Neither side is complete alone. The point is to distribute labor so each side does what they do best.
Links: AI as a Cognitive Tool, Economics, Risk and Entrepreneurship, Praxis, LLM Grounding Problem, Cognitive vs. Motor Skills, Claude Opus 4.6 Research, Gaming, Measurement, Causality, and Free Will, Karpathy LLM Wiki Convergence, The Three-Layer Method, Level 6 — Direct Execution, Career
Not replacement. Not augmentation. Integration. The human and the AI form a single working unit where labor is distributed based on comparative advantage. Each side contributes what they do best, and the result exceeds what either could produce alone.
This isn’t a metaphor — it’s an economic statement. From utility theory: trade happens because both parties have different capabilities and different costs. The human-AI collaboration is a trade: the human provides what’s expensive for the AI (judgment, grounding, physical action), and the AI provides what’s expensive for the human (speed, scale, tireless execution, parallel processing).
Entrepreneurial Judgment (CEO Function)
Grounding in Physical Reality
Intent and Direction
Quality Judgment (“Taste”)
Physical World Execution
Social and Emotional Intelligence
Cognitive Execution at Scale
Operational Coordination (COO Function)
Pattern Recognition and Retrieval
Creative Problem-Solving Under Low Stakes
Tireless, Consistent Operation
The video’s framing: “The humans set direction, evaluate quality, and make judgment calls. The agents execute, coordinate, and scale.”
This is correct but incomplete. The full distribution:
| Function | Human | AI | Why |
|---|---|---|---|
| Direction | Define goals, intent, what to build | Decompose goals into executable tasks | Humans know what’s needed; AI knows how to break it down |
| CEO Judgment | High-stakes risk decisions, entrepreneurial bets | Present options with analysis, recommend but don’t decide | Slow feedback loops make this hard to train; human bears consequences |
| COO Operations | Override when judgment needed | Routing, triaging, constraint enforcement, coordination | Structured, data-driven — AI’s native domain |
| Grounding | Provide physical context, validate real-world fit | Process and organize information within text domain | LLMs can’t distinguish lived experience from verbal claims |
| Cognitive Execution | Review, approve, iterate | Write code, generate content, analyze data, build | AI is faster, cheaper, tireless |
| Physical Execution | Hands-on work, physical presence, real-world interaction | Emerging robotics, but mostly indirect (controlling systems via software) | Different architecture needed — see cognitive vs motor |
| Quality | “Is this actually good?” — taste, domain expertise | “Is this technically correct?” — tests, validation, consistency | Correctness ≠ quality; humans judge the latter |
| Pacing | Must continuously update mental models of what’s possible | Improves automatically with each model release | The boundary between human and AI roles shifts every few months |
From Video 2: skills like judgment, taste, and domain expertise “now have 100x leverage because they are multiplied by the number of agents that person can direct.”
This is the economic argument for the cyborg model:
Revenue per employee at AI-native companies (5-7x traditional SaaS) reflects this. It’s not that the humans are 5-7x better. It’s that each human is leveraged by agents.
The emerging org model: 2-3 humans + a fleet of specialized agents, organized by outcome, not function. The humans set direction, evaluate quality, and make judgment calls. Everything else is delegated.
The capability boundary between “human territory” and “AI territory” is moving fast:
“Your January mental model of what AI can and cannot do is already wrong.”
A 2025 METR randomized control trial found experienced open-source devs were 19% slower with AI tools — while believing they were 24% faster. The gap is organizational: bolting AI onto existing workflows causes a productivity dip (the J-curve) because the workflow hasn’t been redesigned around the tool. Most organizations are stuck at the bottom of the J. The ones breaking through are the ones who redesigned end-to-end: specs, review processes, CI/CD, org structure.
Dan Shapiro’s five levels of AI coding (2026) map where the industry stands:
| Level | Human role | AI role |
|---|---|---|
| L0: Autocomplete | Writes code | Suggests next line |
| L1: Intern | Architects | Discrete tasks |
| L2: Junior dev | Reviews all code | Multi-file changes |
| L3: Manager | Directs, approves PRs | Implements features |
| L4: Product manager | Writes spec, checks outcomes | Everything |
| L5: Dark factory | Writes spec | Spec in → software out |
| L6: Direct execution | States intent | Executes the process — no code needed |
90% of developers who think they’re AI-native are at L2. The vault’s prediction: L6 is next — agents executing business processes directly, not writing software to execute them. See Level 6 — Direct Execution.
This means:
The cyborg model isn’t static. The distribution table above will look different in 6 months. The human’s job is to keep finding the frontier where their judgment, grounding, and physical presence create value that agents can’t yet provide — and to stay ahead of the boundary as it moves.
One of the most significant shifts: the distinction between “technical” and “non-technical” workers — which has organized knowledge worker hiring and compensation for 30 years — is dissolving.
The new divide isn’t technical vs non-technical. It’s people who can clearly articulate intent and evaluate output vs people who can’t. Domain expertise and judgment matter more than the ability to write a for-loop.
Karpathy’s compression of the human role: “You can outsource your thinking, but you can’t outsource your understanding.” The operational form of that — how the human’s understanding gets encoded (Spec), defended (Verifier), and persisted (Environment) — is laid out in The Three-Layer Method. The articulate-intent-and-evaluate side of the divide above is precisely the Spec + Verifier work; the side that can’t is the side that never built the Environment to hold it.
