Graduate Student Researcher, ILLC
MS in AI, Univ of Amsterdam.
Data scientist.
Fitness Enthusiast.
Speaks fluent sarcasm.
A fitzcarraldo of monachopsis.
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
extremely unprofessional. if kimi wants to make it as a frontier lab, they need to act like one: perhaps silently route people to worse models, and maybe write a blog post about the collapse of humanity
The DiLoCo team at Google DeepMind and Google Research is proud to release Decoupled DiLoCo, the next frontier for resilient AI pre-training.
Decoupled DiLoCo enables training with datacenters across the world, using heterogeneous hardware, and never halting the system despite hardware failures.
It’s a shame what’s happening in Bangalore.
I started my AI Consulting company in Pune but soon left Pune for Bangalore due to this exact regional chauvinism.
We’re a team of 70 people now and as a company pay our fair share of taxes to the Karnataka government. Our team comes from all parts of India and calls Bangalore their home. We also contribute in our own individual capacity to the development of the city by spending and taxation.
I know a few Kannada words but rarely use it in my day to day. Only languages I know are English, Hindi and Odia. I don’t mind learning Kannada but definitely wouldn’t learn if I’m forced.
I’d rather contemplate moving elsewhere if this unabated hooliganism continues.
We’re already tolerating terrible traffic, shitty infra, and unclean neighborhoods.
Is living in peace too much to ask for?
"Scaling laws vs Model Architectures" from @GoogleAI.
Lessons:
- Not all arch scale the same way.
- Vanilla Transformer does pretty well 😀
- Touching the attention too much is "dangerous". 😔
- Perf at base may not translate to large+ scale.
pdf: https://t.co/UgwpzHRNWY
Philips ships new MRI analysis software based on deep learning. Short transfer cycle of technology from FastMRI challenge to product. Congratulations Nicola and team!
It was bound to happen sooner or later: Reviewer on a recent paper claims "the authors don't really understand the contributions of Veličković et al." 😅
New paper with Chris Yu & Henrik Schumacher: https://t.co/SDlisHgmXu
We model 2D & 3D curves while avoiding self-intersection—a natural requirement in graphics, simulation & visualization.
Our scheme also does an *amazingly* good job of unknotting highly-tangled curves!
[1/n]
Happy to announce our work on Coordinate Independent Convolutional Networks.
It develops a theory of CNNs on Riemannian manifolds and clarifies the interplay of the kernels' local gauge equivariance and the networks' global isometry equivariance.
https://t.co/Vb0KHlTJ7S
[1/N]
Tired of rigorous scientific processes? We offer the fastest publication cycle ever with no review (!!) at our @iclr_conf social: Bad Hypothesis Contest. Submit your 5 min fake science contribution:
https://t.co/dJEosWn516
with @YugeTen and @karen_ullrich