The mistake that fixed my AI budget
Output tokens cost more than input tokens on every tier, and a chatty model answering a yes or no question is the most expensive habit you can have without noticing. What I changed, and what it actually saved.
DHAHAB Dispatch · free weekly
A weekly letter about building leverage without a budget: AI agent workflows, automation that survives contact with reality, and the honest economics underneath it. Written by a full-time CS and applied math student who pays his own API bill, which is exactly why the numbers in it are real.
Free forever. One email a week. Unsubscribe in one click.
The rules
Already written
These are finished, not planned. Subscribe now and you get them in order, one a week, starting with the first one.
Output tokens cost more than input tokens on every tier, and a chatty model answering a yes or no question is the most expensive habit you can have without noticing. What I changed, and what it actually saved.
One agent grinding a checklist is slower than five agents working five independent problems. How I split work into parallel lanes, and the two kinds of task that must never be parallelised.
Walk-forward testing on my own data killed five of six ideas, including two that looked excellent in-sample. The one that held up returned a modest positive expectancy in simulation, and I explain exactly why that is still not a promise.
Not by working 25 times harder. By building the parts that repeat once and then letting them repeat. A tour of the system, including the parts that were a waste of time.
A small model on my own machine drafts. A bigger one only judges. Where that split saves real money, and the class of work where sending it local quietly costs more than it saves.
My co-authored work on bilingual coding instruction did not have to stop at the conference. How the finding turned into a curriculum for English-learner high schoolers, and what that took.
Automated writing is fast. Unchecked automated writing is a liability with your name on it. The quality gate I run before anything generated is allowed out the door.
The real accounting, not the highlight reel. What made money, what made nothing, and which of my own projects I would not build again.
Who writes it
I am a computer science and applied math undergraduate at San José State, co-author on two SIGCSE 2026 papers in CS education, and an undergraduate researcher in the CSEd lab. I also build the things I write about: a security scanner, a set of offline apps, an outreach automation rail, and a swing screener whose results I publish whether or not they flatter me. The Dispatch is the working notebook, cleaned up once a week. It is free because the audience is the point.
FAQ
Once a week. If there is nothing worth sending in a given week, nothing gets sent. Filler is worse than silence.
Yes, and every issue currently planned is free. If a paid tier is ever added, everything already published stays free and nothing you already receive gets moved behind a paywall.
No. The list is not rented, sold, or shared with anyone. Every issue has a one-click unsubscribe link and unsubscribing takes effect immediately.
No. Trading comes up as one subject among several, and only as a case study in testing ideas honestly. The Dispatch does not publish trade alerts, signals, or price targets, and nothing in it is financial advice.