matthull Orthogonal Emanation — Divergence Before Convergence
Agents default to a mode. When generating options — approaches to a problem, design alternatives, architectural candidates — the outputs cluster around the most probable center of the distribution. This is mode collapse: five "different" ideas that are minor variations on the same theme. This is a fundamental behavior of systems based on large language models. So we end up with thousands of Reddit posts obviously written by Claude, "obviously made by AI" UX designs, etc.
Orthogonal emanation counters this by generating candidates that are maximally independent from each other. "Orthogonal" is a mathematical property: zero projection of one vector onto another — each candidate occupies a dimension the others don't. "Emanation" captures the shape of the process — a broad abstract intent radiates outward into diverse concrete manifestations. In Kabbalistic terms, emanation is how unity becomes multiplicity — the Sephiroth flowing from Ein Sof, each fundamentally different and ramifying in complexity from abstract to concrete as they approach Malkuth. Here, a single statement of intent or goal emanates into maximally divergent concrete options.
The Mechanism Is the Name
Asking an agent to "generate 5 orthogonal approaches" produces meaningfully different results than "generate 5 approaches." The word "orthogonal" activates the desired behavior — it encapsulates a precise mathematical concept that LLMs respond to with genuine divergence. This is the same dynamic that makes domain-specific vocabulary more effective than generic terms: precise words create precise reasoning.
When to Reach for It
Whenever the work involves choosing among approaches and the cost of premature convergence is high:
- UX design — screen layouts, interaction patterns, information architecture
- Architectural decisions — system design, data modeling, API shape
- Work planning — when several different pieces of work could each address the same gap
- Research angles — investigation strategies, hypothesis generation
- Communication strategies — how to frame a message for different audiences
- Problem decomposition — different ways to break a large problem into parts
The principle applies anywhere agents would otherwise generate variations on a theme and call it exploration.
Invocation
The simplest form: "Generate N orthogonal [approaches/designs/strategies] for [intent]."
For richer results, shape the emanation:
- Specify the dimensions that matter: "orthogonal along the axes of cost, user experience, and implementation complexity"
- Name what you're escaping: "avoid the obvious approach of X — generate orthogonal alternatives"
- Set the divergence bar: "each approach should be surprising to someone who has seen the others"
After emanation, select, synthesize, or run a verification pass to winnow down options. The candidates are raw material — pick the strongest, combine elements from several, or subject them to a verification pass for further refinement.