A 20 YEAR OLD MIT STUDENT JUST PUT HIS AI ROBOT TO WORK IN AN AUSTRALIAN TOUR COMPANY - THEN IT ENCOUNTERED A KANGAROO ATTACKING A DOG.
The student reportedly built the robot with help from Anthropic and decided to test it in the real world instead of a controlled lab.
During a tour, the robot suddenly noticed a kangaroo grabbing a dog and moved toward the situation.
No script. No controlled environment. Just a completely unpredictable real world interaction.
The robot had to perceive what was happening, react to the animals and physically navigate the chaos in real time.
This is what Physical AI looks like when the demo ends and the real world starts.
The crazy part? A 20 year old is already putting an AI humanoid into an actual tourism business.
AI isn't just learning to think anymore. It's learning to operate in the physical world.
WHILE YOU WATCH HER FIX HER STOCKINGS IN THE MERCEDES
You’re sitting there staring.
She doesn’t even register you exist.
She doesn’t give a fuck.
Leather. Light. Stockings. A cabin you’ll never touch even on credit.
You’re already getting off on someone else’s life like a beggar at a shop window.
And lying to yourself that “this is possible.”
5% live like that.
The other 95% is a performance for people like you. The camera cuts - she gets on a tram. Goes home.
No Mercedes. No fantasy you built in your head.
You’re not watching luxury.
You’re watching a set they fed you so you’d keep sitting there, consuming a life you never built.
Most of you will rot like this: empty fridge, someone else’s interior on your phone, and the proud lie that “money isn’t everything.”
Money is everything. Just not for you.
They sold you the market. I shove the receipt in your face.
you handed a task to a subagent and wrote the delegation the way you would brief a colleague
it arrives carrying five things, and your conversation is not one of them
the prompt in its own file, the task as it was summarised,
your claude.md, a git status, and any skills you preloaded
that is the entire briefing
the constraint you established in chat, the table that is wrong,
the thing you said not to touch, all of it stays in your window
a fork inherits the conversation, a subagent does not,
and from the outside the two are indistinguishable
so the number of your spoken rules that crossed over is zero
and the fork in the task gets decided by a worker that has no way to ask which one you meant
the fix is one sentence, not one field, put the constraint inside the delegation next to the task
the briefing is not thin because someone summarised it badly
it is thin because you were never in it
THIS ONE PROMPT TRICK BUILDS A FINISHED WEBSITE IN A SINGLE SHOT
the secret is you stop describing the site and start handing the model a shot list
most people type something like a premium arctic puffer jacket store with a video hero and a clean catalog
the model has to invent the whole structure so it guesses different every single time you run it
instead you write the page as an ordered list of sections and give each one a job
01 hero, full bleed video, nav sitting over it, one product in frame, nothing competing with the subject
02 ticker, single line, specs only, loops forever
03 catalog, 3 columns on desktop, 2 on tablet, 1 on mobile, image first, price last
04 story, one image, one paragraph, done
now the model isn't designing anything, it's filling a structure you already locked in
that's the whole reason one prompt comes back finished instead of almost right
write the shot list first, then take the shot
tomorrow I post the full section by section template I use to one shot any landing page