The DJ set pipeline is the cyborg model in miniature. Every role in the distribution table above maps cleanly:
The pipeline: Camelot analyzes tracks → Human writes the set score (track order, transitions, EQ notes, energy arcs) → Mix engine executes the mix from instructions → Light mastering pass → Done.
Who does what:
| Cyborg Role | DJ Application |
|---|---|
| Direction (Human) | Track selection, set philosophy (“texture over harmony”), energy arc design |
| CEO Judgment (Human) | “Break Camelot intentionally here” — creative risk-taking the AI can’t evaluate |
| Grounding (Human) | Knowing what sounds good on a dancefloor, what a crowd responds to |
| Quality/Taste (Human) | Final ear-test — “does this transition actually work?” |
| Pattern Recognition (AI) | Camelot detecting keys, BPM, structural events, beat grids per track |
| Cognitive Execution (AI) | Time-stretching, beat-matching, crossfading, gain staging, rendering |
| Tireless Operation (AI) | Processing 11 tracks × analysis + mixing without fatigue |
The key insight: The human initially assumed they needed a mastering tool (LANDR, Dolby.io). But the DJ Set 1 README already contained all the mixing instructions — it was a program, not a cheat sheet. The right tool was a programmatic mix engine that could execute those instructions. Mastering became trivial because the hard decisions were already encoded.
This is the cyborg model’s leverage equation in action: one human’s creative judgment (track selection, transition design, energy philosophy) multiplied by AI execution (analysis, mixing, rendering) produces a result neither could achieve alone. The human can’t time-stretch 11 tracks to beat-perfect alignment. The AI can’t decide to “break Camelot intentionally because the texture matters more than the harmony.”
See: Set Mastering Pipeline, Programmatic DJ Mixing Tools, DJ Set 1
The cyborg model isn’t just a way of working — the vault itself is a working model of how it works. The structural parallel between this vault and the mammalian brain is precise enough to be explanatory, not just metaphorical.
The mammalian brain has specialized regions: visual cortex processes images, Broca’s area handles language production, the hippocampus forms new memories, the motor cortex controls movement. Each region is optimized for its input type — different architectures for different problems (see Cognitive vs. Motor Skills). The hard problem was never the specialization. It was the coordination: how do these separate systems talk to each other?
The answer turned out to be structured pathways — white matter tracts, the thalamus as a routing hub, the association cortex binding concepts across modalities. The connections between regions are as important as the regions themselves.
The vault has the same architecture:
| Brain Structure | Function | Vault Analogue |
|---|---|---|
| Thalamus | Routing hub — nearly all sensory input passes through it before reaching specialized regions | INDEX.md — the master map everything passes through |
| White matter tracts | Direct connections between specialized regions | Cross-links (**Links:** sections between pages) |
| Association cortex | Binds concepts across modalities — where “seeing red” meets “feeling hot” meets “the word fire” | Tag index files — grouping related pages across domains (a tag like economics connects philosophy, gaming, and business) |
| Hippocampus | Forms new associations, consolidates short-term memory into long-term | The ingest process — raw input is processed, linked, and integrated into the existing knowledge structure |
| Prefrontal cortex | Executive function — what to attend to, what to suppress, what to act on | The human directing the LLM + CLAUDE.md schema |
| Sensory input | Raw, unprocessed signals from the world | raw/ folder — unprocessed source material (articles, transcripts, PDFs) |
| Working memory | Limited capacity (~7 items), temporary, active processing | The LLM’s context window |
| Long-term memory | Persistent, structured, retrievable via associative paths | The vault itself |
The critical insight: the LLM’s context window functions exactly like working memory. It’s limited. It’s temporary. It can only hold so much at once. When the conversation ends or the context fills up, everything in it is gone — unless it was written to the vault.
The vault structure is what makes long-term memory work. Not by cramming everything into the context window simultaneously, but by providing navigable retrieval paths so the LLM can load the right information when it needs it. The brain doesn’t hold all memories in working memory either. It has paths back — and those paths are the cross-links, the index, the tags.
This is exactly why the Karpathy LLM Wiki pattern beats RAG. RAG reconstructs the retrieval path from scratch on every query — like a brain that lost all its white matter tracts and has to re-derive every association from raw sensory data each time. The vault’s pre-built link structure means the associations are already there, maintained, ready to traverse. The LLM reads INDEX.md (thalamus), follows links to the relevant pages (white matter tracts), and synthesizes across them (association cortex). The path exists before the query.
The full cycle mirrors sensory processing → memory consolidation:
raw/ (article, transcript, PDF)/vault-sync updates INDEX.md, tag files, and bidirectional links — strengthening the pathwaysThe human’s role throughout is the prefrontal cortex: deciding what to attend to, what matters, what emphasis to give. The LLM does the encoding, filing, and cross-referencing — the bookkeeping that makes the memory system actually work. This is why wikis maintained by humans alone always die: the maintenance burden (the hippocampal/association work) grows faster than human willingness to do it. The LLM doesn’t get bored.
/vault-heartbeat as the first self-tuning skill, once the right fitness metric emerges. Videos exist on using markdown to define fitness functions for skills — worth ingesting when found.