Love Gradio, a real superpower for machine learning practitioners
Gradio@Gradio
This is built using Gradio 6's super easy new HTML component! You can play with the demo here: huggingface.co/spaces/multimo…
English
Alexandr Notchenko
1.8K posts
Alexandr Notchenko
@Gang1man
Engineer ⋂ Scientist ⋂ Maker Prev: Co-founder and CTO at https://t.co/YweWtzzgM8 PhD grad from @Skoltech Founder of ODS London and @ods_ai
London Katılım Ağustos 2010
4.4K Takip Edilen784 Takipçiler
@haseab_ What you don't like about screen pipe ? github.com/mediar-ai/scre…
Lets collab and find a way to migrate local rewind database into new open source software, last time I checked it was still encrypted.
English
Rewind AI shutting down by December 19th
I've tried alternatives in the past few days, and none of them are as high quality as Rewind
Fuck it, i'm going to create the replacement, and it will be open source
Dan Siroker@dsiroker
I'm excited to share that Limitless has been acquired by Meta! Here’s why we joined forces, what this means for customers, and what comes next.
English
Qwen models are very impressive, love the transparency.
Min Choi@minchoi
Qwen Image Edit w/ Camera Control is wild 🤯 Quickly rotate the camera, switch between bird's eye and worm's eye views using just clicks. Here's how plus 7 wild examples:👇
English
Is it just me or @scikit_learn website and docs are down ? scikit-learn.org
English
@willdepue In UK there is just London, that's it, very convenient.
English
god has cursed everyone cool and attractive to live in new york and everything interesting and important to happen in san francisco
English
Really cool, always wanted to do something like that, hyper specialized train loop for one model type. Need to see how fast it compared to vanilla pytorch implementation.
MrNeRF@janusch_patas
First fully ML-framework-free 3D Gaussian Splatting implementation in LichtFeld Studio. I’ve completed the migration of the full training pipeline to a custom CUDA-based tensor library. No PyTorch, no LibTorch, no autograd. Every gradient is implemented by hand, either through CUDA kernels or minimal abstractions on top. This makes it the first full training setup for 3D Gaussian Splatting with zero dependencies on existing ML frameworks. It’s not just about independence, it's about control! We now manage every byte of GPU memory, which opens the door to tighter optimization and finer performance tuning. The framework footprint is minimal, without pulling in gigabytes of ML runtime code that was never designed for real-time or graphics-driven applications. A few modules, such as the metrics and 3DGUT interfaces, are still being ported, and some operations are temporarily naïve, so performance is not yet on par with master. But this refactor lays the groundwork for: - A fully self-contained binary - Fine-grained memory optimization - Easier experimentation without the weight of an ML stack We’re getting close.
English
Alexandr Notchenko retweetledi
Scaling Elections with GPUs and Mojo 🔥
I love occasionally optimizing obscure algorithms no one cares about. One of those, also having O(n³) complexity, like matrix multiplication, is the Schulze method for ranked-choice voting. So last year, at one of the AGI hackathons I tried accelerating it across CPUs and GPUs, with CUDA C++, Numba, and Mojo 🔥
It was an epic event and I really wanted to share the results sooner, but without AMD GPUs they didn't feel complete! We are talking about plurality of choice, after all! Now, after gathering AMD MI355X results - the story is out. The Mojo variant didn't require a single line change, ran out of the box, and at scale delivered 4x the throughput of H100.
Don't take the article too seriously. It's just a fun exploration of modern GPU tech beyond LLMs — tropical semirings, blocked Floyd-Warshall, and the participation paradox in voting!
Blogpost: ashvardanian.com/posts/scaling-…
Code: github.com/ashvardanian/S…
Overall it was one of the best cozy gatherings that year - @clattner_llvm talked about Mojo's GPU support, @RajaXg explained the reasons for the success of the CUDA ecosystem in the last decade, @tri_dao gave the first public talk on FlashAttention-3... now superseded by his FlashAttention-4, and @dylan522p shared some insights on the data-center market!
Thanks to @bztree, @verdagon, and Pradeep Ramani for collaborating with me on this one! Thanks to @JvNixon, @khoomeik, and @kylejohnmorris for bringing the AGI house together and hosting all of us and to @nebiusai for providing the compute! Can't wait to gather again!
