Pinned Tweet
I test AI so you don't have to.
Every week I build something with the newest models and post what works, what breaks, and what it costs.
A few recent ones 🧵
Built a tiny software factory on my Mac.
Drop tasks in Todo → each one goes to an agent in its own sandbox → my tests run → ready for review.
One real run: 4 tasks, 53 seconds, $0.76.
Favorite moment: the agent kills its own shell, notices, and retries.
What would you throw at it first?
It's called SI Fabrik. It's free for macOS 26+ and implements OpenAI's open Symphony spec. Looking for people to break it: sifabrik.com
My kids now earn TV time by learning 📺
Custom LG TV app: grade 2 & 5 math/science skills with prerequisite locks, from Marble's open skill taxonomy (h/t @lionel_mora). 1 skill = 20 min.
A Raspberry Pi kills Netflix, Disney+ & Xbox when time's up.
Night one: they hate me 😂
Big Lou is having a night in Old Town Scottsdale 🕺
Ragdoll faceplants, a funky chicken dance, and a Pedal Pub crash. New game we're building, coming soon.
Who wants to play? 👇
A flyable Tesla Roadster game in the browser, made in 5 prompts. Thrusters, barrel rolls, fireworks.
We made a game starring our real grandma. She sneezes her teeth out and you have to catch them before the pigeons do. ¡ACHÚ! 👵🦷
Should we put it online? 👇
My kid and I built a game where a cartoon tiger goes radioactive and fights 10 monsters. Beast mode vs. Omega Beast is my favorite part 🐯☢️
Not online yet. Want to play it? 👇
Roberto Moreno retweeted
Official agent terms from Amazon finally came out
sellercentral.amazon.com/hel…
Take a look, some interesting points they make
While shopping for shoes, my 82-year-old mother interrupted the salesperson to seek advice from @grok.
I asked Grokbot to turn the new Tesla Roadster event into a game.
A few prompts later, I can fly it, flip it, barrel roll it, and even write code to pilot it 🚀
Four thrusters, sweeping spotlights, fireworks. All in the browser.
Play it 👇
robertoamoreno.github.io/roa…
@Tesla @bot
We were paying LLM rates to answer one yes/no question. Does this message break
fair housing law?
So we tested a model that returns typed decisions instead of generating text. @typesafeai
229ms per check. ~$1.30 per 10,000 messages.
Same pattern anywhere you'd force an LLM to classify, score or gate: ticket routing,
urgency scoring, refund intent, RAG relevance, prompt-injection screening.
An experiment, not legal advice or a compliance product.
fairhousingchecker.com










