Am I the only one who when vibe-codes feels like "the LLM whisperer"?

Transfer learning and self-supervised learning have changed how we build neural architectures. Reusing representations from one task saves time and data, but tuning for a new problem is never plug-and-play. Always check what actually transfers. 🤔 #ItzikThoughtLoop

Since it resonated with the audience, I’ll recap my main argument against AGI here. ‘General intelligence’ is like phlogiston, or the aether. It’s an outmoded scientific concept that does not refer to anything real. Any explanatory work it did can be done better by a richer scientific frame. 1/3

How AI Taught Itself to See Self-supervised learning is fascinating! How can AI learn from images only without labels? In this video, we’ll build the method from first principles and uncover the key ideas behind CLIP, MAE, SimCLR, and DINO (v1–v3). Video link: youtu.be/oGTasd3cliM

In a sort-of social experiment, instead of assigning "reviewers" to the submitted papers, we assigned "mentors". Although the job content was the same, the quality of the reviews and the satisfaction of the authors improved quite significantly (even those whose paper was rejected). Think about it.

Meta brought AI to rural Colombia. Now students are failing exams restofworld.org/2025/colombi...

The system is completely broken. blog.neurips.cc/2021/12/08/t... Review results are random and affect the achievement record of thousands of researchers globally. If review quality cannot be controlled the responsibility would fall on ACs to recheck flagged papers, which doesn't seem feasible either

Reinforcement Learning Teachers of Test Time Scaling arxiv.org/abs/2506.08388 We introduce a new way to teach LLMs how to reason by learning to teach, not solve! Here, a teacher model is rewarded based on how effectively its explanations help the student model recover correct solutions.

Guys … I’m all for critiquing the way LLMs are implemented in society and debunking folks who make wild claims that they’re sentient. But definitively claiming the brain doesn’t use statistical learning is not the right take. Don’t be armchair neuroscientists please. Seeing some wild opinions today.

As I see valuable work being rejected all the time, I tend to think that: - Good reviewers feel like accepting the papers, and try to find enough reasons to do so - Bad reviewers feel like rejecting the papers, and try to find enough reasons to do so They may seem equivalent but oh they are not

I wish reviewers would stop seeing ablation studies as a chance to reject a paper, emphasizing the impracticability of the proposed method since its accuracy is unstable when conditions vary... it's an ablation study you dummy, that's what's supposed to happen! 🤦‍♂️ #only_in_topcvconf #bully_reviewers

These boxes are not moving. A mind-bending optical illusion by Japanese artist Jagarikin.

[Please share🔄] Did you read blog posts / preprint about #ICLR2025 LLM experiment? Were you in the #ICLR2025 review process? As I was not in the process, I'm asking for opinions to improve review systems Thank you for your help🙏 🤗 huggingface.co/blog/yoshito...

Verified my Bluesky account🦋 I was hesitant to do that, but TIL the following update > Update as of December 12, 2024: If you change your default Bluesky username (with the .bsky.social suffix) to a website URL, your old .bsky.social username will be reserved for you. bsky.social/about/blog/4...

In the multi-object setting, objects that appear earlier in the caption or that are larger in size are prioritized by the CLIP encoder. arxiv.org/abs/2502.19842

I got to test the LLM "help" they provided and it was quite useful! It suggested being more specific on key areas and I thought I genuinely helped me improve the quality of my reviews, without changing my original assessment in any way. Imho it was very well tailored.

Credibility, not capability. The most important thing we build in technology and academia is not capability, but credibility. It does not matter how fast we calculate, how smart we are, or the number of products or papers we make, if we cannot answer "Why should anybody believe anything we say?"

🧠 Keeping LLMs factually up to date is a common motivation for knowledge editing. But what would it actually take to support this in practice at the scale and speed the real world demands? We explore this question and really push the limits of lifelong knowledge editing in the wild. 👇

Visual Geometry Grounded Transformer (VGGT) predicts cameras, point maps, depth maps, and point tracks for up to hundreds of images in less than a second on a H100 GPU. github.com/facebookrese...

I don't think I can go back to surveying papers without LLM summaries for quick filtering. But the fact that there's people who've never done a survey without an LLM kind of worries me. Like people who've never made a sum without a calculator or spoken a foreign language without a translation tool.

Similar vibe 😅

"our politics is split between a left that defends government even when it doesn't work, and a right that wants to destroy government even when it is working" www.youtube.com/watch?v=Vwjx...

IMHO, AI doesn't do jack. People who want to profit from it do LOL Really, we need to stop using the lazy word AI and talk about specific people, technologies and business models. AI doesn't mean anything, I don't use it on my CV anymore and even stop listening when it pops up in a convo.