English
Thank you for your work. It inspired a lot of researchers throughout the years including me. Even less popular papers like OpenDR - made a lot of impact down the line, inspired me to work on differentiable rendering in my PhD. Many more years of great papers and code!
Michael Black@Michael_J_Black
SMPL is 10 years old and has done what we hoped — it changed the way the field estimates and models 3D humans and their motion. I’m delighted that the original team has been recognized today at @ICCVConference with the Mark Everingham Prize. The prize is given to individuals or teams who have worked to further progress in the computer vision community as a whole. Mark Everingham understood that to have an impact, it is not enough to simply publish a paper. SMPL’s success is due to lots of hard work to provide the community with code, data, and support. My deepest thanks go out to all the members of the @PerceivingSys department who have supported SMPL and related technology over the years. It has been a team effort of many dedicated people and we share this award with you. Mark understood that big changes require community effort. Consequently my big thanks go to all the users of SMPL and related tools. You have pushed the field forward as a community in ways that no small team could. I’m constantly inspired by your work. Computer vision has changed a lot in 10 years but people keep finding new uses for SMPL, most recently in training humanoid robots. There are many more applications to come in games, interactive entertainment, sports, and biomechanics. Congratulations to my coauthors Matt Loper, @naureenmahmood, Javier Romero and @GerardPonsMoll1. smpl.is.tue.mpg.de
English
@DrZeeshanZia Not even taking into account reduction in quality of service or direct harm caused by deploying half baked systems in civic service, healthcare and engineering.
100% a lot of harm happening right now.
English
Far too many people get hung up on the Sky Net / Terminator scenario, where there are plenty of harms taking place today:
- kids have committed ChatGPT assisted suicide.
- AI unlocks a new level of ransomware by enabling hackers to go through your data (e.g. identify cheating spouses etc)
- gives ability to low-skilled people to build complex weapons and engage in terrorism
I could go on and on. These are immediate harms that technologists can tackle.
Harj Taggar@harjtaggar
Those who can, do; those who can’t, get really into AI safety.
English
Vibe engineering is a misnomer, it should be called stochastic coding.
English
@iamgrigorev Make Screen recording too and put it on youtube, will be great way to log the process and be very educational to others.
English
I am thinking of writing the next blogpost about these topics:
Optimizing training throughout with FP8
I will show how to write FP8 kernels
How to implement DDP
How to implement FSDP, with distributed Muon
How to implement TP
Gradient accumulation
Gradient checkpointing
I think using “How to scale your model” on a real consumer gpus connected with pcie and writing that from scratch on pure PyTorch would be really useful
English
@AndrewMohawk I used their email forwarding mechanic for my domains, it worked fine. Is it going to be a Gmail clone or a MailChimp clone?
English
What’s the worst that can happen blog.cloudflare.com/email-service/
English
@gwulfs Almost exactly what I did in my startup 3 years ago, data and artefact store in R2, and taking the cheapest servers anywhere in the world to train on, at the time it was lambda labs and paperspace.
English
Good time to start investing into HDMI capture stick with AI on board. Because this is getting ridiculous.
Peter Steinberger 🦞@steipete
Wanted to make a screenshot of gemini as it wrote "Rewriting in Rust for no particular reason..." but seems macOS 26 has some new security features... or AGI doesn't want to be mocked.
English
Great selection, I would also add Jaynes - Probability Theory: The Logic of Science.
yingzhen@liyzhen2
I read the first 4 books extensively during my PhD, highly recommended 👍 I'd also highlight the 5th book as my first read re deep learning. Mind-blowing for a young math undergrad (me) at the time, made me decide to go for ML
English
Mostly agree with takes in this blog (blog.jxmo.io/p/there-are-no…) about role of the datasets in advancement of ML models over the last decade and a half. But I would disagree on advancement claims, before robotics and formal verification a lot of "progress" was from learning labels.
English
@y0b1byte Pseudo-code in the paper, python code listing into appendix, go back to classic approach. Also if you know there was a major war in 2010s for the appropriate research artefact storage. Figshare, ResearchGate, Academia.edu, but the true way is Uni hosting of course.
English
Alexandr Notchenko retweetledi
We are thrilled to support the UK’s new #OpenBind consortium, a government-backed initiative to create the world’s largest open dataset for AI drug discovery.
English