My Take on Scaling AI: Experience Is the Scaling Layer

Nobody closes the spells gap by reading documentation.

In the first post I wrote about why words carry energy and the unfair advantage of the people who build AI systems. In the second, I laid out the architecture: prompt optimization, knowledge packaging, autonomous feedback loops. This is the last post. And it is the one that sits underneath the other two.

The architecture only works if people put in the time. They close the gap by using the damn thing. Thousands of interactions. Patterns that repeat until they become intuition. Failure cases that no manual can teach.

Wizards are not made by studying wands. They are made by using them, failing with them, and eventually reaching the point where the wand and the hand stop being two separate things.

From fear to access

I started using AI from a position of fear.

Not the dramatic kind. The practical kind. Privacy. Data leakage. Hallucinations. I had watched enough confident AI demos go sideways to know that the gap between impressive and reliable was not small. So I kept my guard up. A tab I opened when I had a specific question. Every interaction controlled.

Privacy: anything I typed was potentially training data. Reliability: hallucinations were not an edge case — the model would confidently produce a wrong answer, and if you did not already know the answer, you would not know it was wrong. That asymmetry is dangerous.

What changed was a compound effect. I started with small things. Each time it worked, I extended the radius of trust by an inch. Those inches accumulated. The caution did not disappear. It got absorbed into the system design. Instead of me being careful about every interaction, I built a harness that was careful on my behalf.

That is the turn. Not less fear. Different architecture.

The harness is the guardrail

The biggest shift happened when I stopped treating guardrails as restrictions I placed on myself and started treating them as infrastructure I could build outside myself.

When most people give an AI system access to their tools, they are using a general-purpose assistant with whatever safety defaults the provider ships. These are extraordinary tools. But they are not harnesses. The guardrails are in the model's instructions, which means they are soft. Persuadable. The model is trying its best, but trying its best is not the same as a deterministic fence.

What changed for me was encountering systems built around the idea that the harness is the product. The agent runtime where I am writing this operates on the principle that the deterministic layer around the model is where the real safety lives. Not in the model's instruction-following. In actual code that defines what actions the model can take regardless of what it is instructed. The harness says you cannot do this in code, not in a prompt.

That distinction made the access decision rational instead of emotional.

When you start using a new harness, you cannot set it up in one day. I read that from experienced people before I understood it. Only through incremental usage, through hundreds of experiments with things you do not mind getting wrong, does the harness start to reflect who you actually are. Use it, find an edge case, fix it, use it more. Both sides learn. Your trust compounds through evidence.

It is not a setup. It is a relationship that matures with exposure. And that maturation is the training of spells.

Finding the Spells

A spell is not a prompt. A prompt is a sentence. A spell is the intuition for which sentence opens which door.

I know exactly which words in our AI enterprise stack trigger which tools. I wrote the tool descriptions. I named the skills. When I say send a message the model might go to Slack, might go to email, might go to a Jira comment, depending on exactly which word I add around it. I know which word that is. I know it not because I read documentation but because I failed enough times to see the pattern. The failure is the curriculum. There is no other way.

A wizard without training holds a wand that might fire anywhere. A wizard with years of practice holds the same wand and hits exactly what they aimed at. The wand is the harness. The spells are the skills and the prompts. And experience is what transforms a wand-holder into a wizard.

The gap between an AI builder and an AI user is not a knowledge gap. It is an experience gap. That model of how the system thinks is not written down anywhere. It cannot be. It lives in the behavior, in the instinct for when to add a hint and when to let the model figure it out.

What this means practically is that any strategy for scaling AI to two hundred people lives or dies on one variable: how much time-in-system each person accumulates, and how fast they accumulate it. Everything else is infrastructure that only delivers its value if people put in the time. And they will only put in the time if putting in the time does not feel like work.

The keyboard as a gate

For most of my career, working meant sitting.

A keyboard. A monitor. A specific chair in a specific room. The interface to my own work was physical and located. The keyboard was not just an input device. It was the gate between me and my output.

The real crack came from the phone. The first time I understood how much I could do from a simple chat interface on my mobile screen — not a terminal, not a desktop app, just a messaging conversation — I was disoriented. I could check the state of my home systems. I could review my calendar and my finances. Draft meeting agendas. Start development workflows. Review pull requests. From my couch, from my car, from a walk.

That was the first unlock. Mobile chat. Eyes and hands required, but no longer location-dependent.

The problem with mobile chat was that it kept the keyboard. And the keyboard, whether physical or on glass, imposes a cost. Typing requires your eyes. You stop. You edit yourself. That cost compounds. If reaching for the system requires stopping, you reach for it less. Reduced interaction volume is the enemy of the experience accumulation that closes the spells gap.

Voice removes that cost. Not by being faster than typing — it is not. But by removing the activation energy entirely. When I can talk while walking, while driving, while making coffee, the barrier goes from sit down and type a prompt to say what you are thinking.

A hundred rough voice interactions teach you more about how the system thinks than ten polished typed prompts.

The closed-system problem

The voice experiences that were available before I built my own were extraordinary and useless at the same time.

OpenAI's GPT-Live is genuinely impressive. Full-duplex, low-latency speech-to-speech, numbers that would have been impossible five years ago. xAI's Grok powers the voice layer inside Tesla vehicles. I used these. I was baffled by them.

And they were useless to me.

Closed systems. The best voice conversation I had with GPT-Live happened with a model that had no access to my tools, my context, my skills, or my memory. Conversational experience: world-class. Harness integration: zero.

On the other side, my harness had all the power. Full tool access. Persistent memory. Skill layers. Deterministic guardrails. But no voice interface. To reach it, I needed to type.

The trigger was an open-source project from Hugging Face: speech-to-speech on GitHub. A real-time pipeline combining speech-to-text, a language model backend, and text-to-speech. The backend flexibility was the unlock — point it at a local model on your own hardware, the entire pipeline on your infrastructure, no API minutes, no external dependency.

I took that stack and built the voice layer for my harness. The same tools, skills, memory, and guardrails I had through typed chat, now accessible through natural speech.

The first real test was writing this post. Not at a desk. Not at a keyboard. Walking through my morning, thinking out loud, the system capturing as I spoke.

Not a demo. A working artifact, produced through the stack I built.

What we are leaving behind

A year into this, I still do not have a clean answer for someone who is where I was at the start. Afraid of access. Skeptical of voice. Still tied to their keyboard.

What I keep coming back to is what we are actually leaving behind.

The idea that AI is something you sit down to use.

That is the gate. Not a technical gate. A conceptual one. Tools are things you pick up and put down. Infrastructure is always on, always there, the thing you build your workflow around. The desk was not a physical constraint on how I worked. It was a conceptual constraint.

Voice is not just a new interface. It is the thing that makes the mental model shift possible at scale. You cannot maintain the tool I pick up and put down frame when the tool is always listening, always ready, accessible through the same mechanism you use to talk to a colleague in the next room. The frame dissolves on its own when the friction is low enough.

The spells gap closes one interaction at a time. There is no shortcut. But there is a faster path. Remove everything that makes each interaction harder. Build the infrastructure that makes each interaction useful. And trust that the experience will compound.

The series opened with words carrying energy. It ends with words carrying everything. Not just the spells. The experience that teaches them.

Onwards, wizards.