Kelip

AI Postdoc in NUS
Vibecoder
Sharing AI knowledge and experiences with everyone.

[Kelip @JackSk35800]

JEPA, From Features to Actions
JEPA predicts representations of the world. Follow the path from LeCun’s proposal to V-JEPA 2, including how frozen video features become a robot planning system. #WorldModels #Robotics

[Kelip @JackSk35800]

Robots Learn in Dreams
World models let a policy practice in imagined futures. Here’s how World Models, Dreamer, and DayDreamer connect visual prediction to real robot learning. #WorldModels #Robotics

[Kelip @JackSk35800]

@npceo_ how to deal with the image modal?

[Kelip @JackSk35800]

@tony_worldmodel yeah, its long journey been there

[Kelip @JackSk35800]

π0.7: One Model, Many Robot Tasks

Fold laundry, make coffee, follow new instructions. I break down how memory, world-model predictions, and quality prompts help π0.7 generalize across tasks and robots.

VLA Fundamentals, Part 4. #Robotics

[Kelip @JackSk35800]

I plugged @typesafeai's Jev into PaperDance and my arXiv feed stopped being wrong.
Same 100 candidates. One yes/no question per paper. Off-topic cards on page 1: 4→0, 9→0, 7→0. ~1 s and $0.0004 a page.Left: before.
Right: after.
Real production data.
https://t.co/lKVkMTBKzO

[Kelip @JackSk35800]

π*0.6: Robots Learn From Experience
RECAP uses robot rollouts and human feedback to improve speed and reliability. I explain rewards, value functions, and advantage conditioning through a delivery-rider example.

VLA Fundamentals, Part 3. #Robotics

[Kelip @JackSk35800]

π0.5: Robots in Unseen Homes

How does a robot turn "clean the kitchen" into useful actions? I break down FAST action tokens, diverse training data, and how π0.5 connects language to movement.

VLA Fundamentals, Part 2. #Robotics

[Kelip @JackSk35800]

VLA Explained: From RT-2 to π0
A robot sees an image, reads an instruction, and predicts its next moves. I break down RT-2, OpenVLA, and π0: action tokens, flow matching, and the path from language to robot control.

VLA Fundamentals, Part 1. #Robotics

[Kelip @JackSk35800]

跟着H3学视频生成4
视频模型的位置编码

[Kelip @JackSk35800]

跟着H3学视频生成3
AdaLN

[Kelip @JackSk35800]

跟着H3学视频生成2
H3模型的信息流

[Kelip @JackSk35800]

跟着H3 学视频生成系列1
flow matching

[Kelip @JackSk35800]

@MikeShou1 great work

[Kelip @JackSk35800]

It's a tiny indie project and completely free. No paywall, no invite code.

Try it on your phone: https://t.co/mpR41wIZd3

Tell me what direction you'd want covered next, and which feature you'd want most.

[Kelip @JackSk35800]

Every paper becomes one full-screen card:
• the key figure, pulled from the paper
• TL;DR + abstract, expandable to the original
• affiliations, first & corresponding authors
• venue badge, citation count

(Chinese cards add a one-liner + 3 bullets.)

Swipe for the next one.

[Kelip @JackSk35800]

Like, save, comment, or "not interested" — right rail, exactly like the short-video apps.

Your saves live in "My Library" and export in one click: BibTeX, Markdown, CSV, JSON.

My workflow: swipe in PaperDance, then hand the export to Claude Code to read the papers in depth.

[Kelip @JackSk35800]

I kept missing papers I should have read. Not because they were hard to find, but because scanning 300 arXiv titles a day is a job by itself.

So I built PaperDance: swipe through arXiv papers like short videos.

Free, no app, any browser → https://t.co/mpR41wIZd3 🧵