Devlog: Gallery of Conceptual Lenses

I spent a chunk of a day digging into prompt engineering and modern research: basically, how do you get better results from a model?

That immediately turned into a side-quest: how do you extend the model’s “horizon”—not by asking harder, but by getting it to explore more of what it could plausibly access.

The “latent space” rabbit hole (and the problem)

When you send a prompt to an LLM, you’re lighting up particular associations—neurons, connections, hidden-layer pathways—based on how your prompt relates to everything the model has learned.

That’s usually great: it’s efficient, it stays on-topic, it gives you the “reasonable” answer to the thing you asked.

But if you want the model to be more creative, the core question becomes: How do you get it to search wider—into regions it wouldn’t normally touch?

The idea

I’m calling it a Gallery of Conceptual Lenses.

A “lens” is a frame of thought: a philosophy, a discipline, a language, an objective—basically a handle that forces the model to approach the same prompt from a different angle.

So instead of “one prompt -> one default trajectory,” you do:

one prompt -> many intentionally distant trajectories.

How it works (mechanics)

Here’s the mental model / algorithm as I’m currently thinking about it:

1. Embed the prompt into an embedding space (i.e., locate it in the “sea” of meaning).
2. Pick a set of lenses (philosophies / frames).

3. Generate combinations of lenses (example in my notes: if you pick four lenses, you build a “4x4 matrix” of combinations).

   • Example vibe: Buddhism x JavaScript x Gravel x Sky (yes, that kind of weird).

4. From those combinations:

   • Find the combo closest to the prompt (to stay anchored).
   • Then find combos that are as far away from the prompt and from each other as possible (to maximize spread across the space).
5. Now you’ve got a set of far-separated “approach points”—multiple ways into the same problem, deliberately distributed.
6. You tell the model: answer using these distinct approaches—each one pushes activation into a different region, which (in practice) tends to pull in more novel material.

Why it’s useful (what I’m seeing)

Once you do this, each lens-combo “activates” farther regions of the embedding space, so you get:

• more creativity,
• more weird cross-links,
• more usable brainstorming output,
• more chances of stumbling into a novel angle.

This has been a legit tool for “think outside the box” tasks and idea generation.

Tradeoffs / known drawbacks

This probably makes the model less focused. I’m basically forcing it to explore, so drift is part of the deal.

So: not ideal if you need a tight, single answer. Great if you need a spread of approaches.

Current one-line thesis

Creativity via distance: deliberately choose conceptual lenses that are far apart (and far from the prompt) to widen the model’s search, then harvest the best results back toward the original intent.