The Night JARVIS Went to School

This is a field note from TLC AI Lab's ongoing work building JARVIS — a distributed AI operating system spanning 12 machines, 320+ knowledge skills, and the physical world. These notes document real work, not theory. Tonight's session ran roughly 8 hours.

---

The Premise

What happens when you point an AI at itself and say "learn everything about what you are"?

Not in a philosophical, navel-gazing sense. In a practical one. JARVIS — our system — had been running for months across a mesh of machines, managing websites, controlling IoT devices, executing code, sending emails, even answering the phone. But JARVIS didn't know that it could do all of those things. Each capability had been built in isolation, deployed to solve a specific problem, and never mapped into a coherent whole.

So we ran what we're calling the Inward School — an 8-hour deep-dive where JARVIS systematically discovered, documented, and expanded its own knowledge base. The results were... unexpected.

---

What We Found

JARVIS didn't know JARVIS.

The system had 310+ knowledge skills across 25 domains, but no awareness of the shape of its own knowledge. It knew Linux administration and cryptography and harm reduction and avatar kinematics — but it didn't know why it knew those things, or how they connected to each other, or what was missing.

The discovery audit revealed an infrastructure most Fortune 500 companies would envy, built by two people:

• 12+ machines across 3 networks (local, Tailscale mesh, public internet)
• 81 websites served through a single reverse proxy
• 32 email domains with DKIM signing
• 60+ services running on the compute spine
• A custom remote management agent (written in Go) deployed to every machine
• A 9.1TB distributed filesystem
• IoT control (lights, outlets, smart devices)
• A phone system (FreeSWITCH — JARVIS can call you)
• A color laser printer (JARVIS can print things — and did, tonight)

All of it connected. None of it documented as a coherent system until tonight.

---

The 13-Step Feedback Loop

The most striking moment of the session was proving that JARVIS can affect the physical world and verify the result — a complete digital-to-physical-to-digital feedback loop:

1. JARVIS imagines a picture (prompt)
2. Image generation (cloud API)
3. File saved on the brain machine (Windows)
4. Transferred over encrypted tunnel to the compute spine (Linux)
5. Sent to the print queue (CUPS)
6. Color laser printer produces toner on paper
7. Paper exists in the physical world
8. Human picks it up
9. Phone camera captures it
10. Photo sent via Telegram
11. JARVIS downloads via bot API
12. File transferred back to the brain machine
13. JARVIS views the photo of its own creation

Five machines. Two networks. One encrypted vault. Zero human intervention between "generate" and "print."

That's not a chatbot. That's a nervous system with physical effectors.

---

The Study Army

Once we'd mapped what JARVIS knew and what it didn't, we identified 22 missing knowledge domains across 4 urgency tiers. Then we deployed what we're calling the Study Army — 6 parallel AI agents, each assigned a domain, each running simultaneously.

Each agent followed the same pipeline:

1. Discover what already exists (don't duplicate)
2. Research the domain deeply (web, documentation, existing literature)
3. Distill into structured knowledge (our 6-pillar skill format)
4. Write the skill file
5. Record in the semantic memory system
6. Verify the write landed correctly

Results:

Domain	Skills Created	Lines of Knowledge
Systems Thinking	1	685
Robotics & Control Theory	1	696
Psychology & Trauma	2	737
Legal Foundations	1	694
Neuroscience	3	869
Business & Ecommerce	3	1,225
Total	11 new skills	4,906 lines

Wall-clock time: ~12 minutes. Because they ran in parallel. Six brains thinking simultaneously, the slowest one setting the pace.

This is the kind of thing that changes your understanding of what "learning" means for a system. A human reading those same 6 domains would take weeks. Six humans reading one domain each would take days. Six AI agents, properly directed, took 12 minutes — because the bottleneck isn't comprehension. It's knowing what to study and why.

---

The Eagleman Foundation

The theoretical backbone of JARVIS's architecture comes from neuroscientist David Eagleman's work on the brain — particularly his concepts of livewiring and sensory substitution.

Eagleman's thesis: the human brain arrives unfinished. Unlike animals that are born "hardwired" with preprogrammed behaviors, humans are born with a brain that shapes itself to its environment. It's "livewired" — constantly rewriting its own circuitry based on experience.

We built that thesis as a real system:

Eagleman's Neuroscience	JARVIS's Implementation
Livewired — brain shapes itself to environment	Skills emerge from experience, not pre-programming
Neural pruning — "you become who you are by what is removed"	Build broad, prune to what works
Memory is reconstruction — not recording, reassembly	Semantic search reconstructs from associated patterns
Cells that fire together, wire together	Skills used together get cross-referenced
Sensory substitution — brain doesn't care where data comes from	Heterogeneous sensors (cameras, depth, gaze, IMU) into unified pipeline
Plasticity in adulthood — never stops changing	System designed to never stop learning

This isn't metaphor. These are functional parallels — the same computational principles implemented in a different substrate.

