Never going to understand fake-altruism with founders, while their VC's are straight up cutthroat for ROI. It's ALL about the money. Please. Don't let your origin story delude you.

Lizzie Magie.

Being a billionaire is a pretty decent moat...

Opus 5.5 High on the $20 plan is worth about 90m of intense work. Man, this blows.

Having a $200 @CloudRiftAI grant is super helpful right now. I wish they had H200's, but the 5090 is good enough to make some progress on this Kernel challenge.

NGL I love working remote, but if I was younger and wanted to be in tech I would head straight to SF. There is literally nothing like the connections you will make and the opps you get. It's insane. Also, HIGHLY suggest making friends with even one YC founder. IYKYK.

I've gotta be the only person on X that has never tried @Muse

In a few days I'm gonna rent an H200 for the Open Kernel comp. There's no way I'm sitting in 3hr queues to test + use my test/submit cap.

Welp. I found my session limit on the $20 Claude plan. Grrrr. I really didn't want $100/mo plans at both labs, one of them has got to go... Making good progress on the Open Kernel Challenge btw. Will w/u more later.

Working on MoE, finding incremental gains and still in 300+ place. 2 days ago I was #2 on the MoE chart. Competition is rough.

So far Opus 5.5 is very impressive on the $20 plan. I thought for sure I'd one-shot the entire week over a 100 minute session...not the case. Hopefully post-IPO we can keep these levels of subsidization. It feels like they'd be fighting churn flipping that switch.

Wild to see so many OAI and Ant workers leaving after vesting equity (/s, btw) It's almost like they were there for the outsized gains not the mission...actually, that's exactly what it is.

Alright, switching to Opus 5.5 to see if I can get around this bottleneck I have on Open Kernel Challenge.

11+ hours into the Open Kernel Challenge. Best scored results so far/. 1.51× for attention, 1.33× for the full decoder layer, 1.27× for prefill. All against the challenge’s SGLang baseline on H200. Still so much do, this is hyper competitive.

My best @Meta Open Kernel Challenge prefill submission went from 1.10× to 1.27× against the baseline after switching the expert GEMMs to Humming. H200, NVFP4 weights, BF16 activations.