VIBEcoded kernel gave a 3T param model 2x speed increase
>>110101928ok.
>>110101928don’t believe you, bullshit
2x speed-up for a 3T param model? how did they get 2.5x speed-up for a 5T param model? I didn't even know there's 7T param models already, how did they get a 3x speed-up for it?
>>110102073Larger models (especially in the multi-trillion-parameter regime) often see higher relative speedups from the same class of systems/software/hardware optimizations because fixed or poorly scaling overheads become a smaller fraction of total runtime, while compute intensity and utilization improve.
Anyone used this to increase the speed of a Bitcoin mining software?
>>110102570why would anyone tell you if they did
>>110101928*cuck coded, also this won't stop the AI bubble from popping
>>110102570Figure that yourself
>>110102649>>110102687Fair enough
>>110101928the y-axis shows each speedup being significantly more expensive to compute per token. what's going on here exactly?
haha I though with my background in systems and ML I could get that kind of job when I graduate
>>110105395you just described what's going on
>>110101928WE WON, VIBEGODS
