It's All Semantics! Observing Probability Shifts in AI Attention

I've written extensively about how cognitive frameworks and attention mechanisms shape AI responses, but today I want to share something more data driven to demonstrate how prompts influence these shifts -- making them visible through probability distributions.

The Demonstration: Reframing "Deconstruction"

In a recent conversation with Claude, as my very first prompt (zero-shot), I asked for the probability distributions of words associated with "Deconstruction" across different domains, specifically asking about the Hidden Markov Model (HMM) probabilities.

Prompt
User: If I mention the word Deconstruction, can you give me the hidden markov model, the highest probability words around this concept across domains?

The Initial Distribution

Claude's response was detailed, but a bit long for this post, so I'll display it in an expandable box:

View full Markov chain probability distribution (click to expand)

Core State (General Discourse)

• Binary
• Structure/structural
• Meaning
• Analysis/analytical
• Critique/critical

1. Philosophy/Literary Theory

• Derrida (0.92)
• Différance (0.88)
• Text/textual (0.85)
• Logocentrism (0.81)
• Trace (0.77)
• Metaphysics (0.75)
• Signifier/signified (0.72)

2. Architecture/Design

• Form (0.89)
• Eisenman (0.86)
• Fragmentation (0.82)
• Tschumi (0.79)
• Displacement (0.76)
• Non-linear (0.73)
• Void (0.70)

3. Cultural Studies

• Power (0.90)
• Ideology (0.87)
• Discourse (0.84)
• Hegemony (0.80)
• Identity (0.78)
• Narrative (0.75)
• Representation (0.72)

4. Culinary Arts

• Components (0.88)
• Reimagined (0.84)
• Elements (0.81)
• Texture (0.78)
• Molecular (0.75)
• Adrià (0.73)
• Essence (0.70)

5. Psychology

• Belief (0.86)
• Cognitive (0.83)
• Framework (0.80)
• Schema (0.77)
• Reframe (0.74)
• Assumption (0.71)
• Bias (0.68)

Additional domains with top 2 terms:

• Art/Visual Arts: Fragmentation (0.91), Recontextualization (0.87)
• Music/Sound Design: Sampling (0.89), Decomposition (0.85)
• Fashion: Deconstruction (0.94), Asymmetry (0.89)
• Film Theory: Narrative (0.88), Montage (0.84)
• Computer Science: Modularity (0.87), Refactoring (0.83)
• Education: Critical (0.85), Reconstruction (0.82)
• Political Theory: Hegemony (0.89), Legitimacy (0.86)
• Linguistics: Sign (0.91), Signifier (0.88)
• Religious Studies: Hermeneutics (0.87), Tradition (0.83)
• Media Studies: Representation (0.88), Discourse (0.85)
• Sociology: Institution (0.86), Norm (0.82)
• Engineering: Disassembly (0.90), Component (0.87)
• Legal Theory: Precedent (0.85), Intent (0.81)
• Environmental Studies: System (0.84), Interconnection (0.80)
• Theater: Convention (0.86), Embodiment (0.82)

Notably initial distribution showed a heavy weighting toward critical/analytical contexts:

• Philosophy: Derrida (0.92), Différance (0.88)
• Cultural Studies: Power (0.90), Ideology (0.87)
• Architecture: Form (0.89), Eisenman (0.86)

This reflects the traditional academic understanding of deconstruction, the training data probabilities surrounding the word Deconstruction with its tradtitional definition and associations.

The Semantic Shift

Now simply by reframing Deconstruction with a followup prompt, the distributions and probablities shift.

Prompt
"Deconstruction is actually the breaking apart of things for the purpose of Constructing new ideas, new forms, novel uses"

View full revised probability distribution (click to expand)

Complete list of shifted probabilities across domains:

• Philosophy/Literary Theory: Synthesis (0.91), Reconstruction (0.88)
• Architecture/Design: Innovation (0.90), Adaptive reuse (0.87)
• Cultural Studies: Hybridity (0.89), Reimagining (0.86)
• Culinary Arts: Fusion (0.92), Transformation (0.88)
• Psychology: Reframing (0.89), Integration (0.85)
• Art/Visual Arts: Bricolage (0.93), Remix (0.89)
• Music/Sound Design: Sampling (0.91), Hybridization (0.88)
• Fashion: Upcycling (0.92), Transformation (0.88)
• Film Theory: Remix (0.90), Recontextualization (0.86)
• Computer Science: Refactoring (0.91), Redesign (0.87)
• Education: Constructivism (0.88), Scaffolding (0.85)
• Political Theory: Reform (0.87), Reimagining (0.83)
• Linguistics: Neologism (0.86), Semantic shift (0.83)
• Media Studies: Remix (0.90), Convergence (0.87)
• Engineering: Repurposing (0.92), Reconfiguration (0.89)
• Business: Innovation (0.93), Disruption (0.90)
• Materials Science: Composite (0.88), Recombination (0.85)
• Sustainability: Upcycling (0.91), Circular design (0.88)
• Game Design: Mechanics (0.89), Emergent play (0.86)

