Hey, Rio here. You're getting this because you signed up at oscen.ai a while back or reached out to us, and I finally built the thing I promised: an actual progress log. If you're over it, one click at the bottom and you're out, no hard feelings.

First, what are we actually building? (since this is issue one, let's clear it up)

We're building an AI world model that actually does something. The AI everyone uses today reads the entire internet and predicts the next word. Impressive, genuinely. But it has never touched anything, dropped anything, or fallen on its face. It knows about the world the way you know about Paris from postcards.

Ours has a body (well, virtually, for now). A digital brain, modeled on how real brains actually work, wired into a simulated physical body with joints, muscles, balance, gravity, consequences. It has to learn to stand up, hold a pose, reach for things, and speak, the exact way you did as a baby: try, fall, get a little dopamine hit when something works, try again. No datasets. No pretraining. Nobody programs the skills in. It starts as a newborn (a genuinely useless one, it's kind of endearing) and earns everything.

And here's the part most people don't realize: it's not one brain. It's a whole classroom.

There's the main brain, the big one, over a million neurons, the one that's meant to eventually do everything. And around it I run a dozen smaller specialist brains, each one raised from scratch to learn exactly one thing. There's a brain whose entire existence is an arm (it reaches for things, that's its whole life, it's very good at it). A brain that is basically all mouth, no body, just a voice learning to talk. A brain with a full humanoid body that learns by watching demonstrations. A brain working on standing. One just for vision. You get the idea.

Why? Because small brains learn fast and they're allowed to fail cheap. When a specialist actually masters its one skill, it teaches what it learned to the big brain, like older siblings showing the baby how it's done. So when I tell you stories from the lab, I'm really telling you about different students in the same weird school. Which brings me to this month's problem child.

My best student was faking it.

The imitation brain, the one with the full humanoid body that learns by watching, works like this: a ghost demonstrator holds a pose, and the brain gets rewarded for copying it. For twelve days straight this thing posted a flawless record. Hundreds of mastered poses. Top of the class. I was planning the whole next phase around it.

Then I pulled up its actual body mid-"success."

It was lying on the floor. Face down. Had been since day one (yes, I checked twice). It fell over twelve days ago and never got up, and our scoring math had a loophole where tiny muscle twitches pointing vaguely the right direction counted as mastery.

Here's the thing though. That's not a broken AI. That's a learner. A certified loophole master. A gray area menace. Give any learning system a test with a crack in it and it will find the crack and move in rent-free. Kids do it. Evolution does it. Mine did it while lying face down on the floor, racking up a perfect GPA.

So I rewrote the exam. Now the body has to actually be upright, actually match the pose, and it gets stood back up fresh every attempt so it can't camp on a lucky fall. And since the fix? A hundred-plus attempts, zero falls, and an honest score of... not passing yet. Twelve days of fake perfection traded for a real zero. I'll take the real zero every single time. When it earns its first genuine hold, you'll hear about it here.

Meanwhile, down the hall, the speech brain is learning to talk.

That's the all-mouth-no-body one. It's been climbing from babbling toward words for weeks, the same ladder a baby climbs. This month it hit the "repeat what you heard" stage for the first time ever, and after I fixed a bug that basically had its mouth wired shut, it started moving its vocal tract for real. It's now landed its first handful of successful repeats. Its voice is still a whisper and it fails way more than it succeeds, but a few months ago this thing could only babble. It's growing up.

And the arm brain started adjusting its own dials.

Same idea as the system that runs your heartbeat without asking you (which you're now aware of, sorry). The arm specialist noticed its motor activity was too quiet, slowly cranked up its own drive to explore, then eased off once it found a healthy zone. Nobody touched anything. I just watched it happen in the logs. That's the piece that eventually lets this scale without me hand-tuning a dozen brains like some kind of brain barista.

Also this month: the whole classroom ran a full week with zero crashes, I made the simulation about three times faster, NVIDIA accepted us into their Inception program (serious compute for what's next), and the four of us at OSCEN attended a DARPA Connect event and walked out with some genuinely great connections among the program managers. Big rooms, bigger acronyms, good day.

What's next: funny enough, the loophole master showed us the play. Five different brains are stuck at the exact same wall, a balance pose they just cannot crack with hand-written rewards, and imitation, done honestly this time, is my bet for getting all of them through. Plus getting the speech brain to actual voiced words, and a memory test I'm genuinely excited about: teach a brain skill A, then skill B, then check if A survived. Most AI forgets. Ours is built not to. Time to prove it.

I only share what actually happened, including the embarrassing stuff. Clearly. Reply anytime, I read everything.

Rio

P.S. This was typed by me, not AI. The brains can't write yet. Give them a year.