At 3:55 this afternoon I watched the Anthropic skills talk for the second time.
Not because I missed something. Because I wanted to make sure the thing I thought I heard the first time was actually there.
It was.
Infrastructure has a ceiling. Skills do not.
if this is your first time here, this is a daily newsletter about raising my son in a world that will not require him to work. start anywhere.
A base model is a general-purpose system. It handles almost anything at an acceptable level. Draft an email, summarize a document, write code, translate a sentence. The capability is broad. The intelligence is ambient.
But broad has a ceiling.
The ceiling is not about raw intelligence. It is about specificity. A general-purpose model operates at the level of general problems. When the problem is specific, when it requires knowledge of this client's terminology, this workflow, this database, this particular set of pain points encoded over time, broad is not enough. Broad cannot distinguish between what matters and what does not. Broad does not know the difference between a friction that builds something and a friction that just wastes time.
What changes a general-purpose system into a specialized operator is not upgrading the model. It is attaching the right skills to it.
The skills know the domain. They carry the logic specific to this context. They encode the decisions made over time about what matters, what to ignore, what language to use, what sequence makes sense. The base model provides the general intelligence. The skills provide the specific capacity. Strip the skills away and you have a capable, generic system. Keep them and you have something built for this exact world.
This is not a metaphor. This is architecture.
The insight that locked in at 3:55 was this: the model is not the moat.
Every company building on top of AI right now is racing toward the same base model. The underlying intelligence is becoming commodity. The cloud is commodity. The compute is commodity. Anyone with a browser and an account can access the same foundational capability. The differentiation is not in the model. It is in what gets attached to it: the skills, the context, the domain logic, the specific choices made over time about what this system is for and who it serves.
Infrastructure commoditizes. Attachments do not.
Three days ago I wrote about Lucian and the pocket.
He grabbed a tissue peeking out, looked at it, set it down, and moved on. I put it back slowly, right in front of him. He watched it disappear. Then he lost interest. I wrote that he is testing the world as a retrieval system, and that the question I am watching for is when his relationship with objects flips from where did this come from to where should this go. I said I did not know when that moment arrives. I said I wanted to be there when it does.
I am still watching.
But this afternoon opened a different dimension of the same question.
It is not enough to watch for when the question flips. The more important question is: what have I attached to him before it does?
Lucian will grow up in a world where the base model of human cognitive capability is ambient. It will be infrastructure. Every employer, every institution, every interface will assume access to it. The credential-to-career ramp was built for a world where showing up with general capability was enough: the degree, the skill set, the employment history as proof of generic competence. That ramp is already narrowing. What comes after is a world that does not ask whether you can retrieve information. It already knows you can. Every device assumes it. The question the world will ask instead is what has been attached to that general capability that makes it specific, authored, and real.
For most of the industrial era, parenting was preparation for the base model. Teach the child to read and write. Teach them to follow instructions. Show up on time. Perform the assigned task at an acceptable level. These are general-purpose human capabilities. They map to a labor market that needed general-purpose human labor. The credential system was the verification mechanism: this person has developed general capability to the level that certifies them for this class of work.
That system was not a scam. It was the right architecture for its moment.
The moment is changing. The base model of human capability is no longer what the world is short on. The world is not struggling to find people who can perform general tasks. It is struggling to find people who have had the right things attached to them: the capacity to hold an outcome in mind and move toward it without being told what to do, the ability to distinguish what matters from what does not in a world designed to distract, the presence to be inside a real moment with a real person without outsourcing the experience to a device.
Those are not general-purpose capacities. They are attachments. They are not installed by infrastructure. They are installed by relationship, by repetition, by being with someone who already has them and being shaped by the proximity.
Ten days ago I wrote about the foam ramp I built on the couch. The instinct that fired when Lucian nearly fell was not the result of the ramp. It was the result of eleven months of presence: a nervous system already built, already attuned to his movement, already ready before any plan could be made. The ramp was curriculum. The presence underneath it was what made the curriculum possible.
The ramp is base model. The presence is the skill attached to it.
I keep returning to this framing because it cleans something up I have been circling for months. The question was never what to teach Lucian. It was what to attach to him, before the world offers its own attachments, its own framing, its own defaults about what matters and what to want and what success looks like.
The world is generous with defaults. The school system. The media environment. The algorithm. The peer group. All of them arrive with attachments already installed, engineered for something other than Lucian's specific formation. They are good at reaching the general-purpose child. They are not designed to know this particular child, his specific shape, his specific wiring, his particular way of meeting the world.
Only proximity does that. Only the parent who has been there, on the floor, in the mess, in the ordinary repetition of the same morning, the same song, the same reach for the same pocket, gets to attach something specific.
This is what presence is actually building in a child. Not just that Lucian learns to sort signal from noise. That he develops the capacity through being close to someone who already does. Discernment is not a curriculum item. It is an attachment. It is what happens when a child spends years inside the peripheral vision of a parent who is paying attention.
The Compute constraint makes the same claim from the architecture side: use AI as a tool with a job, not an oracle with authority. Do not surrender authorship to the model. The attachment that matters is the one that keeps authorship with the human. That is not a policy. It is a practice. And it is transmitted by proximity, not by instruction.
Sacred compute discipline, the insistence on remaining the author rather than the machine's interpreter, is itself an attachment. It is installed across years of watching a parent navigate intelligence tools without losing themselves in them.
At 3:55 this afternoon the architecture became clear: the base model is not the moat. The attachment is. In the AI economy, the differentiation has already shifted from the model to what gets built on top of it. Every parent paying attention should notice that the same shift is underway in human development.
In 2044, Lucian will be nineteen. The base model of human intelligence will be ambient infrastructure. What will not be ambient, what will still be specific, earned, and irreplaceable, is what has been attached to him across eighteen years of proximity with people who knew the difference between what matters and what does not.
I am the first skill library Lucian has access to. That is not a responsibility I can delegate. The base model is not the moat. The attachment is. And I am the first one installed.
What I am holding onto:
Three days ago I said I was watching for the moment Lucian's question flips from where did this come from to where should this go. I am still watching. But tonight I am carrying something new alongside it. The recognition that the flip is not automatic. It happens in the children who had the right things attached to them before it arrived. I want to be deliberate about what I am installing, from the floor, before the world sends its own attachments with better marketing.
I remain of the mind that specificity is the only thing that does not commoditize. Not intelligence, not access, not the credential. The specific capacity attached to this specific child over eighteen years of proximity. That is the bet.
The parent is the first attachment. Everything else is infrastructure.
Cheers, JanCarlos
hit reply and tell me one thing you are attaching to your child that the school system cannot. I read every one.