Memory that knows where it came from.
Its own actions, what it watched others do, and its failed predictions are kept apart, in order, with their origin.
Replay adds learning. Never new evidence.
N-TEK · Independent AI research
One mind. Every world.
NEXMIND is intelligence that learns while it acts: one mind across games, a full autonomy stack for flight, and an aircraft of its own.
INEXMIND General
Three worlds. One mind.
One NEXMIND learns 2048, Rocket League and Minecraft at the same time, through one gamepad. Each world has its own eyes. Everything else is shared, and nothing tells it what the worlds have in common.
Measured against itself.
2×stronger 2048 play than its own first games
4.6×more world dynamics captured with grounded perception
IIInside NEXMIND
The mechanisms that make it learn.
Four pieces of the architecture, running live. Each is a deliberate answer to how a mind should learn while it acts. Try them.
Grows when surprised. Prunes what it doesn’t need.
When the world shows it something its model cannot explain, a dormant component wakes up. When a component stops earning its keep, it is pruned. The model’s size follows the evidence.
Complexity has to pay for itself.
Structure that earns its place.
A challenger runs in the shadow of the active model and is judged only on data that arrives after it was proposed.
Promoted atomically. Rolled back instantly.
The plan bends. It doesn’t break.
Between planning steps, a variational corrector revises the next command 120 times a second, inside strict bounds.
Fast correction, anchored to the plan.
IIINEXMIND Deployment
A mind with a body.
The same principles, engineered for flight: verified sensors in, a deterministic safety authority between every thought and every actuator, and NOMAD Mk. II, the aircraft it is built to fly.
Every frame is verified.
Every camera, depth and thermal frame is signed, ordered and checked for replay before any part of the mind is allowed to believe it.
authenticated RGB-D · radiometric thermal · tamper and replay testsObjects, not pixels.
A C-JEPA world model written natively in Rust learns object slots from camera video and follows them through occlusion, with gradients reaching across 4,096 frames.
Burn · CubeCL · Vulkan · pixel-to-slot encoder · thermal JEPAA belief it can defend.
Its map of the world changes only through authenticated evidence receipts, so every update traces back to something it actually saw.
finite spatial belief · signed retraction auditFive timescales, one decision.
Five temporal layers each run exact expected-free-energy inference. Slow layers send priors down; fast layers send evidence up.
EFE as variational inference · prior and evidence busesIt changes itself, carefully.
Parameters and structure learn online, through Bayesian model reduction, History and Surprise banks, and immutable generations that can roll back instantly.
structural evolution · atomic generation hand-offNothing learned touches the controls.
An independent, deterministic safety authority admits or refuses every single command, with a hold-and-decelerate fallback.
isolated supervisor · one-use command leasesProven in simulated flight.
Helicopters of the 700 kg class, flown in simulation with online learning and safety gating active: search and find, and more than 34 minutes of continuous autonomy. Here, the rotors spin up and NOMAD lifts off into forward flight.
ArduPilot boundary · blade-element rotor · Pitt–Peters inflow · Gazebo FieldLab
- 700 kg
- helicopter-class autonomy in simulation
- 34+ min
- continuous autonomous operation
- 589
- parts in the whole-aircraft CAD
- 37×
- stiffer engine-bay frames at R12
IVThe destination
Toward an Artificial Entity.
NEXMIND is the first step. The goal is an entity whose experience stays with it, learned into its parameters as it lives, not held in a context window.
- 01
Scars.
A painful outcome rewrites the weights. Afterwards, it is a different system.
- 02
Teaches.
It learns from its own consequences, and from watching others act.
- 03
Sticks.
Nothing parked in a prompt. Carried in its parameters, across sessions and worlds.
Memory you can wipe. Experience you can’t.
Give both systems a few experiences. Start a new session. Then show them the fire.
context window · 4 slots
- —
- —
- —
- —
Empty window.
Shown the fireReaches for it
Shown the fire · predicted harm 35%Reaches for it
A real neural network, learning by gradient descent in your browser. NEXMIND learns the same way at scale: prediction error updates its weights, painful and surprising moments are kept whole and replayed more often, and checkpoints carry what it learned across restarts.
VEnd to end
One person. Every layer.
N-TEK is built from the GPU kernels up to the airframe, natively and end to end. Select a layer.
A whole aircraft designed in CAD, from exterior to internal structure, with finite-element checks and airflow studies.
- FreeCAD
- FEA
- CFD
- Blender
Back the work
N-TEK is one person. Imagine it with compute.
I’m İbrahim Alican, an aerospace engineering undergraduate at Istanbul Aydın University. I do this research alone, on a single workstation, without institutional funding. Everything on this page came from one desktop GPU.
- 01
Compute
The bottleneck. GPU hours, cloud credits or hardware turn straight into bigger models and longer experience.
- 02
Fund
Grants, sponsorship or investment turn a solo effort into a lab, and NOMAD from CAD into hardware.
- 03
Collaborate
Active inference, world models, continual learning: I want sharp eyes on this work.
- 04
Hire
Roles and internships where this research can grow. I build end to end, from theory to Rust to airframes.