WTF is behind ox alpha ????
this thing just rolls up on openrouter with 1m context, multimodal, coding god, and oh yeah it can serve 100 trillion tokens a day !!!
like⦠how do you even HAVE that much compute lying around???
@straceX Absolutely mind-blowing what humans have engineered. To shrink incredibly complex technology down to a scale thatβs almost impossible to visualize is unimaginable
The reason starts with how memory is addressed...
Memory is byte-addressable.
Basically, the CPU doesn't normally get an address for an individual bit. It gets an address for a byte.
So even though "true"/"false" only needs 1 bit, a normal variable gets at least a byte of storage.
But here's where it gets interesting...
You can pack booleans into individual bits.
The reason we don't do that for every "bool" is that accessing an individual bit is more work.
The CPU has to get the byte containing that bit and then extract it using masking, shifting, etc.
And modern CPUs are insanely fast.
So we're often willing to spend those extra 7 bits to keep access simple and fast.
But then I found out something cool in C++:
"std::vector<bool>" actually does bit-packing.
So:
bool arr[100];
typically takes 100 bytes.
While:
std::vector<bool> v(100);
can store those 100 values in roughly 13 bytes.
That's a pretty crazy difference once you're storing millions of flags.
So those 7 bits aren't necessarily "wasted."
You're trading memory for simpler access.
The future of computing may be determined less by how many operations a processor can perform and more by how efficiently the system can move information to those operations.
@harshilmathur@Razorpay@nvidia@awscloud One model that actually understands how money moves, not just another layer of AI bolted on.
8β10% better success rates is no small thing.