@theo@robinebers Do a video on that. I think that harnesses themselves should approach zero. I.e. the frontier models are smart enough to do things on their own given the right tools. Anything else as far as harness “features” are tech debt / bloat
Update: the far-pair conjecture in this paper is false. With one partner time per direction, Clifford-algebra matrices shrink the measured rank to (q+1)/4^q of m(Δ−R), even over Q. Partner times can't be dropped. The capacity-free question is still open.
https://t.co/ddNb4ELDsm
@xennygrimmato_ asked how to couple corrections that vary with evaluation time. Our note uses a determinant argument to do that, with layout bounds under bounded incidence, expansion and a capacity condition. The capacity-free case remains open. https://t.co/MzfOeOh1c1
@notkevinzhang Jev within orchestration to make LLM’s “behave”. Route, gate, stop, verify, rank and interrupt swarms b4 wasting inference
Possibly use as a pre-ranker in MTP / spec decoding
I came out of the Twitter woodworks the first time in years to reply to this lol
email [email protected]
@DanPriceSeattle CPI is based on the products and services the average person buys. A CEO isn’t gonna buy 2x as much baby formula cuz they got a raise. The VELOCITY of money matters just as much as quantity when it comes to inflation. Gov’t could print $500T but if it sits still nothing happens
@DanPriceSeattle CEO’s getting a pay raise or big companies being bailed out won’t affect the CPI much, bc those entities won’t demand more of the everyday consumer products and services that the CPI is based on. That money would prolly go into their company or sit static in investment accounts