Multimodal most likely. They are able to predict time to fire accurately and are able to do complex manouvres in the air. Sounds extremely impressive to me.
Maybe in the future wars will involve sending multimodal LLMs in charge of fighter jets to strike enemy country's infrastructure targets with minimal loss to human life.
If there’s something pilots really don’t like, it’s when someone else does things without them knowing exactly what, when and why. There’s a reason you separate pilot flying from pilot monitoring.
Automated systems are obviously crucial but they need to be so reliable and predictable that the pilot can form a mental model of how they work. If not, even simple systems like MCAS can wreak havoc. It’s like having a silent copilot fiddling on the controls - even the ”right” decision can cause problems. This is similar in AVs - when you have L3 autonomy you end up in awkward inbetween split-brain in-between states, that can result in new failure modes.
Maybe in the future wars will involve sending multimodal LLMs in charge of fighter jets
Probably not LLMs but yeah, it feels like this is the type of tech that belongs in a UAV without life support systems. We are probably in the last generation or two with humans in the cockpit anyway.
Comments
What kind of models do these run? Vision? What “predictions” are being made?
There is slightly more information on Helsing website: https://helsing.ai/centaur
Multimodal most likely. They are able to predict time to fire accurately and are able to do complex manouvres in the air. Sounds extremely impressive to me.
Maybe in the future wars will involve sending multimodal LLMs in charge of fighter jets to strike enemy country's infrastructure targets with minimal loss to human life.
If there’s something pilots really don’t like, it’s when someone else does things without them knowing exactly what, when and why. There’s a reason you separate pilot flying from pilot monitoring.
Automated systems are obviously crucial but they need to be so reliable and predictable that the pilot can form a mental model of how they work. If not, even simple systems like MCAS can wreak havoc. It’s like having a silent copilot fiddling on the controls - even the ”right” decision can cause problems. This is similar in AVs - when you have L3 autonomy you end up in awkward inbetween split-brain in-between states, that can result in new failure modes.
Probably not LLMs but yeah, it feels like this is the type of tech that belongs in a UAV without life support systems. We are probably in the last generation or two with humans in the cockpit anyway.
The risk being that your enemy cannot reasonably steal all your fighter jet airmen but the can absolute steal a transformer model.
LLMs? Sounds backwards to me. It’s not like there is a lot of text to process up there.
LLMs work on tokens. There's no reason those tokens can't be something other than text.