Pinned
stephen balaban
1,910 posts
stephen balaban
@stephenbalaban
Building supercomputers for AI. Co-founder and CTO of Lambda.
Joined September 2010
This is the strongest evidence that I’ve seen so far that we will achieve AGI with just scaling compute. It’s genuinely starting to concern me. I used to think that we would run into roadblocks: end of scaling laws, maybe we don’t have the right model architecture, power densitySora performance scales with compute
00:00- Fuck it, we just dropped our H100 on-demand prices to 1.99. From now on, Lambda will always have the cheapest price in the entire world, if we don't, just tweet at me and we will drop the price. H100s here: cloud.lambdalabs.com/sign-up
“Nvidia H100 and A100 GPUs - comparing available capacity at GPU cloud providers” llm-utils.org/Nvidia+H100+an… - Sometimes shader golf makes me think the minimum description length of our universe might not be that long.for(float e,i,a,w,x,g,h;i++<90.;o+=.01-.02/exp(max(s,e)*3e3)/h){vec3 p=vec3((FC.xy-.5*r)/r.y*g+2.,g);p.zy*=rotate2D(.5);e=p.y;h=e+p.x*.3;p.z+=t;for(a=.6;a>.001;a*=.7)p.xz*=rotate2D(5.),x=(p.x+p.z)/a+t+t,e-=w=exp(sin(x)-3.)*a,h+=abs(dot(sin(p.xz/a*.3)*a,r/r));g+=e=min(e,h*.5-1.);}
00:00 - That’s nice, but try $1.4B in revenue with 24 people or $3.7B with 28 people. Let’s see who can name these two companies.$100M in revenue with less and 10 employees is this generations' 4 minute mile.
- Yale has a perpetual bearer bond from the Dutch government from 1648 that pays 2.5% interest, they go to the Netherlands every few decades to collect the back interest. It’s the oldest security still generating payments.
- This was Lambda’s first supercomputer. Built in 2015 to run the Dreamscope style transfer app. It was 32 NVIDIA GTX 980s. We were running out of cash from a $40K/mo cloud bill. We called it the Deep Beast Machine and it saved Lambda. Today Lambda has 25K GPUs. never ever quit
Replying to @stephenbalabanI love that you just kept building bigger and bigger computers. Congrats on the raise! - Replying to @flavioAdI just did this in Lambda Chat’s DeepSeek R1 and it got it perfect the first time. I’m not sure your conclusion is accurate. Which hosted DeepSeek are you runnning? Hosted Replit: …xagon-lambda-chat-deepseek.replit.app Prompt: Write an html JavaScript app that shows a ball bouncing inside a
00:00 - Lambda Chat now has DeepSeek R1 671b and it's FREE to use. DeepSeek's R1 model, hosted in the USA. We don't train on your data. Use it here:
- NVIDIA popped because everyone underestimated the amount of GPU compute the world needs. I’m honored that Lambda was mentioned in NVIDIA’s earnings call today. Soon, every human will be using AGI powered by GPUs. And Lambda is going to be the #1 GPU cloud in the world.
- This is how Copilot is “programmed”. This is how your future AGI co-workers, friends, and lovers will be “programmed”. The future is here, it’s just not evenly distributed.
- I no longer use Google when programming and default to thinking: "I should ask ChatGPT about X." LLMs are a far superior method vs. the standard (read documentation / blog post / stack overflow) loop. It's like having a thoughtful pair programmer with you at all times. A thread:
- Lambda has raised a $500M GPU-backed financing vehicle. Designed to fund GPUs for hundreds of thousands of developers with No Contract Required. This is half a billion dollars for the 'gpu poor'. When new technologies appear in the market, they require new systems of financeToday, we are announcing $500 million in debt financing to scale Lambda’s On-Demand GPU Cloud. Learn more: bwnews.pr/4aoHCZQ
- Lambda has raised a $320M Series C led by Thomas Tull's USIT. This new financing will be used to expand the number of NVIDIA GPUs available in Lambda Cloud and build features that will absolutely delight you.
00:00
