indirectjump

[indirectjump @indirectjump]

@hu_yifei You need NVFP4 + the flashinfer a2a backend enabled. The out of the box a2a kernel is dog slow and so is the Marlin int4 grouped gemm

[indirectjump @indirectjump]

@thdxr @meansoabstractn Lol just completely wrong. Deepseek inference margins are public. And they don’t even have NVL72s that literally give you 10x gains in inference throughput.

[indirectjump @indirectjump]

@ben_j_todd In what world is inference compute 50% unless you count free users?

Your B200 numbers are way off too, since SemiAnalysis showed DS R1 can get >9k TPS/GPU with NVL72 + PD disagg.

[indirectjump @indirectjump]

@finbarrtimbers They’re moving up the stack to capture more margin. Fully intentional and imo the right decision if you have the best model.

[indirectjump @indirectjump]

@StijnSmits @eliebakouch xhigh thinks for 10000 years while opus takes 10 seconds

[indirectjump @indirectjump]

@tokenbender The thinking time is just too long unfortunately

[indirectjump @indirectjump]

@gallabytes You can just predict the distribution in RGB space instead of expectation of the distribution

[indirectjump @indirectjump]

@rosinality I think they really should’ve tested this with English/Chinese, having shared tokenization between Latin languages helps hugely imo

[indirectjump @indirectjump]

@zephyr_z9 Seems easy to raise with bonds considering their revenue

[indirectjump @indirectjump]

Returns in society are accruing to the tails more than ever and your average SWE is not ready for the wave that’s about to hit them

[indirectjump @indirectjump]

@zephyr_z9 @teortaxesTex If you’ve been to KR/JP it’s very obvious they were never going to win just based on culture

Same reason NYC hasn’t produced any notable work in this space

SF has enough non grifters and risk takers still

[indirectjump @indirectjump]

@teortaxesTex Increasingly top heavy

Impact & income will continue to concentrate exponentially more I think

[indirectjump @indirectjump]

@FangYi11101 Tabelog is not that great tbh

Lots of 3.8-4.0 places that are mid

Beli in US is probably higher hit rate than Tabelog in JP

[indirectjump @indirectjump]

@teortaxesTex Better hardware = stack more SRAM

Already been tried (Cerebras, Groq, SambaNova)

Diffusion doesn’t improve this if you decode 1 token at a time

[indirectjump @indirectjump]

@_xjdr It saves them a couple iterations of synth data -> train -> repeat doesn’t it?

[indirectjump @indirectjump]

@suchenzang 4x the flops, what’s the point you’re trying to make here?

[indirectjump @indirectjump]

@redtachyon I mean sure this works if you did ICPC but what you do on a daily basis is generally somewhat poorly correlated to interview questions

[indirectjump @indirectjump]

@rosinality https://t.co/XX7zGJWjiz

[indirectjump @indirectjump]

@yacinelearning It’s a big hack

The reason GRPO like objectives are unstable with MoE is that a single gradient step can cause you to greatly exceed the clip bound due to rerouting, and then you won’t update to fix it bc you’re clipped. The right way to fix it is to stop clipping like CISPO.

[indirectjump @indirectjump]

@joel_bkr @scaling01 Elicitation is harder than people think!

[indirectjump @indirectjump]

@ViaFloo @apples_jimmy Spending time on safety is exactly why they can do it

All modern RL techniques come from safety (oversight specifically)