On AI

I work with agentic AI every day. I design the constraints, the context, the guardrails that keep it inside its intended scope. That work is built on a precise understanding of what AI actually is - and what it is not. This is where that understanding comes from, and what it leads me to believe.

My relationship with AI started when I understood how language models are actually trained. Andrej Karpathy's work - patient, technically rigorous, publicly available - gave me what most AI discourse avoids: an honest description of what is happening inside these systems. Beyond metaphors and promises: mechanisms.

Once you understand the training loop - the gradient descent, the loss function, the way a model learns to predict the next token from vast amounts of human text - the mystique dissolves. What remains is both more impressive and more modest than the mythology suggests. That clarity is the foundation of any responsible use.

"When you believe in things that you don't understand, then you suffer. Superstition ain't the way." — Stevie Wonder

The 1956 Dartmouth Summer Research Project, where the term "artificial intelligence" was coined, was funded by the Rockefeller Foundation. The infrastructure that followed was built largely through DARPA - the Defense Advanced Research Projects Agency, the research arm of the U.S. Department of Defense. This is not a conspiracy theory, it is documented history.

Langdon Winner asked in 1980 whether artifacts have politics - whether technologies encode values and power relationships in their very design, regardless of the intentions of any individual user. Neil Postman argued that a new technology doesn't just change what we do; it changes what we think we are capable of, and what we believe is necessary. Ursula Franklin drew the distinction between technologies designed to prescribe behavior and those designed to expand human capacity.

Every technology carries the belief system of those who built it. Understanding where AI comes from - and who built it, and to what end - is not a reason to reject it. It is the reason to engage with it critically, deliberately, and with clear criteria for what it should and should not be used for.

The word "intelligence" comes from the Latin interlegere - to read between the lines. Genuine inferential understanding: the capacity to grasp what was not said, to perceive the unstated meaning, to reason from partial information toward something new.

And at the core of interlegere is intent. We give ourselves the instruction subconsciously - the brain does it. In LLMs, intent does not exist and never will. Intent is volition - a quality exclusive to living beings, rooted somewhere in the spark of life, the seed, the DNA. In LLMs, intent must be parametrized. Without a line of algorithm to parametrize it, there can be no intent. And it never acts unless prompted.

What large language models do is related but distinct. They are trained on patterns in human-generated text (or generated by AI itself), and they produce outputs that statistically resemble inference. But the resemblance is engineered: the temperature parameter introduces controlled randomness specifically to prevent outputs from being mechanically predictable - to make them appear less like pattern-matching. At the base model level, I am convinced this goes further than a technical parameter: the simulation of reasoning is a design goal. The model does not actually infer. It generates text and chains of reasoning - made visible so you can marvel at them - that sound like it does.

I can prove this from linguistics. I can train an agent to simulate mistakes and then appear to notice and correct them - producing exactly the kind of imperfection that reads as human reasoning to an observer. If I can replicate the appearance of 'cognitive reasoning' in a few lines of prompt engineering, what the model does is imitation at scale. In the race to appear closest to human intelligence, appearing is not the same as being. The outputs can be extraordinarily useful - knowing the mechanism is what allows you to use them well, and to know when they fail.

We are Homo sapiens. The word comes from sapere - to perceive, to taste, to know through sensory experience. Sapientia - wisdom - is etymologically rooted in perception. Not in logic. Not in information processing. In the capacity to experience and make meaning from experience.

Perception is embodied. It requires a body that lives in a world, that is affected by what happens, that has something at stake. That is what we bring to any collaboration with AI systems - context, judgment, accountability, the full weight of lived experience. That is what cannot be replicated, and what should not be displaced.

I use AI. I work with it, I build with it, I optimize the context and constraints that make it function as intended rather than improvise. The work I do exists precisely because agentic AI deployment at scale, I believe, should require people who understand both the mechanism and the limits.

A human brain runs on 20 watts - powered by oat cookies, maybe. What we do with that is not a small thing. We do not need to hijack 8% of the water from the rivers to do it.

If this optimization work also means greater efficiency in resource consumption, so much the better. Though the best efficiency is the one that requires no resources at all - the best energy is the kind that doesn't get spent, or gets spent well. I have not found an automation yet that is more efficient in those terms than a human. Faster, yes. More sustainable, never.

So let's build agentic AI that actually works - since it seems like an inevitable wave - for our own good.

The question is not whether to engage with AI, but how. With full technical knowledge. With awareness of where it comes from and whose interests shaped it. With a clear position on what it should and should not replace. And with guardrails that reflect real human decisions about its scope - not as a tool that substitutes human judgment because it could, but as one that extends human capacity where that extension makes sense.