Simo Ryu
@cloneofsimo
Followers
16K
Following
2K
Media
2K
Statuses
5K
"Im gazillionare founder of google but let me do IC work cause why not" Its honestly so fucking based
Sundar Pichai: “Sergey is spending more time in the office. He's literally coding, and some of my fondest memories over the last year is sitting with Sergey on large screen looking at loss curves as we train these models.” When Sergey is back to coding and watching training
10
5
273
Always curious and full of energy, Beagles turn every walk into an adventure 🐾🎉.
240
930
14K
Because the progress is so incremental, i don't think people realize that we have already essentially reached agi (minus physical task unfortunately). If you had to make a tradeoff, would you hire average person over LLM? Its just AGI isnt as interesting nor what we need. We
@tszzl Winning the IMO gold was seen as an AGI level difficulty problem
36
5
214
sora x disney is going to be pure magic. dream come true to be working with them.
Really happy to be working with Disney to bring some magic to Sora and image gen! Disney is the best storytelling company in the world, and our users really, really want to generate content with their characters.
20
15
287
Ive been using 5.2 for some time, its really good model, do try out!!!
2
0
20
Heres my standards for 2026: - uses muP over SP - uses muon over adam - can clearly explain difference between independence and zero mutual information - explain the problem with unscaled projections in transformers - explain 3 flaws with chinchilla scaling law, and what one
Here's my standards for 2026: - good-looking east asian guy eg John Cho - well dressed - thick hair and perfect hairline - light brown eyes - shy and introvert - wit and romantic - soft and gentle - read and write - into deep tech - think of universe sometimes - movie lover -
11
11
428
Problem with society is dumbass people go politics while actual problem solvers just solve problems
@cloneofsimo i think every person (under democratic rule) should try out activism at least once in their lives, for a single issue they care deeply about Not that it'll work, but -- what's the system for, right?
6
1
102
If alien trained autoregressive language model and deployed into space, but its running on biochemical chip, and if we interact with it, how will we know if its not a form of life?
We have just used the @Nvidia H100 onboard Starcloud-1 to train the first LLM in space! We trained the nano-GPT model from Andrej @Karpathy on the complete works of Shakespeare and successfully ran inference on it. We have also run inference on a preloaded Gemma model, and we
3
0
29
It is SO FUCKED up whats going on in SK. Nobody is aligned. Everyone wants to squeeze their bit of pension, exploit the system, etc. Starting next year, SK citizens cant transfer cash over 10k$ without approval 💀💀 https://t.co/SNVqVWCSoO
1
0
25
Why the fuck do we have one government Like we should have 4 governments that compete with each other to provide better services and actively try to acquire citizens Government workers should get paid proportional to profit government makes, so they would cut cost, etc.
309
27
734
As execution gets cheaper and cheaper, ideas gets more and more expensive.
1
4
46
This guy will one day recreate earth with javascript
6 years of thinking I can make a better portfolio. 1 year of actually building it. My new portfolio is LIVE 🚀 Please, don't break anything! 👉 https://t.co/rVPv9oWkjL
0
1
37
Crazy to think there is still no scientific explanation on why ice is slippery nor why we sleep Humans know astonishingly little about the world
15
2
103
Financially successful people never brag about how successful they are, Smart people never brag about how smart they are, Wise people never brag about how wise they are. But adorable cat owners always brag about how adorable their cat is, and somehow thats socially acceptable.
4
0
65
Nice work!
Optimizers are only as good as our ability to predict good hyperparameters for them when used at scale. With robust hyperparameter transfer, we find Muon and Shampoo consistently beat AdamW by 1.4x and 1.3x in compute for training language models up to 1.4B parameters. 1/n
2
3
77