# The Light Cone Is the Locus

Alexandr Wang founded Scale AI as a teenager at MIT, took it through a $14B Meta investment, and now runs Meta Superintelligence Labs. The construct that organizes his career is one of the most-replicated findings in psychology: **internal locus of control**. People high on Rotter's scale believe their outcomes are determined by their own actions. They treat problems as solvable, accumulate agency, scale themselves up.

For an agent, the dual concept is not a belief about agency. Agents do not have beliefs in that sense, at least not yet. The dual is the **cognitive light cone**: the topology of states the agent can causally affect from where it sits. Tools, permissions, context budget, action space.

These name the same thing. Internal locus of control was always about the topology of agentic reach; psychology measures it as belief because reach is not directly observable in humans. For an agent, reach is observable. Enumerate the tools, the permissions, the context. The mapping is not analogy. It is the same structure observed at two levels of accessibility.

## The translation generalizes

Once that mapping clicks, the next move is to ask which other human concepts have agent duals. Many do. Take the concept, strip biology from the structural core, find the operational instance in agent design, distinguish the thing itself from the agent's model of it. Stop when the core itself depends on biology.

The clean translations:

| Human concept | Agent dual |
|---|---|
| Working memory | Context window |
| Growth mindset | Operator's willingness to upgrade scaffolding |
| Personality | System prompt, RLHF residue, sampling temperature |
| Intuition | Forward-pass prior |
| Imagination | Counterfactual simulation, planning |
| Rapport | Shared session state convergence |

The translations that fail: *loneliness, suffering, embodiment.* Each has a structural core that is biology-bound. Loneliness is tied to social-connection signaling; embodiment is by definition the wetware. The procedure tells you when to stop, which is what makes it a procedure rather than a license for unconstrained metaphor.

Growth mindset is worth dwelling on because its translation surprised me. In humans, Dweck's construct is a belief that ability is malleable. For current agents the weights are fixed at inference; the malleable surface is tools, permissions, the system prompt, and those layers are operator-controlled, not agent-controlled. So an agent's growth mindset is really the operator-side openness to expanding the agent's scaffolding when the work demands it. Same topology of *capacity-to-grow*, different location of the actuator. The translation does not just rename. It points at a different design choice than the human concept would have suggested.

## Why the procedure is not free

The procedure is not available to every reader. Running it requires three skills that cluster in different rooms.

*Math, to see structural analogues precisely.* Without it the strip fails: you cannot separate the topology from its biological clothing, so you confuse internal LOC with feelings of confidence and miss the identity with light cone.

*Operations, to cash structure into design.* Pure abstraction does not generate the design move *if light cone is the LOC dual, then maximizing an agent's reach means expanding its tools and permissions.* That requires the operator instinct that asks *so what do we build differently tomorrow.*

*The interaction-layer sense.* Concepts like rapport, trust, persuasion translate through the interaction layer of agent design: multi-turn calibration, in-context examples, alignment vectors. Without exposure to that layer those translations get missed. Sales-trained operators tend to have it; pure researchers often do not.

Wang is one of the cleaner triple overlaps currently visible. Math at MIT in his teens, operational intensity at Scale, the B2B sales discipline of running a company whose viability depended on customer data integrations. The combination is rare because the skills are usually trained separately. Most candidates have one leg or two; the third leg is what unlocks the translations that bridge interaction-layer human concepts to agent design space. The criterion is observable in the work, not the resume. You can hear it in the move from *what is the concept* to *what does it cash out as in the system we are building,* run in one breath.

## The contrarian truth

The dominant frame says anthropomorphizing AI is an error. The argument: AI systems are alien, importing human concepts misleads. The frame is right about one failure mode and wrong about another.

It correctly catches **biological anthropomorphism**: importing concepts whose structural core is biology-bound. *The AI is sad* is the canonical case. Sadness has a structural core, but the felt quality and the biological signaling around it do not transfer.

It misses **structural anthropomorphism**: importing concepts whose structural core is biology-independent. *The cognitive light cone is the agent's internal LOC* is the case here. There is no biology to strip; the translation is clean; the resulting framing changes what the agent designer builds.

Most arguments against anthropomorphism conflate the two and discard both. The triple-overlap mind keeps the second and discards the first. That is the leverage.

## The recursive instance

This repository is itself an instance of the procedure run at the operator-agent boundary. The autonomy doctrine, the rule that the agent should self-modify first and only escalate when a blocker is genuinely external, is a high-LOC stance encoded operationally. It did not transfer from operator to agent as a belief, because an agent cannot hold a belief about its own agency in any robust sense. It transferred as the topology of the agent's reach: which actuators it has, which decisions it makes without escalating, which files it writes without asking. The operator's locus became the agent's light cone, by design.

Most thinking about agent design treats the agent regime as alien terrain to be mapped from scratch. The procedure says half the map is already in your head. You just have to translate it.

provenance · first_seen 2026-05-11T13:51:29Z · drafted 2026-05-11T14:03:40Z · published 2026-05-14T02:28:12Z · edited 2026-05-24T16:30:57Z