iiinova — where your digital self keeps evolving

Reposted by Vlad Niculae

Annabelle Michael Carrell @ab-carrell.bsky.social · 14/07/2025

Your data is low-rank, so stop wasting compute! In our new paper on low-rank thinning, we share one weird trick to speed up Transformer inference, SGD training, and hypothesis testing at scale. Come by ICML poster W-1012 Tuesday at 4:30!

2257

Reposted by Vlad Niculae

Annabelle Michael Carrell @ab-carrell.bsky.social · 19/07/2025

So you want to skip our thinning proofs—but you’d still like our out-of-the-box attention speedups? I’ll be presenting the Thinformer at two ICML workshop posters tomorrow! Catch me at Es-FoMo (1-2:30, East hall A) and at LCFM (10:45-11:30 & 3:30-4:30, West 202-204)

054

Reposted by Vlad Niculae

justine zhang @tisjune.bsky.social · 27/06/2025

new working paper! we (me, Su Lin Blodgett, @ninamarkl.bsky.social) examine how recent marketing of LLMs extends older discourses that cast workers as bundles of skills, and unpack the false promises of empowerment these discourses embed, in times of precarity tisjune.github.io/papers/aarhu...

tisjune.github.io

3353

Reposted by Vlad Niculae

Andreas Vlachos @andreasvlachos.bsky.social · 06/06/2025

Looking forward to this year's edition! With great speakers: Ryan McDonald Yulan He @vn-ml.bsky.social @antonisa.bsky.social Raquel Fernandez @annarogers.bsky.social Preslav Nakov @mohitbansal.bsky.social @eunsol.bsky.social Marie-Catherine de Marnefffe !

063

Reposted by Vlad Niculae

Kacper Wyrwal @wyrwalkacper.bsky.social · 13/02/2025

Excited to share our ICLR 2025 oral "Residual Deep Gaussian Processes on Manifolds"! With @vabor112.bsky.social & @arkrause.bsky.social, we introduce manifold-to-manifold GPs that can be composed together, generalising deep GPs to manifolds. Applications include wind prediction & Bayes opt! 1/n

Schematic illustration of a scalar-valued residual deep GP with L hidden layers. The last layer is a scalar-valued GP on the manifold. If it is not present, the model is manifold-valued. If it is replaced with a Gaussian vector field (GVF), the model is a vector field on the manifold.

1379

Reposted by Vlad Niculae

CJ @virmalised.us · 04/01/2025

i can't believe how long we've spent fooling ourselves about the value of fully specified, massive matmuls instead of embracing the gods of sparsity

052

Vlad Niculae @vn-ml.bsky.social · 24/04/2025

Recruiting a PhD candidate at U. of Amsterdam (funded, 4yr). We will use ML&NLP, prob. models, and user studies, to make adaptive scientific-assistant systems that communicate & justify decisions in ways helpful to experts. More: vene.ro/jobs.html Apply by May 18: werkenbij.uva.nl/en/vacancies...

vene.ro

Vlad Niculae

0116

Reposted by Vlad Niculae

Alex Thiery @alexxthiery.bsky.social · 20/11/2024

Variational approximation with Gaussian mixtures is looking cute! So here it's just gradient descent on K(q||p) for optimising the mixtures means & covariances & weights... @lacerbi.bsky.social

2337

Reposted by Vlad Niculae

Gabriel Peyré @gabrielpeyre.bsky.social · 29/01/2025

This review paper by @guillaume-garrigos.com on SGD-related algorithms is a fantastic resource, offering elegant, self-contained, and concise proofs in a single, accessible reference. arxiv.org/pdf/2301.11235

118940

Reposted by Vlad Niculae

Mark Riedl @markriedl.bsky.social · 14/12/2024

These phenomenon have been observed since early vision systems. It is important to report these things, though. Maybe it will permeate and we won’t keep making the same mistakes over and over

1123

Reposted by Vlad Niculae

Clément Canonne @ccanonne.github.io · 14/12/2024

This is such a beautiful algorithm (and a nice analysis): to check if an array is sorted vs. far from being sorted (many entries need to be changed), just: - pick an element uniformly at random in the array - "forget" where it was - try to find it again via binary search Repeat this a few times.

0286

Reposted by Vlad Niculae

justine zhang @tisjune.bsky.social · 06/12/2024

"AI can be bad but also it can be good" is just a really dumb way to talk about anything...it's the grade-school exercise of "make a list of pros and cons" but pressed into service for producing a sense of inevitability and making the medicine go down

141

Reposted by Vlad Niculae

Mark Riedl @markriedl.bsky.social · 04/12/2024

OpenAI in 2024: “No AI for weapons or military” “Do use our AI to make weapons to hurt yourself or others” “Military is fine, but no AI for weapons” “Sure put it on battlefield drones” www.technologyreview.com/2024/12/04/1...

technologyreview.com

OpenAI’s new defense contract completes its military pivot

A new partnership with Anduril, announced today, will deploy AI on the battlefield. It represents an overhaul of the company’s position in just a year.

619971

Reposted by Vlad Niculae

Jack Hessel @jmhessel.bsky.social · 03/12/2024

Blue skies 🦋 , hot (?) takes 🔥 Constrained output for LLMs, e.g., outlines library for vllm which forces models to output json/pydantic schemas, is cool! But, because output tokens cost much more latency than input tokens, if speed matters: bespoke, low-token output formats are often better.

281

Reposted by Vlad Niculae

Hellina Hailu Nigatu @hellinanigatu.bsky.social · 02/12/2024

I hope I am not late to the party (was away post-quals chilling) but here are some thoughts on why this is bad IMO: First, a disclaimer that I am writing this as an African who is a speaker of multiple African languages, NLP researcher of African languages, and HCI researcher focusing broadly on..

912560