@dhh@anduriltech@Snapdragon@PalmerLuckey omarchy on bluechip is a cool combo. how's sleep and battery on snapdragon linux these days? that always seemed like the part holding arm laptops back
@AndurilObserver feels like the answer is putting the multirole part in software and swappable payloads instead of the airframe. keeps the per unit cost down and you still get the flexibility. curious if fury's bays are built for that
@Saronic what does swapping mission payloads look like on marauder? like is it a standard container interface so the same hull can go from logistics to sensors in a day, and does the autonomy need retuning when the load changes that much?
@BuildMonumental a site just for staging and training operators is such an underrated part of deploying robots. how long does it take to get a new operator up to speed?
@DavidSKrueger yeah coding and math are easy to verify so that's where everything gets measured. the stuff that's hard to grade is exactly where we know the least
@satnam6502 one program that targets cpu, gpu and fpga is such a cool dream. the hard part always seemed like memory, every target wants the data laid out differently
@ico_TC partial reconfig feels like such a good fit for open tools bc the vendor flows make it so painful. what's the main thing missing right now, the bitstream side?
@thuHanShen this is super useful. for running stuff locally memory traffic is what actually decides speed, so seeing that per model before you download it is huge
@aaron_defazio fixing the late stage weirdness is a big deal, that's usually where long runs get unstable. is it mostly about how weight decay interacts with the lr schedule?
@NextScience commercial approval is a huge step. how fast is the decoding though? robotic hand control needs pretty low latency or it feels laggy and people just stop using it
@levidiamode 308 days straight is insane lol. the edge llm lab is a good one, on a laptop most of the speedup ends up being memory not compute so it's cool to see where simd actually helps