Skynaptor · the engine
We rewrote the inference engine.
General-purpose engines are built to run anything. Ours gives up that generality and gains speed in exchange, because it is built around one thing: what a mission really asks of an AI.
The idea
A decision is not a sentence.
Is it a person? Which option? Where is it? An ordinary engine makes the AI write a sentence, and then a program reads it back. But the answer is already in the model before the first word. We read it from there.
The usual way: write a sentence, then read it back
Skynaptor: straight to the answer
Skynaptor: a second question on the same image
Why it is so fast
No magic. Just an engine with one job.
Decisions in milliseconds
On a single graphics card a decision is read at once, without generating a word.
Many requests, one card
Several requests move forward together in the same round, so the card is never idle.
It looks ahead
A small model predicts what is coming, and the big one only has to check it.
It does not start over
What it already knows is kept and shared, instead of being worked out again every time.
Much memory, little space
It keeps a huge amount of information in mind on a single card, because each piece takes a fraction of the usual space.
It knows who is asking
Whoever is waiting at a screen goes first. Background work takes the capacity that is left.
Fast, but not in command
The AI decides what to look at. It never touches a motor.
On your servers it prepares the mission. On board it recognises what the drone sees, without the internet. Between the AI and the autopilot there is always a safety filter that follows fixed rules.
- AIEyes
Always watching what the camera sees.
- AIQuick judgement
Is it a person? If in doubt, it takes a better look.
- AIStrategy
It can propose a change. A person has the last word.
- RulesHands
A short list of allowed actions, and nothing else.
- RulesSafety filter
The only way to reach the autopilot.
- RulesAutopilot
It flies, as it would with any pilot.
Your model. Your servers. Your call.
Use ours on your hardware, or connect the platform to one of your own.