Kaizhao Liang
@KyleLiang5
Followers
620
Following
10K
Media
301
Statuses
4K
Class of 2020 @IllinoisCDS 5 years at @SambaNovaAI Grad student @UTCompSci since 2023 Interested in new optimizers and neural architectures
Austin, Texas
Joined December 2018
TLDR: 1⃣ line modification, satisfaction (theoretically and empirically) guaranteed 😀😀😀.Core idea: 🚨Do not update if you are not sure.👨💻🤗📚 @cranialxix @lqiang67 @Tim38463182
12
37
258
Reading Neurips reviews now 🤪🤪🤪
0
0
13
RT @nateliason: Waymo is a godsend for working parents. Need to hand off the baby between meetings?. Put them in a Waymo and send them to….
0
1K
0
Just realize cautious is basically a hard version of hypergradient lr scheduling 🤨.Instead of last batch of gradient, it uses momentum 😛.
Fun fact: When you're using a good learning rate, the gradient should be almost *perpendicular* to direction of the last step (opposite of the intuition of many gradient descent diagrams that make it look like the gradient is following a smooth path). You can derive this by
0
0
3
RT @rosmine_b: Fun fact: When you're using a good learning rate, the gradient should be almost *perpendicular* to direction of the last ste….
0
9
0
Knew this would be coming in some shape or form….
Some news: We're building the next big thing — the first-ever AI-only social video app, built on a highly expressive human video model. Over the past few weeks, we’ve been testing it in private beta. Now, we’re opening early access: download the iOS app to join the waitlist, or.
0
0
1
😀😀😀.
optimization theorem: "assume a lipschitz constant L. ". the lipschitz constant:
0
0
1
RT @itsPaulAi: Wait NVIDIA has just released new SOTA open source models?!. Available in 4 sizes 1.5B, 7B, 14B and 32B that you can run 100….
0
168
0
RT @eliebakouch: We've just release 100+ intermediate checkpoints and our training logs from SmolLM3-3B training. We hope this can be use….
0
59
0
Maybe a bit of oversimplification, but I am genuinely amazed that at the core of everything are variants of simple task objectives like classification (next token prediction) and regression (flow matching), nothing beyond these two seems to make significant difference 😝.
1
0
5
RT @MaziyarPanahi: Perfect Sunday: I just used Kimi-K2 by @Kimi_Moonshot to vibe code a @Gradio app! 🔥 . You can use "Anycoder" Space by @_….
0
17
0
RT @LakerNewhouse: [1/9] We created a performant Lipschitz transformer by spectrally regulating the weights—without using activation stabil….
0
78
0
RT @Gradio: Mixture of Experts? No no, Roundtable of Experts! . Introducing Consilium - When AI Models Debate Around a Table!. Results are….
0
15
0
RT @VentureBeat: Meet AnyCoder, a new Kimi K2-powered tool for fast prototyping and deploying web apps
venturebeat.com
For novice developers or even those with expertise who want to spin up a new project fast, AnyCoder seems like a great place to start.
0
14
0
RT @_akhaliq: You can install anycoder as a Progressive Web App on your device. Visit and in the footer click set….
0
11
0
In hindsight, would have been cooked if they did not exit in time. .
character.ai
Chat with millions of AI Characters on the #1 AI chat app. Where will your next adventure take you?
0
0
1
0
16K
0
I am showing this to everyone that says #AI is a bubble 🫧.Also is this what 200K GPUs can enable?🧐🧐🧐
0
0
0