Alex Garabito

AI slop researcher

[Guillermo Garabito @GuilleGarabito]

Amazing! Idol by @speridlabs is now #1 trending on @huggingface! 🚀

https://t.co/PJlm2C9f3O

[Alex Garabito @alexgarabt]

@adelavega__ ahahahahahsh

[Alex Garabito @alexgarabt]

i guess you can just do things

[Speridlabs @speridlabs]

More downloads than Mistral, Pika and Kimi models is crazy, Idol is the 2nd in trending in @huggingface

[Alex Garabito @alexgarabt]

@chema_garabito @speridlabs @AleexCah dont forget us when you are famous

[Chema Garabito @chema_garabito]

Idol from @speridlabs is 2nd in huggingface trending

Has been downloaded more times than Kimi, Mistral and the new video model of Pika, that is crazy

[Alex Garabito @alexgarabt]

today we got boomer AI

im sure every mum and grandma will love it

[OpenAI @OpenAI]

Images, upgraded—introducing ChatGPT Images 2.6.

Now with GIFs and reference images.

Plus, two new models are coming to the API with 50% faster image generation.

[Alex Garabito @alexgarabt]

@koguGameDev we are working on also outputting meshes and multi view conditioning for the next release.

but is so cool to see your demos, really great work!

[fal @fal]

Idol by Sperid Labs is now available on fal

Turns a single image into a 3D Gaussian splat in about 10 seconds
Choose the density, from 32K up to 1M Gaussians
Exports a .ply splat plus a .glb preview, for photos and renders

[Alex Garabito @alexgarabt]

@migero_usausagi @YeheLiu @lichtfeldstudio https://t.co/aN0bQ8zNBB

[Chema Garabito @chema_garabito]

@SPabolu @speridlabs not yet, but it will, stay tuned!

[Alex Garabito @alexgarabt]

@the_ember_mind @speridlabs yeah, that’s tralalero tralala

[Guli Moreno @GuliMoreno]

Spoiler: it's working 🤯. All made from pictures fetched from the @IKEASpain website and then passed through @speridlabs IDOL via @huggingface agents.

All done by Opus 5.5 - I didn't actually go into Ikea or Sperid or HF.

[Alex Garabito @alexgarabt]

@adelavega__ 🗿🗿🗿

[Alex Garabito @alexgarabt]

@chema_garabito @AleexCah developed this just so I could create more brainrot

[Alex Garabito @alexgarabt]

Okay, now ai slop can be 3d too.

[Speridlabs @speridlabs]

Meet Idol, an open image-to-3D model. A 3D-DiT that generates complete 3D objects from just an image.

Pre-trained from scratch, contains only 600M parameters, it runs on consumer GPUs, using less than 7.2 GB of VRAM and can generate 1M splats in 3 seconds. Try it today.

Fully open: weights, code and demo (Apache 2.0)
https://t.co/Cf0slnSI5C

[Alex Garabito @alexgarabt]

try it!! i just generated this ones. @AleexCah is a genius

[Speridlabs @speridlabs]

Meet Idol, an open image-to-3D model. A 3D-DiT that generates complete 3D objects from just an image.

Pre-trained from scratch, contains only 600M parameters, it runs on consumer GPUs, using less than 7.2 GB of VRAM and can generate 1M splats in 3 seconds. Try it today.

Fully open: weights, code and demo (Apache 2.0)
https://t.co/Cf0slnSI5C

[Alex Garabito @alexgarabt]

@AleexCah that funko pop is mogging you

[Speridlabs @speridlabs]

Meet Idol, an open image-to-3D model. A 3D-DiT that generates complete 3D objects from just an image.

Pre-trained from scratch, contains only 600M parameters, it runs on consumer GPUs, using less than 7.2 GB of VRAM and can generate 1M splats in 3 seconds. Try it today.

Fully open: weights, code and demo (Apache 2.0)
https://t.co/Cf0slnSI5C

[Alex Garabito @alexgarabt]

@PyTorch @speridlabs @nvidia Hi thanks for answering, already seeing movement in the issue, thanks!

https://t.co/EXBegKqgN3

[Alex Garabito @alexgarabt]

I have contacted @nvidia to work out the proper fix.

I opened an issue in @PyTorch https://t.co/9n8l0lD1JH

& wrote a mitigation for pytorch, a byte-span guard before cuDNN's backward
https://t.co/w4W3rTvKFv

Meanwhile i recommend pass the incoming gradient through .contiguous() or force FLASH_ATTENTION.

The full story is here https://t.co/XIyoi8SEMY

[Alex Garabito @alexgarabt]

I found a bug in @nvidia cuDNN's attention backward (CWE-190) that silently breaks large scale training on default @PyTorch cu130 wheel. No crash, past a certain tensor size the model stops learning and starts to unlearn. It happened with one of our models @speridlabs

+

[Jon Barron @jon_barron]

Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:

[Alex Garabito @alexgarabt]

Every cuDNN I tested, 9.10 → 9.27, fails with the exact same fingerprint. Deterministic, same number on every machine.

And @PyTorch cu130 wheel defaults to cuDNN 9.13

Affected archs are:
sm_103 (B300), sm_90 (H100/H200), sm_80 (A100), sm_86, sm_120/121
Other architectures are untested and may be too.

All results and a reproducer you can run yourself: https://t.co/1fLJj5q3xp