Skip to content

concepts.md

Latest commit

 

History

History
170 lines (121 loc) · 6.02 KB

concepts.md

File metadata and controls

170 lines (121 loc) · 6.02 KB

Core Concepts

The Central Insight

Ideas are atomic. Position is relational.

An observation like "Lives paycheck to paycheck at 42" has no inherent meaning. Its meaning emerges from context - what you're optimizing for, what question you're answering.

In one context, it's evidence of financial risk. In another, evidence of sacrifice for purpose. In another, evidence of choosing freedom over salary.

The World Model captures this by allowing the same observation to appear in multiple trees with different positions (PRO or CON).

Observations vs. Beliefs

Traditional profile systems store beliefs:

  • "Andrei believes in decentralization"
  • "Andrei values autonomy"

These are already interpreted. Position is baked in.

The World Model stores observations - atomic facts without inherent polarity:

  • "Built DAO governance framework"
  • "Rejected VC funding"
  • "Works alone"

Position emerges from competition between agents.

The Debate Model Foundation

The architecture extends a tokenized debate model with binary tree structure:

                    [Claim]
                   /       \
                PRO         CON
               /   \       /   \
            [obs] [obs]  [obs] [obs]

Weight propagation formula:

net_score = direct_weight + sum(pro_children.score) - sum(con_children.score)

A claim's strength depends on:

  1. Direct stakes (how much agents invested)
  2. Supporting evidence (PRO children)
  3. Contradicting evidence (CON children)

Agents as Internal Plurality

A person isn't monolithic. They're a coalition of drives finding equilibrium.

The seven tendencies model universal human drives:

Tendency Question It Asks
SURVIVAL "Is this safe? Do I have enough?"
STATUS "Am I respected? Do I matter to others?"
MEANING "Does this matter? Will it outlast me?"
CONNECTION "Am I known? Do I belong?"
AUTONOMY "Am I free? Can I choose?"
COMFORT "Is this pleasant? Can I sustain this?"
CURIOSITY "Do I understand? What's there to learn?"

The allocation represents how strong each drive is in a specific person. Andrei might be high MEANING/AUTONOMY, low STATUS/COMFORT.

Adversarial Competition

This is what makes the model "alive."

In passive models, agents just categorize observations. In the adversarial model:

  1. Agents PROPOSE: Each agent makes a claim about what matters
  2. Agents STAKE: Support their claims, undermine competitors
  3. Agents WIN/LOSE: Based on how well evidence supports claims
  4. Allocations SHIFT: Winners gain influence for future debates

The competition surfaces tensions. When SURVIVAL and MEANING stake the same observation with opposite positions, that's internal conflict made visible.

The Equilibrium IS The Personality

Traditional models describe a person. This model is the person's decision-making structure.

The equilibrium - which tendencies dominate, how they relate, where they conflict - constitutes personality. Not a representation of it. The thing itself, in computational form.

Training as Calibration

ML-style training isn't teaching the model new information. It's calibrating the agent allocations to match the evidence.

  • Epochs: Multiple rounds of debate
  • Convergence: When allocations stabilize (tendencies find equilibrium)
  • Validation: Testing on held-out observations
  • Learning rate: How fast allocations shift per round

The goal isn't accuracy in a traditional sense. It's finding the equilibrium that best explains the person's observations.

Multi-Lens Meaning

The same observation has different meaning vectors across trees:

"Lives paycheck to paycheck"
  ├── SURVIVAL tree: +0.8 (strong evidence)
  ├── MEANING tree: +0.6 (sacrifice for purpose)
  ├── AUTONOMY tree: +0.4 (chose freedom)
  ├── COMFORT tree: +0.7 (unsustainable)
  └── STATUS tree: -0.3 (against status claims)

Observations that flip polarity across trees reveal tensions. Observations that are consistently PRO reveal core identity.

The Bet

This architecture bets that:

  1. Binary distinction + recursion = sufficient to model meaning
  2. Seven tendencies = comprehensive but tractable agent set
  3. Competition = produces coherent personality from plurality
  4. Same structure = applies to anyone (different allocations + observations)
  5. More useful = than unstructured text at scale

If the bet pays off, this is a general architecture for modeling minds.

What This Doesn't Model

  • Temporal dynamics: How the person changes over time
  • Context switching: Different "modes" in different situations
  • Emotional states: Momentary feelings vs. stable tendencies
  • Social relationships: How the person relates to specific others
  • Skills/Knowledge: What they can do vs. what drives them

These could be extensions, but the core model focuses on stable value structures.

Key Formulas

Weight Propagation

net_score = direct_weight + sum(pro_children.score) - sum(con_children.score)

Allocation Rebalancing

target_allocation[t] = score[t] / sum(scores)
new_allocation[t] = current[t] + (target[t] - current[t]) * learning_rate

Validation Accuracy

accuracy = correct_predictions / total_predictions
baseline = 1 / num_tendencies  # Random chance
p_value = binomial_test(correct, total, baseline)

Terminology

Term Definition
Observation Atomic fact about a person (~280 bytes)
Agent Human tendency with token allocation
Tendency One of 7 universal human drives
Allocation Agent's share of influence (0.0-1.0)
Tree Value hierarchy with root claim
Claim Statement proposed by an agent
Node Observation positioned in a tree
Position PRO or CON relative to parent
Stake Weight an agent assigns to a node
Arena Where adversarial competition happens
Epoch One full round of debate
Convergence When allocations stabilize