Hey I'm Konrad, I work in ML for a top 0.1% neuroscientist
At 16: Cold emailed 200 neuro professors. Grew a moustache to look older in interviews. Landed spots in 2 labs.
At 17: Led a project for a top-published professor. Built a cold-calling startup on the side. Met unicorn founders. Got a girlfriend.
At 18: Backed by a Waterloo VC. Running pilots of our neuro project in clinics. Scout at Front Row Ventures.
In Waterloo for now. If you're with a lab or AI startup, let's talk
@winglian Is the 8% speedup mostly from making fewer trips to GPU memory? And does it still help as much with long inputs, where attention takes up most of the time?
@Henry_Ndubuaku The blog mentioned Q/K settles in the first quarter of training, could you freeze routing at that point and train only the write paths to cut training compute?