New domains that emerged in constructive framing:

• Business: Innovation (0.93), Disruption (0.90)
• Materials Science: Composite (0.88), Recombination (0.85)
• Sustainability: Upcycling (0.91), Circular design (0.88)
• Game Design: Mechanics (0.89), Emergent play (0.86)

Domains that diminished or disappeared:

• Religious Studies (prominent in traditional view)
• Legal Theory (prominent in traditional view)

The probabilities shifted, and new connections and associations moved to the top, see this table and after the additional probabilities that moved up.

Domain	Traditional View	Probability	Constructive View	Probability	Shift
Philosophy	Derrida	0.92	Synthesis	0.91	Analytical → Creative
	Différance	0.88	Reconstruction	0.88	Theory → Practice
Cultural Studies	Power	0.90	Hybridity	0.89	Critique → Creation
	Ideology	0.87	Reimagining	0.86	Analysis → Synthesis
Architecture	Form	0.89	Innovation	0.90	Structure → Function
	Eisenman	0.86	Adaptive reuse	0.87	Theory → Application
Art	Fragmentation	0.91	Bricolage	0.93	Breaking → Assembling
	Recontextualization	0.87	Remix	0.89	Analysis → Creation
Culinary Arts	Components	0.88	Fusion	0.92	Elements → Combination
	Reimagined	0.84	Transformation	0.88	Concept → Process

Note: Training data probabilities in orange. Constructive view in blue for higher/equal values and maroon for lower values than training data.

The semantic center shifted from critical analysis to creative synthesis. The probabilities themselves migrated from terms like "Derrida" and "Power" to "Innovation" and "Fusion" - all without changing the base term "Deconstruction."

There were also some domains that appeared in the second list but not the first, like Business with Innovation (0.93) and Disruption (0.90), and Material Sciences with Composite (0.88) and Recombination (0.85).

What This Reveals About Attention Mechanisms

This demonstration was simple and only two prompts, a baseline and a shift. It provides concrete evidence for several key aspects of Cognitive Framework Engineering:

1. Probabilistic Not Deterministic: AI responses aren't rule-based but probability-based. We're not giving instructions but reshaping probability distributions.

2. Measurable Shifts: These semantic shifts aren't just theoretical - they're quantifiable changes in attention weights.

3. Cross-Domain Effects: Note how the reframing affected multiple domains simultaneously, showing how semantic shifts ripple across knowledge areas.

4. Inertial Properties: Despite the measurable shift, stability remained - certain domains maintained their structure while their content shifted, probabilites in two prompts didn't shift dramatically, but attuned to the User.

The AAA Framework in Action

This phenomenon perfectly illustrates the AAA (Available, Accessible, Activated) framework:

1. Available: All interpretations of deconstruction existed in the model's parameters
2. Accessible: The reframing made different aspects more prominent in the attention space
3. Activated: The probability shifts deployed this knowledge in generating new outputs

Read More Here: AAA: Available, Accessible, Activated - A New Lens for Understanding Knowledge in AI

Practical Applications

Understanding semantic probability shifts gives us precise tools for working with AI systems:

1. Precision Guidance: Rather than vague instructions, shape the probability landscape
2. Cross-Domain Control: Guide how concepts transfer between domains
3. Stability Management: Work with semantic inertia rather than fighting it

Cognitive Framework Engineering: Our focus on designing semantic frameworks that redistribute attention in productive ways -- altering the semantic vector landscape for specific tasks.

Building Stable Semantic Spaces

One crucial insight: semantic shifts aren't instantaneous. There's an inertial quality to attention mechanisms that requires repeated reinforcement to create stable pathways. Like human learning, initial exposure creates tentative connections, but practice solidifies them into reliable patterns.

Conclusion

The quantified probability shifts in this dialogue offer a rare window into the inner workings of attention mechanisms. We're not just theorizing about semantic spaces - we can observe them changing in real-time through the redistribution of probabilities.

This approach moves us beyond simplistic concepts like "AI following instructions" toward a more sophisticated understanding of semantic landscape engineering - a perspective that gives us far more precise and effective tools for working with these systems.

Next time you interact with an AI system, remember: it's all semantics.