Sahil Chelaramani retweeted
I don’t think people realize how selfless and wonderful the StackOverflow’s founders were.
Unlike reddit, x etc they insisted ALL of their content will be 100% open, nothing will ever be locked up in walled garden.
They made it a point by even offering structured data downloads!
They didn’t cared people cloning the website. They didn’t worry about putting the whole monetization thing at risk.
It was a true hacker spirit in it’s purest form.
Most folks won’t be able to grasp how hard this is. You are on top of the world. Millions of people depend on you daily. You have the rarest opportunity to rack in billions. And you still stay true to your principles.
This exactly in turn later enabled AI revolution. I still remember downloading entire SO archives and preparing it for pre-training while Reddit and others went onto lawfare and firewalling the content. Note that this was the content their users - not reddit employees - had created for free and voluntarily to help others and the world. But they locked it all up for profiting from it personally.
SO did none of that. I think they knew they will be sacrificed at the altar. Yet it will exactly serve the intent of their founding principles to help others and the world, now through AI, even at bigger scale.
All of today’s models are built upon efforts of such true pure hackers who spent their days and years to create training data for AIs. I was a proud SO contributer for many years and there are couple of times while testing models I did felt that models might have used my answers 😇.
Hats off to @codinghorror and @spolsky and so many builders and hackers at SO for moving the world forward, one revolution at a time 🫡.
Sahil Chelaramani retweeted
There has never been a better time to go into math: youtube.com/watch?v=LKrwhjOI…
Sahil Chelaramani retweeted
There are a lot of things wrong with this world…
but too much intelligence is not one of them.
Sahil Chelaramani retweeted
Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats to @alexandr_wang and the MSL team for all your great work on these models.
Absolutely disgusting football this match. Argentina didn't deserve to be here. Shit sportsmanship. Thankfully cheaters didn't prosper prosper Good riddance to this Argentine team. Vamos!
Open-weight models are closing the AI gap. I don't care if I can't run them locally. I care that no single company controls them. A healthy ecosystem means competition, choice, and providers aligned with our values. Stallman described this beautifully youtube.com/watch?v=lrcdhzr2…
Open-weight models mogging closed-weight models on software benchmarks was not on my weekend bingo card, but I'm absolutely here for it 🚀
Sahil Chelaramani retweeted
“Life to me was never about accomplishment, but about Happiness”
My man, Steve Wozniak, is an inspiration!
Sahil Chelaramani retweeted
Too late to explore earth, too early to explore the universe.
Right on time to explore super intelligence though :)
Sahil Chelaramani retweeted
This explanation from @edfrenkel (with @lexfridman), of the development of complex numbers, is a much better explanation than I could provide of the kind of breakthrough research that LLMs, as currently defined, are perfectly set up to be terrible at.
youtube.com/watch?v=00q7td_X…
Sahil Chelaramani retweeted
And... the first-ever ICLR test-of-time award goes to "Auto-Encoding Variational Bayes" by Kingma & Welling.
Runner-up: "Intriguing properties of neural networks", Szegedy et al.
The awards will be presented at 2:15pm tomorrow; there will be retrospective talks. Please attend!
























