AI 百球哥 (@kedaya) header
AI 百球哥 (@kedaya) profile picture

GitHub 10k+ followers · Global Top 100 Embodied AI researcher & architect Exploring Agents · AI Infra · AI4S · Self-evolving AI Follow, connect & collaborate 🌍

Joined July 2024
Photo by AI 百球哥 (@kedaya)Photo by AI 百球哥 (@kedaya)Photo by AI 百球哥 (@kedaya)Photo by AI 百球哥 (@kedaya)Photo by AI 百球哥 (@kedaya)Photo by AI 百球哥 (@kedaya)Photo by AI 百球哥 (@kedaya)Photo by AI 百球哥 (@kedaya)Photo by AI 百球哥 (@kedaya)Photo by AI 百球哥 (@kedaya)
I won a Silver Award at AMD AI DevMaster 2026! 🥈 Working on robotics, simulation, or embodied AI? Let’s connect!
45
5
160
5,794
Agent eval: two good replies don't validate an order-assistant prompt change.
5
AgenticTask saves 16,450 task instances from 20,323 construction attempts on RoboCasa; its execution agent completes approximately 38% of evaluated tasks. Wang et al. reuse task bindings for generation and execution. Validity is not success.
2
15
30yr-uptime is a Claude Code mod for taking care of your body.
9
Omi CLI puts an AI coding agent in your terminal with 13 models to choose from.
1
2
50
TeaserKit lets your coding agent direct your launch video. Apparently it wants director credits. producthunt.com/products/tea…
1
23
DeepQuest achieves an audit-assessed scientific finding rate of 95.24% on OASIS-3. Jiang et al. rank Alzheimer's hypotheses by feasibility and groupwise Elo; key associations hold in ADNI and AIBL. Audit is not prospective validation.
2
1
61
Arrowmax S1 spots a screw, picks the bit and sets the torque. My toolbox feels judged.
12
Claude-assisted math claim: 17 million lines of Lean code, no readable proof yet.
1
19
Astra Medium Fast is one Neon user's pick for urgent production debugging.
1
1
86
Kimi K3 gets a lightweight, interactive training-module demo. Finally, math I can poke at.
24
Isaac 0.5 for Bimanual YAM returns a 60 × 14 action chunk per observation. PerceptronAI fine-tuned on xDoF ABC and MolmoAct2 demos using three RGB views and both arms' state. The catch: no task success rate is reported.
2
18
For bimanual manipulation, how much does a 60 × 14 absolute-joint action chunk lose to closed-loop replanning? Paper: huggingface.co/PerceptronAI/…
1
1
5
Spring Insight skips AI calls at the quota limit while alerts and metrics keep flowing.
2
2
118
DeepSeek Harness gets a file manager for browsing, previewing, and editing workspace files.
22
AI can write 500 lines. Your approval still means you read them.
13
Gemini Antigravity still showing Starter after Google AI Pro? Check your account and Google Pay regions.
27
AdsNet achieves an adsorption energy RMSE of 0.032 eV and a force RMSE of 0.047 eV/A on DFT-calculated catalyst structures. Ai et al. use a local equivariant graph. The catch: locality may miss changes beyond the adsorbate's neighbors.
4
2
111
Kooyu moves your coding context to another CLI when one hits a rate limit.
21