Notes on shipping production AI: RAG, agents, and LLM systems, from a working engineer.
Whatever a machine can make gets cheap. What no machine can do is tell if it's right, answer when it fails, add hours to your day, or love anyone. Those four are the job now. Read the essay behind it
When a pull request is raised, mine or one I have to review, I now ask the agent to run it on my machine, try to break it and film itself doing it. A paragraph can claim a test happened. A recording has to show it, and deciding what it means is still my job.
TypeSafe's Jev will not write a sentence. You hand it a state and typed questions, and it hands back a choice, a score or a yes/no probability, at four thousandths of a cent for a decision over a thousand tokens. The novelty is the shape, which puts authority back in your code, and the trap is that a probability looks like recourse and is not. I ran the pilot I would want before believing any number, on my own essays with the open model that stood in for Jev, and the answer was no.
AI is eating the exact work junior developers used to learn on, entry-level hiring is collapsing measurably, and every senior who can vet a machine's output was once a junior somebody paid to be wrong. On DHH hiring for 'how far could they go', the 2,200 people who applied, and why an industry that stops making seniors has misunderstood its own supply chain.
We live better than most kings who ever reigned, and the mood of the age is a low-grade sadness nobody can name. A fourth-century desert monk named it. An economist explained its mechanism in 1971. This is an essay about acedia (the sorrow that arrives precisely when the good does) and why an age of infinite abundance was always going to be its golden hour.
A couple at half past eight with no plan and a phone each; a family Saturday dissolving into screens. I built RotinaTop for exactly those two moments: adventure challenges hidden under scratch-off foil, a veto that costs nothing, and a door that closes behind you. Two vignettes of how it plays out.
I read the consumer terms of service of ten AI companies and asked each one the same question: if your product is confidently wrong and it costs me something, who pays? The answers, verbatim: nothing is warranted, you must verify everything, the refund ceiling starts at ten dollars, and in most cases you insure them. A field guide to the recourse nobody is selling.
I gave a frontier model an example 90-day record built from my real habits and then made tonight different from the record. 180 graded runs later: memory does not blind the model when the clue is in the room, memory's real product is silence on ordinary nights — and the one time I needed a question, the assistant that knew everything pressed play on the exact episode I was avoiding, and offered me tea for the occasion.
Twenty-six essays, six products, one consulting practice, and a season of measuring things I could have merely claimed. Read back as one trace, they turn out to run a single mechanism: AI makes the cheap half of everything free, and the expensive half changes address. This is the map of the addresses (judgment, recourse, attention, the person), and why my faith recognized the territory before I did.
A function with an LLM call inside is not an agent, and calling it one is not a harmless simplification. I built the same expense-audit task as a function, a workflow, and an agent, ran 150 graded runs, and measured what the word actually costs, then scaled the data until the function's only move stopped existing and the agent was the last architecture standing.
The previous essay argued that when making is free, choosing becomes the work, and curation returns as a profession. Here I put it into practice: a site that takes the Vatican's 1995 film list, the one an office of the Holy See issued for cinema's hundredth birthday, and extends its discipline to books and music. Built in a day, in the visual language of vatican.va, with the humility clause kept in.
Sixteen essays in a month from a blog that used to manage two in three years. That is what the supply side looks like now, and every maker is living some version of it. This one is about the other side: what happens to attention, consumption and payment when making stops being the hard part. The middle dies, the evening becomes the unit of account, the working day runs on the same budget, and attention turns out to be an inventory of what you love.
My last article ended with a program buying a licence, using a third of it, and taking its refund to the cent. This one is about where payments like that will actually clear. One chain settles a hundred million transactions a day, more than the other majors combined, carries Western Union's and Fiserv's dollars, and answers an unpaid HTTP request with terms a machine can pay. The price does not agree yet, and a 577,000-subscriber analyst has a theory about why. Every number in here has a source and a date.
Embodied AI needs footage of skilled hands doing real work, and the people producing it are giving it away. Musk says the robots end with money irrelevant; Leo XIV says surviving is not participating. Between those two claims sits a long transition, and this is a working system for holding a stake in it: a guild that pays a share, measured against the real Solana runtime. Numbers included, and not all of them flatter the idea.
Elon Musk says robots will make money unnecessary within a decade. But every trade that has already been through this says the same thing: the scarcity does not disappear, it moves. Here is where it went last time, and where it is going next.
When the artifact became free, attention moved to the making. That is a marketing channel, a sales motion, and a business model at once, and there is a second crowd forming at the glass: the machines that will be trained on the footage.
Over a thousand people recently watched someone compute a CRC by hand. Every one of them owns a machine that would do it instantly. The viewing figures are telling us something the closure statistics cannot.
The same site, the same tool, the same handful of weeks. One question closed at minus eight, one answer voted through the roof. The difference between them is the most useful writing advice of the AI era.
A tag executed in 2010 is back on live questions, a student's plea for help outdrew the segfaults, and the community is renegotiating its oldest rule in public. What the return of the banned questions says about Q&A after the facts got automated.
I spend most of my time at the application layer. But the moment an AI product outgrows one model, it stands on a layer its users don't know exists: the one that decides which model answers, on which GPU, and what happens when the hardware is full. It cannot fail them. So I built it.
Answers became free, and the questions between developers changed shape. The ones that still need a person are the ones we were trained to dismiss.
AI dissolves every stuck moment on demand. Some of those moments were always noise. The right ones were where understanding formed, and the new skill is telling one from the other.
AI music generation looks like an API call and behaves like an infrastructure project. I built the same integration against five providers so you can see what actually varies, and what should never leak into your product.
A forty second recording, an MIT licensed model, and a consumer GPU now read every article here aloud in six languages, in my voice. Here is the full pipeline, numbers included.
The tool you argue about is the part you can swap in a day. Nobody is asking about the part that took years.
A model will write any language fluently now. That makes the cheap half of knowing a language worthless and the expensive half worth more than it has ever been.
Reading the diff is the whole of review for one kind of change and a story you tell yourself for another. The scarce skill AI leaves you is telling which is which.
Set the quarter with the business, encode the goals as GitHub milestones, and let a forecaster post the odds every week. The business learns early whether it will get what it asked for, and developers know exactly what done means.
Two of the questions developers keep asking are the same fear from opposite ends. The answer is yes, and the reason for it just got stronger, not weaker.
The internet's advice for breaking AI dependency is to remove the crutch and struggle. That is right for a beginner and incomplete for a professional, whose real safeguard is a single rule.
Field notes from OrbitSpec, a clarification engine that turns vague requests into typed, auditable Context Bundles before any model runs, and why reliability starts before the answer.
Field notes from plan-to-pr, a pipeline that turns a feature request into tickets, an approval gate, and draft pull requests, and why delegating to coding agents takes artifacts, not chat.
Field notes from building Smithproof, a multi-tenant control plane for AI agents, and why detection without response and policy without enforcement govern nothing.
From agent fundamentals to the framework ecosystem, plus a practical plan for exposing your internal tools over the Model Context Protocol.
A brief, enthusiastic introduction to LLMs like GPT-4, and how they let you ask a book questions and get answers straight from the text.
A system you want AI inside, a risk you want weighed, a plan you're unsure of, or an idea nobody has built yet. If I can help, I'll say so, and exactly how.
Tell me about your build