A 22 YEAR OLD MIAMI KID FOUND A GALAXY PROJECTOR MAKING $1,816 A DAY ON TIKTOK, BUILT A SHOPIFY STORE WITH CLAUDE IN ONE AFTERNOON, AND CLEARED $31,247 IN HIS FIRST 9 WEEKS WITHOUT TOUCHING A SINGLE PACKAGE
carlos is 22, miami beach apartment, opened https://t.co/nP8AmqDYVK and sorted 10,000 products by revenue
galaxy projector 2-in-1 starry sky, $22.98 to $35.97, $1,816.52 average daily revenue, updated 18 minutes ago
he didn't guess what to sell
the data told him
pause at 0:10 on the revenue chart, the spike hits $4,500 in a single day, that is one product, one tiktok trend, one supplier in china shipping directly to the customer
claude wrote every product description
claude built the ad copy
claude answered every customer email
claude told him what to fix every sunday
carlos shipped zero packages
touched zero inventory
hired zero employees
total cost to run the store: $21/month
week 1: $2,847
week 4: $7,391
week 9: $31,247
the $3,000/month team that used to do this work got replaced by a $21 subscription and one afternoon of setup
the product ships itself
claude runs the store
carlos checks the dashboard
the window is open, follow and bookmark before it closes
@makodama switching tools is easy. switching how you think about the tool is what actually moved his numbers.
curious — what made you finally read about codex?
A 27 YEAR OLD SHANGHAI DEV DELETED CHATGPT FROM HIS COMPUTER IN FEBRUARY, REPLACED IT WITH CODEX, AND SHIPPED 11 PRODUCTS IN 4 MONTHS THAT CHATGPT COULDN'T FINISH IN 14 MONTHS
liang is 27, pudong apartment with sticky notes covering every monitor bezel, paid for chatgpt pro for 16 months and shipped 3 products, switched to codex in february and shipped 11 in the next 4
the math alone should end the argument
but the math isn't even the interesting part
chatgpt is built to have a conversation about code
codex is built to be responsible for code
those are not the same product
chatgpt writes a function and stops
codex writes the function, checks what it breaks, fixes what it broke, runs the tests, catches the edge cases, and tells you it's done
chatgpt loses the thread after 3 files
codex holds the entire codebase and knows what changes when you touch one line three folders away
chatgpt asks you what you meant
codex makes a decision and keeps moving
liang gave both tools the same project in february
full stack dashboard, user auth, stripe integration, data visualization layer
chatgpt produced 60% and stalled on the auth logic
he waited two days trying different prompts
nothing worked
codex finished the entire project in one session
ran its own tests
caught a bug in the stripe webhook handler before liang noticed it existed
fixed it
pushed a clean build
liang shipped that product 9 days after starting
the chatgpt version is still sitting half-finished on his desktop
before codex: 3 products in 16 months, $840/month total revenue
after codex: 11 products in 4 months, $7,340/month and climbing
he didn't get smarter
he didn't work harder
he didn't hire anyone
he just stopped using the tool that talks about building things
and started using the tool that builds them
most developers will read this and open chatgpt anyway because it's familiar
the ones who switch this week will be 11 products ahead by december
the window is open, follow and bookmark before it closes
@andreysuperior $459/month in subscriptions vs $9/month in electricity. the math on this has been obvious for a while. glad someone finally built the proof
An AI Virtual Creator Made $24,000 in One Month.
I saw this today and couldn't believe how simple the setup has become.
A 24-year-old developer from Miami created an AI influencer named Luna.
In her first month online, she reportedly generated $24,000.
The wild part?
Luna doesn't exist.
She has over 1,100 paying subscribers, and one fan spent $1,620 in a single month.
She never records videos.
No human replies to DMs.
Every conversation is generated by Claude.
Every image comes from Flux.
Her voice is powered by ElevenLabs.
The entire character lives inside four Markdown files on a MacBook.
Here's how it's organized:
persona.md → backstory, personality, interests, boundaries
voice.md → writing style and tone
flux.md → appearance, poses, and image consistency
memory.md → remembers previous conversations with every subscriber
Claude loads all four files before every response, so Luna stays perfectly consistent.
A year ago, building something like this would've taken months.
Today, one person can put together the entire workflow in just a few weeks.
AI has made creating digital personalities dramatically easier.
Now the biggest challenge isn't the technology.
It's creating a character people actually want to follow.
@AnatoliKopadze senior engineers with 10 years of experience are failing at this while 19-year-olds with claude and a weekend are shipping. that's what this talk is actually about
A 24 YEAR OLD BEIJING DEV RAN 47 BENCHMARKS COMPARING CODEX AND CHATGPT, CODEX WON 41 OF THEM, AND HE SAYS MOST DEVELOPERS ARE PAYING FOR THE WRONG TOOL WITHOUT KNOWING IT
wei is 24, zhongguancun apartment with three monitors and a whiteboard covered in benchmark results, spent 6 months running codex and chatgpt on the same tasks back to back and stopped switching back in february
the number that ended the debate for him:
same 47 tasks, three attempts each
codex solved 41 correctly on the first try
chatgpt solved 29
that is not a marginal difference
that is the difference between a tool that ships and a tool that needs a second prompt to understand the first one
the comparison nobody wants to make out loud:
chatgpt was built to talk about everything including code
codex was built to think in code the way a senior engineer thinks in code, not as text, as logic, as structure, as consequence
code completion accuracy: codex
multi-file context across a full codebase: codex
debugging without losing track of what changed: codex
writing tests that actually catch the edge cases: codex
following a complex spec without drifting: codex
wei's output doubled in the first week after switching
not because codex is magic
because it stops asking what you mean and starts solving what you asked
chatgpt writes the function
codex understands why the function exists, where it fits, and what breaks three files away if you change it
that context awareness is the entire game in production engineering
the reason most developers still use chatgpt for code is the same reason most people use the wrong tool for everything
it was the first one they found
and switching feels like starting over
wei started over 6 months ago
his last project shipped in 11 days
the same project took him 38 days the year before
the developers who make this switch in the next 60 days will compound that advantage every week
the ones who wait will still be re-prompting chatgpt on step 3 and wondering why the agent keeps breaking
the window is open, follow and bookmark before it closes
@RoundtableSpace karpathy just turned obsidian from a note graveyard into something that actually thinks. this is the setup i've been waiting for someone to publish