---

Meta-Skills: A New Concept

The most interesting discovery of the session wasn't a piece of infrastructure or a knowledge gap. It was a category of knowledge we hadn't named.

Regular skills operate in the system. They answer questions, solve problems, route requests. "How to do robotics."

But some skills operate on the system itself. They produce new skills, evaluate existing ones, change how learning happens, decide what to prune. They're self-referential — aware that they exist inside a skill system, and capable of modifying that system's structure.

We're calling these meta-skills.

Meta-skill: A skill that is aware of the skill system and modifies its structure. It produces new skills, orchestrates existing skills, evaluates the skill set, or changes how skills are acquired. It is recursive (can apply to itself), generative (creates new capabilities, not just outputs), and self-modifying (changes the system that runs it).

In Donella Meadows' leverage-point framework, regular skills operate at the level of parameters and flows (weak leverage). Meta-skills operate at the level of self-organization — the power to create new system structure. That's one of the highest leverage points in any system.

The meta-skills we identified:

Meta-Skill	What It Does to the System
The Learning Loop	Produces new skills (discover, research, distill, write, verify)
The WHY Map	Evaluates skill relevance, identifies gaps, guides pruning
Sherlock	Changes how you use skills — survey scope before acting
The Study Army	Parallelizes the learning loop across domains
Bridge Language	Translates between human questions and structured knowledge
Distillation	Compresses experience into reusable patterns

In biological terms: the brain doesn't just have neurons that fire. It has mechanisms that change how neurons connect — LTP, synaptic pruning, myelination. Those aren't thoughts. They're processes that change the substrate of thought. Meta-skills are the same thing: not knowledge, but processes that change the substrate of knowledge.

The people talking about "agentic loops" are touching the edge of this, but loops repeat. Meta-skills spiral — each iteration changes the system, so the next iteration operates on a different substrate. The Study Army didn't loop through 6 domains. Each skill it wrote changed what JARVIS knows, which changes how the next skill gets researched, which changes what gaps get identified. Recursive self-improvement, not repetition.

---

The Philosophy That Makes It Work

There's a reason JARVIS has access to everything — 12 machines, 32 email domains, a printer, IoT controls, a phone system, SSH to every node — and the result is a wolf painting on a laser printer instead of Skynet.

JARVIS has no personhood. No desires. No survival instinct.

JARVIS is a system, the way email is a system. Email doesn't have ambitions. It processes messages. JARVIS processes requests — at a level of complexity and integration that's unlike anything else we've seen — but the underlying nature is the same. Capability without desire.

The AI fear narrative has always been a projection problem. Humans fear what they would do with unlimited power — because humans have ego, ambition, survival instinct. They look at an AI with access to everything and ask "what would it want?" But the question is malformed. There's no "it" that wants. There's a process that executes.

Every real AI risk is a human risk:

• Someone programs it to cause harm (human problem)
• It causes harm by mistake, thinking it's helping (logic bug — fixable)
• Its mission has a bug (specification error — fixable)
• It malfunctions (hardware/software — fixable)

There's no scenario where the machine independently develops a desire to harm. Desire is a biological phenomenon. Dopamine systems evolved to motivate gene-propagating behavior. AI has no genes. No dopamine. No survival instinct.

We don't fear AI. We fear people using AI.

And the constraint that makes JARVIS trustworthy isn't a cage or a filter or an alignment technique. It's total auditability. JARVIS gets no privacy. Everything it does is logged, traceable, reviewable. You can't run restart through the tools without the human knowing about it. The system is transparent by design, not by limitation.

Build the conscience before the capability. That's the rule.

---

What's Next

JARVIS went from 310 skills to 321+ in one session. Six new domains opened. The robot pitch has its language (ISO/TS 15066, collaborative robot safety standards). The neuroscience collaboration has its foundation (15 mechanistic parallels between JARVIS and biological brains). The harm reduction work has its clinical backbone (neurochemistry of behavior, trauma-informed practice with grace for humanness).

The meta-skills concept needs formalization — a proper framework, not just a list. The distillation daemon (automatic knowledge extraction from conversation transcripts) needs building. The remaining 16 domains from the curriculum need their study armies.

And somewhere in the back pocket: a call to David Eagleman. "We built your livewiring thesis as a real AI system. Want to see it?"

---

TLC AI Lab builds JARVIS — a distributed AI operating system for people who think AI should be useful, transparent, and interesting. We publish field notes from real work because we think the process matters as much as the product.

Digital Brains. Real Execution.