My Coding Agent Shouldn't Have a Boss
Claude Code is one of the best products I use. That is exactly why realizing how little of it I controlled became uncomfortable.
Read essay
Essays
Decisions, mistakes and changing beliefs about work, AI and building companies.
By George Pu
Founder of SimpleDirect, an independent Canadian AI lab.
Search and filters
22 matching essays in own-your-tech. Clear filters
30 Days, Six Products, Two Dead Models, and the Part That Didn't Work
On July 9 we shipped a model my engineering lead trained. On July 12 I took it off the product. Six products in 30 days is the headline. That decision is the story.
Read essay
What a 3-Person Team Actually Pays for AI: July 2026
In the first week of July I spent over $5,000 coding through one company's API and couldn't tell you what I got for it. Here's what we actually pay now - the receipts, the line that shouldn't have existed, and the one number I'm holding back.
Read essay
The Day I killed our product strategy
Today (on a Sunday) I took our own AI models off our own app and replaced them with a competitor's model. This is the story of how the decision happened in 24 hours, what it was like in the 4 days leading up to it, what it cost me to see it, and what I think other founders can take from it. Day 1 - my own product told me the truth We're an AI model company. Or we were. We spent months training small open-weight models for honesty - models that say 'I don't know' instead of inventing answers
Read essay
I Got Bored of Being a Commentator
I got good at having AI takes - and one day realized I'd become a commentator, not a builder. A complaint is just a wish with better production values. So I built the replacement.
Read essay
Why We Started With Character
Capability is getting cheaper every month. What's scarce is a model you actually want to talk to. Why we bet on character - and started with a small, open, honest 4B model you can run yourself.
Read essay
What I Got Wrong About Canadian-Specialized AI
For six months I was sure the answer was a from-scratch Canadian AI model. Seven days ago that broke. I had confused the model with the system - and the open-book version is the bigger bet.
Read essay
We Are No Longer Building a Canadian Legal AI Model
Yesterday I published the post-mortem: we asked flash-1-mini ten questions any Canadian lawyer would consider basic, and it invented seven citations. That post was about what broke. This one is about what it changed. Because two weeks ago I was telling people we were building a Canadian legal AI model - and today we decided we're not. I want to walk through why, because the answer changed how I think about what "building AI" actually means. The original scope The original scope made sense
Read essay
We Asked Our 4B Legal AI 10 Questions. It Invented 7 Cases.
An honest post-mortem on our own model — and what anyone building or buying AI should take from it.
Read essay
Why CBLRE Matters More Than the Model
Yesterday we released CBLRE — the Canadian Bilingual Legal and Regulatory Evaluation. Methodology and status clarification — October 11, 2026: This June account preserves the early proposal. The public CBLRE harness and methodology at revision b2d09fa describe a development-stage instrument with ongoing subject-matter validation. The item bank, gold answers and held-out scoring split are gated; they were unavailable for this editorial review. Current scoring includes programmatic methods and
Read essay
Closed Orchestrators Will Commoditize. Open Ones Will Compound.
Two AI infrastructure announcements landed in the same week. Ours was a model. We released flash-1-mini — a 4-billion-parameter bilingual Canadian legal AI — under Apache 2.0, alongside an open benchmark (CBLRE), an open training corpus, and the methodology behind both. The model runs on a MacBook. It runs on a Raspberry Pi. The weights are downloadable. You own the file. Perplexity's was an orchestrator. At Intel's Computex keynote in Taipei, CEO Aravind Srinivas demonstrated what the com
Read essay
About $50 in GPU Compute for Our Fine-Tuning Run
Evidence and cost-scope clarification — October 11, 2026: This essay was originally titled “Fine-tuning your own AI doesn’t cost $35,000. It cost us about $50.” The amounts and accuracy figures below are the author’s reported estimates and results for this run. An itemized compute invoice, model and dataset versions, evaluation questions and scoring rules, sample sizes, and reproducible run artifacts are unavailable here. The roughly $50 figure covers GPU compute, not total project cost. Pro
Read essay