Kevin Russell
speaker
741 appearances
1 recordings
1 series
first heard Jun 2026
last heard 29 Jun
Kevin Russell’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 1 in all, peaking in Jun 2026 with 1.
Appearances
And what I find uh you can finish your thing and then I wanna jump on.
Which super fascinating you say that 'cause you just connected uh one of our findings that Hindi costs sixteen X the amount of tokens to process downstream.
So in all of our tests when we were running the gamut of all the different languages and then comparing the gate, because we can take any model now and uh three out of uh the hardest of the the questions, uh the battery that we put it through, uh like Chat GT five with the highest frontier will confabulate uh two out of four times.
That's huge.
But put it through this gate that we made and it drops it
To uh just missing one out of 24.
Like it dramatically lower.
Um uh but what's happening is like downstream, you're working in the language and it's over overrepresented in English.
And so when it's trying to find all of uh the the associative web uh with Hindi that matches all of English to try to find an output, it costs way more.
But that's because
you're working in the language domain.
That's not where it thinks.
It thinks and it has its meaning upstream.
Upstream.
If you go mid-stack in the residual and you start pushing the vectors from there, you're gonna have a much different output downstream.
And it wouldn't cost you that same 16x for having Hindi verse uh verse English.
It wouldn't
Yeah, well it's coming into the language, like the way that we've like s structured it is much more English bias.
And so the way that they're actually pushing this to the public is costing them the sixteen X.
It's a deliberate choice for that.
Showing 181–200 of 741 · page 10 of 38
← Previous
Next →