Filippo Valli

chief mog officer

[Filippo Valli @siliconvallii]

grok 4.7 is such a downgrade

[Filippo Valli @siliconvallii]

$5T company btw

[Filippo Valli @siliconvallii]

@IndependentEco I guess they forgot about it then

[Filippo Valli @siliconvallii]

Why does Antigravity still have Claude 4.6?
What's the point of not updating?

[Filippo Valli @siliconvallii]

i need perplexity but for power users

[Filippo Valli @siliconvallii]

@MrAhmadAwais when on desktop app?

[100 @yh_me2]

I need to consult the dawg in me about this

[djcows @djcows]

drinking water to understand how a datacenter feels

[atlas @creatine_cycle]

dudes will be skinny and be concerned with open source vs closed source models.
bro you're worried about the wrong weights

[Filippo Valli @siliconvallii]

@hthieblot why am i here

[Filippo Valli @siliconvallii]

Aristotle of VC

[Marc Andreessen 🇺🇸 @pmarca]

Current status: Retardmaxxing.

[Filippo Valli @siliconvallii]

@yongfook it's $18m but ok

[Filippo Valli @siliconvallii]

@mark_l_watson can it be sufficiently good for anybody to actually use it in apps?

i dont want to dunk on it, it's amazing for 0.8b, but it feels so unreliable to me to actually be used in any real world scenario, even the simplest. maybe it gets it right sometimes, but not consistently

[Filippo Valli @siliconvallii]

@jenzhuscott @Alibaba_Qwen @lmstudio @huggingface and you use it for..? no way the 0.8b can be useful with any task, no way

[Filippo Valli @siliconvallii]

@VadimStrizheus how can a 9b model "perform to almost claude level" 😭😭😭

[Filippo Valli @siliconvallii]

@elliotarledge i'm a big fan. stay strong.

[Filippo Valli @siliconvallii]

@elliotarledge will do.

keep it up

[Filippo Valli @siliconvallii]

i was feeling good but then i saw this dude lock in for 50hrs straight.

my day is ruined thank you.

[Elliot Arledge @elliotarledge]

timelapse #87 (50 hrs):
- 2800x speedup
- i suggest you stop for a min and seriously watch this whole timelapse. speed and energy has been very consistent this time around.
- was relying on claude 4.5 sonnet but its only worth using on niche problems, not codebase refactors
- found myself deviating back toward xAI models by the end
- figured i should tackle the final boss of low precision GEMM kernels nvfp8 and nvfp4 which led me to cutlass and cute so im now tackling two chapters (gemm optimization chapter + cutlass/cute chapter) at once
- 30 min mentoring meeting
- figuring out how to get O-1 as soon as physically possible
- vibe coded weak + strong scaling tests for 8xH100 node
- worked my magic on quantized video lms
new monitor + headphones
- merged and simplified simon’s SGEMM_CUDA and pranjal’s h100 kernels to make my life easier on seeing the last of optimizations to apply
- optimized gemv, hopper gemm, ampere gemm, cuda core gemm, softmax, layernorm, topK chapter carefully being shipped
- implemented pipeline parallelism on 8xH100 for arbitrarily large MLP (inference only educational example)
- wont have to worry about quantization chapter until my editors look at it
- havent full polished the open source repos but ill roll those out passively
- everything is going my way except balance

[Filippo Valli @siliconvallii]

Sonnet 4 is too good to be true 😭