NEXMIND learns while it acts. It cuts its experience into events and plans through stepping stones it discovers itself. No goal list. No hand-written steps.
Click anywhere to give it a goal
Event-centric active inferenceAn N-TEK contribution
Reach the door, open it, cross the room. Long goals are made of steps, and today someone has to write those steps for every task. NEXMIND forms its own.
It cuts the stream of experience where the world changes, and defines each piece by what it did, not by what it looked like.
Pieces with the same effect become one event. Too crude, and an event is added; redundant, and two merge. Nobody writes the list.
From where it is now, it chains events toward what it wants, and hands the next one down to act.
Every event starts with no pull. It earns trust each time achieving it leads somewhere better.
Early results · pre-registered
Every comparison below was registered before its runs.
20/2020/20fresh runs improved
A deep reinforcement learner (PPO) on Craftax-Classic scores higher with EC-AIF’s goals than without, on every fresh seed.
Registered controls show the gain comes from the events themselves, not from extra reward or novelty.
Preliminary report, Table 1
up to8.7×8.7×for standard active inference
The same event layer, added unchanged, on DeepMind’s bsuite catch.
Preliminary report, Table 1
1.9×1.9×for NEXMIND
NEXMIND, N-TEK’s own agent, with the event layer against itself without.
N-TEK internal test · pre-registered
Preliminary report · Zenodo · 2026
DOI 10.5281/zenodo.233780080009-0005-0673-8027
Early results. In active development.
I don’t want machines that only remember. I want machines that are changed by what happens to them, the way we are.
NEXMIND learns into its own parameters as it lives: experience carried in its weights, not parked in a prompt.
NOMAD Mk. II
NEXMIND is built for NOMAD Mk. II, an unmanned helicopter N-TEK designs end to end, from the airframe to the mind.
Flown in simulation with online learning in the loop.
A deterministic safety authority admits or refuses every command. Nothing learned touches the controls.
Every camera, depth and thermal frame is verified before the mind may believe it.
Back the work
I’m İbrahim Alican, an aerospace engineering undergraduate at Istanbul Aydın University. N-TEK is my independent lab. A new form of active inference, a public research report, a flight autonomy stack and an aircraft design: all built alone, on a single desktop, without institutional funding.
The bottleneck. GPU hours, cloud credits or hardware turn straight into bigger models and longer experience.
Grants, sponsorship or investment turn a one-person lab into a team, and NOMAD from CAD into hardware.
Active inference, world models, continual learning: I want sharp eyes on this work.
Roles and internships where this research can grow. I build end to end, from theory to Rust to airframes.