Interesting paper that studies the tradeoff between memorized vs. retrieved knowledge in LLM-RAG frameworks. Quantifying Memorization and Retriever Performance in Retrieval-Augmented Vision-Language Models (arxiv.org/abs/2502.13836)

Introducing The AI CUDA Engineer: An agentic AI system that automates the production of highly optimized CUDA kernels. sakana.ai/ai-cuda-engi... The AI CUDA Engineer can produce highly optimized CUDA kernels, reaching 10-100x speedup over common machine learning operations in PyTorch. Examples:

Are people actually using random names for their repos?

An uncensored version of R1 is released 🔥 “R1 1776 is a DeepSeek-R1 reasoning model that has been post-trained by Perplexity AI to remove CCP censorship. The model provides unbiased, accurate, and factual information while maintaining high reasoning capabilities.” huggingface.co/perplexity-a...

This well-known problem is not being dealt with properly. A tutorial in SIGIR2024 proposed a framework in which a model masters fundamental reasoning tasks with basic curated data, without making it learn tons of Internet sh*t. Same as humans do at school. The rest is information retrieval.

Are there any machine learning datasets/benchmarks/competitions where k-nearest neighbors achieves comparable error rates to the top-scoring models?

This benchmark paper on Multimodal Retrieval-Augmented Multimodal Generation provides interesting insight on the task: - Challenges in long-text, high-image-density tasks - The image ordering constraint proves to be an unsolved challenge

It seems like a no-brainer, but someone will try to deploy them anyway 🤷

It is hard to overstate how cool and powerful is flex attention. @chhillee.bsky.social pytorch.org/blog/flexatten… TL;DR: it is an implementation of the attention operator in pytorch that allows in particular to efficiently "carve" the attention matrix. 1/3

I was using an LLM in order to get ideas on ways to implement certain functionalities, and while I was reading the results and nodding, I started feeling bad about how little I was using my brain in the process (like, “I should’ve come up with this stuff”)

As the models get better, the naming schemes deteriorate into madness. 😅

Reducing the mutual information from the multimodal embeddings for different classes yields disentanglement and accuracy improvement Disentangling CLIP Features for Enhanced Localized Understanding (arxiv.org/abs/2502.02977)

A better way of encoding images and text in Large Vision Language Models (LVLM) arxiv.org/abs/2502.01906 By the way, do you say LVLM or Multimodal Large Language Models (MLLM)? I don't think there's a clear naming convention 🤷

Today's the RAG (retrieval-augmented generation) day! - SafeRAG: Benchmarking Security in Retrieval-Augmented Generation of Large Language Models (arxiv.org/html/2501.18...) - RealRAG: Retrieval-augmented Realistic Image Generation via Self-reflective Contrastive Learning (arxiv.org/html/2502.00...)

Indeed, those systems are treated "as is", trying to diagnose the validity or whatnot of their outputs, without caring about what's really going on and how to improve them.

A method for retrieval augmented generation that adds robustness against irrelevant information: 1. Defects Detection: Evaluating the existence of misinformation in the retrieval results. 2. Utility Extraction: Generation of correct answers even from defective inputs. arxiv.org/abs/2501.18365

A few reflections I had while watching this interview featuring Geoffrey Hinton: www.youtube.com/watch?v=vxkB... 1/8

New for #ICCV2025 🌶️ iccv.thecvf.com/Conferences/...

A simple but valid baseline for multimodal RAG (retrieval-augmented generation) arxiv.org/abs/2501.15470

MATCHA:Towards Matching Anything Fei Xue, Sven Elflein, @lealtaixe.bsky.social, Qunjie Zhou tl;dr: diffusion model->semantic+geometric features->transformer-based fusion->enhanced diffusion features->w/ DINOv2->unified feature->geometric/semantic/temporal matching arxiv.org/abs/2501.14945

I'm a bit split... what I don't get is that people are crazy about AGI and chatbots (this does not make sense).. but all the things people don't talk about, especially progress in code generation, RAG, agents, is real and will change the world of programming and compute ... but nobody cares.

I'm maximally confused today... I had a hard time processing that AI companies were fantasizing about how much compute they will need to train AGI... completely delusional, but ok. But that you can crash the market by fantasizing about how little GPU you use... I couldn't have made this up ever.

Germans are so allergic to rankings that they'll insist on all German universities being of equally high quality. Without knowing any better, I based my choice of undergrad uni on this misinformation / "wishful thinking" and ended up in the most unprofessional institution I've ever encountered.

Estoy intentando crear un Starter pack para "investigadores españoles en visión por ordenador", pero no llego al mínimo de miembros necesarios. ¡Si alguien puede echar una mano, lo volvemos a intentar! @andreugirbau.bsky.social @noagarciad.bsky.social