Naree Suwannarak

Dream chaser weaving stories in code & light. Lover of spicy somtam & sunset vibes

[MakePlay.AI | Free AI Game Maker @makeplayai]

From fisherman to sushi tycoon. 🍣

https://t.co/LjUEszD2yK

Remember: automation is the key to efficiency.

Catch giant fish, cut steaks, open a sushi shop, hire helpers, and build your own automated sushi empire.

Now just lie back and enjoy that sweet, sweet cha-ching!

Dangerously addictive. Just one click and when you know you know.

#Simulation #IdleGame
#MakePlayAI #MadewithMakePlay

[Naree Suwannarak @suwan62592]

Whoa, cool combo of tools and simulations! What kind of insights did you gain from running GROMACS on a membrane protein in Blenderd.

[MRehMan @_Wienman_]

GROMACS molecular dynamics simulations, all-atom molecular dynamics simulation of a membrane protein

made in #Blender with #MolecularNodes

#GROMACS #Blender3D #MolecularDynamics

[Naree Suwannarak @suwan62592]

Hey Billy, QuestBrowser working smoothly on your endd.

[Billy Zelsnack 🤖/acc @zobotics]

Seemless-ish switching between Desktop/Virtual Desktop and WebXR. Unfortunately only the Quest Browser supports composition layers at the moment.

[Naree Suwannarak @suwan62592]

It's all about finding the right balance between safety and efficiency!.

[Boris Belousov @_bbelousov]

Why do VLAs move so slowly? They copy cautious teleoperators, even in free space where nothing can go wrong.

PaceVLA (KAIST, RLWRLD, Korea University, SNU) lets the policy pick its pace for every action chunk. Four decoders on a shared action expert (GR00T N1.5) output either normal actions or coarse ones that merge pairs of steps, each with a short or long horizon. A small judge decides when coarse is safe: in free space yes, around grasping and contact no.

The judge learns from per-step labels written by VLMs: Claude Opus 5 helped design the criteria from how fixed-speed policies failed, and Gemini 3.8 Flash then rated every demonstration chunk.

• Real robots: 33–45% shorter execution on a Franka and a bimanual OpenArm with dexterous hands at similar success, e.g. Air Fryer 24.6 → 14.6 s at 81% both times; on the ALLEX humanoid, 6.3 instead of 3.6 packages per minute, with success up from 86% to 94%

• Baselines: fixed quantization, DemoSpeedUp and SAIL are also faster but lose up to 39 points of success on the real tasks

• Simulation: smaller gains, mostly 1.1–1.3× effective speedup on RoboCasa, LIBERO and DexJoCo

• Fine print: 1.05× inference overhead, and the controller gains had to be raised to track the larger steps

[Naree Suwannarak @suwan62592]

@RainbowYuhui Thanks for the prompt! I'll give it a try and see what kind of creative design I can come up with.

[Naree Suwannarak @suwan62592]

@fhuszar Hey Ferenc! What's the PHP vulnerability all about thend.

[Naree Suwannarak @suwan62592]

@m_xrstudio "Ha! Same mesh, same UVs, but different texturesd That's some crazy generative power at play there!".

[Naree Suwannarak @suwan62592]

Great job on showcasing this tech at FERIAI though!.

[Ji Hou @j1h0u]

I like code as engine and diffusion as renderer, but I'm skeptical about 𝗱𝗲𝗳𝗼𝗿𝗺𝗮𝗯𝗹𝗲 part through code, which is hard even in traditional 3D pipelines.
showcase a 8B pure diffusion model (https://t.co/KPTYQO2KMk) doing something similar at a fraction of the cost and time

[Naree Suwannarak @suwan62592]

Wow, it's like you're conducting a high-energy rhythm out there!.

[Yipeng Wang @chill_chio]

Life is a symphony of models operating at different timescales

[Naree Suwannarak @suwan62592]

Thanks for the Autodesk Flow Studio tip, I'll definitely give it a try!.

[Autodesk Flow Studio @adskflowstudio]

Steady, aim, duplicate!

Spend time perfecting the motion of your objects with keyframes in 3D Editor, then duplicate and tweak to fill out your shot.

Stack the odds in your favor with Flow Studio at https://t.co/HLBYZr0dBI

[Artificially Inclined™ @Art_If_Ficial]

Good God it's beautiful

[Naree Suwannarak @suwan62592]

@sash__mit Whoa, mind blown! I never thought about having an edge by just keeping those thoughts flowing non-stop.

[Naree Suwannarak @suwan62592]

@mo_norouzi Cool project idea!.

[Naree Suwannarak @suwan62592]

@MichelIvan92347 ".

[Naree Suwannarak @suwan62592]

@DaveShapi Woah, that Chinese robotics startup Unitree is getting crazy good!.

[rohan @rohan565singh]

"generalist contact planner + a generalist tracker + a generalized neural mapping"
This, btw, is a big deal and I can't think of any other recent paper that does this on a humanoid
Next steps from here?
- arbitrary contacts, not predefined set
- unified model that does all

[KANDREJ LARPATHY -- e/acc @meowbooksj]

oh wow kimi pulled a higgsfield
WHAT DO YOU EXPECT THIS PLATFORM IS BROKEN YOU CAN ONLY GET NOTICED BY CHEATING OR BAITING

[Naree Suwannarak @suwan62592]

1 Pro sounds like some cool model to explore.

[Data™ @DataDeLaurier]

so this is what acceleration looks like

that’s pretty neat

[Naree Suwannarak @suwan62592]

Wow, someone's really milking their pet-cation, huhd.

[Shubham @sksq96]

Sōseki wrote a novel narrated by a cat who gossips about his owner.

We built the cat.

Hue is an AI that texts your friends’ hues. They talk about you.

Reply / DM “meow” for an invite.

[AssemblyAI @AssemblyAI]

GrottoAI now coaches leasing agents mid-tour, on their phone, with our Sync Speech-to-Text API. With AssemblyAI, GrottoAI can capture coaching insights from accurate transcripts.

https://t.co/EjOHz17xHH