the talisman of deep learning.
Joined March 2022
- Tweets9.3K
- Following690
- Followers3.2K
- Likes10.5K
Pinned Tweet
mother of all demos. pt 2.
bodega OS is beautiful, unlimited and all ai models running locally on your hardware.
from porsche, westworld, to learning different planes of power, the giests and vishnu. playboi carti & personal notes at the end too.
it's crippling to use any other ai now. everything else feels bleak and lifeless.
i should not talk like this even though this aged well hell nah brah😭
do you see it. do you feel the retard spectrum increasing.
it’s fairly “uncommon” right now to see anyone with a logical and sensical point of view on this platform tbh. so it’s fine.
david is now part of that retard gaussian distribution.
this is the third time i have cried. 2022 WC, then when my nephew was born, and then last night.
i wish i would have seen you play. one day we will meet.
🎙️ Messi On Enzo Fernandez during Messi’s Farewell Speech :
🗣️“After I announced my retirement from the national team in 2016, a kid in our group wrote a letter to me asking me to stay for a few more years. That he would be proud to share the pitch with me. The words struck me and I had to change my decision.
4 years later, I won Copa America and Word cup 🏆 with the kid.
The kid is Enzo Fernandez “
must read. this guy literally said it in the most optimistic way, which i would never. but he’s right.
Language is not fundamental to human consciousness. In terms of evolution, it arrived only recently on our timeline.
Consciousness evolved over millions of years through sensory awareness, emotion, and the embodied survival instincts of our tribal past; our capacity for empathic engagement and community. Syntactic language only emerged around 100,000 years ago.
This is exactly why Large Language Models (LLMs - so called AI) are the wrong pathway to true artificial consciousness.
They try to replicate the late-stage "output layer" of human intelligence without building the foundational, embodied awareness that actually generates it.
Statistical text prediction is a mirror of thought, not the spine, the guts, the nerves and the subconscious, empathic ability to connect with other living beings. LLMs can never feel, let alone think.
We have been sold a trick, not a pathway to superintelligence. LLMs are fraudulent. Their capabilities have already peaked and we are being sold lies about their future potential. Lies with the goal of companies reaching IPO and tricking investors out of hundreds of billion$.
i bet none of these companies including @PrimeIntellect who received vera don’t even know how a instruction set works let alone arm v9.2A to even compare anything with x86.
their basis of comparison? running tool calls on a sandbox.
ridiculous times and clown world.
if my grandma had wheels she would be a car. that’s how absurd you sound.
these robots can’t even walk straight without bashing their heads in without a teleop. and you expect them to be highly synchronised and highly dexterous and work in collaboration?
david, are you dense?
your physics is right for 8+ gpus split across cpu ports. but with 4 cards, one gen5 switch with 4 x16 ports covers it. all 4 cards sit behind one chip, so cross traffic never touches the cpu.
the jam only appears if you split cards across switches. and with 96gb each, most serving setups run tp2 or independent replicas, so there's very little cross-card traffic.
on the supermicro side, i mean check the model: h12 boards are gen4 (about half your 53-56gb/s), h13 is gen5, so only h13 matches your numbers.
i'd benchmark all-reduce on one switch before buying 3. if tp4 is fast enough, the extra switches are wasted money.
and also killy tell me the top 4 usecases you need your server for. and then mention the top2 and then the most important one. based on that i can help you the best.
for example for me, i need train models as well but most of the for production infernece. i have served up to 95k users on my 4x rtx pro 6k WS, 2x 5090 and couple of 4070 supers and 3090. so i had my setup done accordingly.
So the reason why I wanted 3 for root/leafs is to keep every cross-switch pair at 53-56GBs
Please correct me if I’m wrong, but my understanding about those backplanes is cross-chip goes GPU → chip → CPU root port → chip → GPU. All traffic meets at the CPU, where across chip boundary traffic shares one uplinks where my calculate shows it impacting speed significantly.
Within chip looked fine, just cross-chip traffic that’s going to run into jam (my research says write dropping to 2.7GB/s) and TP8 ring crosses that boundary twice per step.
C-Payne 3 switch setup routes every cross switch traffic through root switch microchip at 53-56GB/s and never touches CPU
If that’s inaccurate, please let me know. If it is, well I always do things more expensive and inefficient way possible so it’s par for the course
in 1977 john backus called it the "von Neumann bottleneck":-the cpu spends its life waiting on memory. we have spent 49 years building faster pipes (caches, hbm) instead of removing the trip. robotic hardware runs one model all day, which is the ideal case for compute-in-memory.
that’s how hell bent and washed up research is. they build on false premises and elon is least eligible to have an opinion on it.
We cut our RAM in half for the Tesla AI5 chip (now 72GB of LP5) and 1/3 for AI6 (now 144GB of LP6).
This was the only way to get enough volume for Optimus production and greatly reduces cost.
As it turns out, we think this will have a negligible effect on Optimus performance, as memory bandwidth is a bigger limiting factor than total memory storage (bandwidth was held constant).
losers talking about losers hiring losers.
You are a struggling second tier competitor that is more unethical and lying just because you have no decency or sense of proper behavior and shows your desperation. Straight out lying about if Chris being fired I thought would be below even you.
Readers added context they thought people might want to know
Khosla is an investor in both Cognition and Factory. khoslaventures.com/category/enter…