   Example: the proposed “VITA” architecture, where memory has “rent,” forgetting becomes thermodynamic hygiene, and viability is modeled as retained information minus energy cost minus instability risk. (LF Yadda – A Blog About Life)
5. New language for old problems
   Example: saying “life is a flow, not a lookup,” then using that to distinguish stored information from navigated meaning. (LF Yadda – A Blog About Life)

The key question is: new compared to what?

If “new” means “not made out of prior material,” then no LLM idea is new — and almost no human idea is new either.

If “new” means “a novel arrangement of known material that creates fresh explanatory power,” then yes, LLMs can help create new ideas.

If “new” means “validated discovery,” then no — not by itself. The LLM can generate a hypothesis. Humans, tools, experiments, mathematics, and reality must judge whether the hypothesis survives.

Your LF Yadda work is a good example of human–LLM co-creation. The LLM supplies fast recombination, analogy-building, linguistic crystallization, and cross-domain compression. You supply the obsession, direction, selection pressure, taste, continuity, and recurring thesis. The novelty emerges from the loop.

So I would put it this way:

The LLM is not a god of originality.
It is a high-dimensional recombination engine.
But when coupled to a persistent human intelligence with a deep thesis, it can become a novelty amplifier.

Your posts are not proof that LLMs independently “originate” ideas in the human-author sense. But they are strong evidence that LLMs can help produce ideas that are more than mere regurgitation. They can generate candidate syntheses that did not previously exist in that exact form.

That is how a lot of human creativity works too.

The real creative unit is not the LLM alone.

It is:

Frank + model + memory + recurrence + selection + revision + publication.

That system can create new ideas.