DOJO

@lucasmaes_ Its so good bro! going to play around with the models

JEPA are finally easy to train end-to-end without any tricks! Excited to introduce LeWorldModel: a stable, end-to-end JEPA that learns world models directly from pixels, no heuristics. 15M params, 1 GPU, and full planning <1 second. 📑: le-wm.github.io

I've been working on a new LLM inference algorithm. It's called Speculative Speculative Decoding (SSD) and it's up to 2x faster than the strongest inference engines in the world. Collab w/ @tri_dao @avnermay. Details in thread.

@initlayers If you're interested in local attention do read these papers one is "Hiera: Hierarchical transformers without the bells & whistle" then another paper I belive it was window attention is bugged

It's a Saturday afternoon and I'm preparing a presentation to explain to my university what I'm doing during my internship. The funny part is most of the technical details probably won't make much sense to them anyway. Still have to make the slides, simplify everything, and explain months of work in a few minutes. Sometimes that's just how these things go.

@initlayers HR NET is based on multi scale feature extraction while DA Net is based on attention between the channels and spatial features

@Dojo__0 Cool. I can feel you. What are these papers btw?

One thing I've realized while reading a lot of papers: many of them are just careful combinations of ideas that already worked before. When you don't follow the chronology, every paper feels extremely complex and novel. But once you trace the lineage, you start seeing the pattern. Good parts from previous work get reused, tweaked, and sometimes that combination produces surprisingly strong results. The real skill is knowing what to borrow and what to change.

@VazeKshitij Let's go bro!! Ah dude you're killing it!!

So......I can confirm that Deepinder Goyal has seen the comments and the mentions tagging me on the temple hiring post. I just had a conversation with them, and the storm that all of you created over there is the first thing that Team Temple bought up in my interview with em. Deepinder himself was asking if the team knew about my profile and if they were looking into me - I don't think I can put it in words just how damn grateful I am for each and every single last one of you guys man! It went well, lasted for 27-ish minutes. From what I can make out at my end, I did well and according to team temple, it was a good conversation my resume would now go to the technical team, and we'll move on from there. There will be a few more conversations with the tech team, and we'll land on an answer about the further proceedings soon enough. IDK man, I AM STARTING TO HAVE HOPE!!!!

Not even a week in, and we've already hit 400+ members! If you haven't joined yet, you're missing out on some really interesting research project discussions happening on our Discord; link in the comments.

@initlayers Hey, I am interested. I will dm.

We're looking for research assistants or early-career research engineers at the Language Technologies Research Center, IIITH. If you're genuinely motivated to learn and build, this could be you. The work involves designing workflows for data ingestion, working with multiple ML/AI models, and building search and indexing pipelines. Familiarity with cloud infrastructure or building optimized standalone applications is a strong plus. You don't need to know everything already. If you're serious and willing to put in the work, you'll learn on the job. There's proper support, mentorship, and technical training to help you grow while you build. If this sounds like you, reach out.

@prajdabre Something like infoNCE would work good but recently at least in my opinion a simple allignment loss and an anti collapse loss( coding rate regularisation) if we simply don't want a large batch size

Slightly challenging ML question: Suppose you are building an embedding model for RAG where for a document and a related query, you generate embeddings which are as similar as possible. You confidently scrape a bunch of document-title pairs and train a transformer model by maximizing the cosine similarity between the embeddings of the document-title pairs. But this messed up completely. Your embedding model aligned a correct document-title pair during testing but also wrong ones. What happened? What is the fix?

🚀 PhD Internship (will be hosting one person): Diffusion LLMs (DLLM) Looking for PhD students with: • hands-on DLLM research • First author published paper on DLLM (NeurIPS/ICLR/ICML/ACL/EMNLP) Send Google Scholar + a brief introduction → DM RT appreciated 🙏 #NLP #ML

@chhuti_is More than 50 casualties..

I did not find a single tweet on Kolkata warehouse tragedy which killed over 25 people. The site is 15 min dive from his residence.

@Neetivaan GTRE engine ka deal kitna dur hai ?

the fine print of the India - US deal is still not out, so wait until everything comes out in detail

@Seizeyoloflow @Neetivaan capital outlay is used only for buying stuff

@Dojo__0 @Neetivaan It will be used for Salaries, pensions I guess !? Wt do you think

Mega Textile Park, the GoI is planning slow de ath to Kanglus and P0rkis. Bhāratvarsh is here to stay ☀️🥶

@Seizeyoloflow @Neetivaan capital outlay is increased signifcantly +30k crores

@Neetivaan Anything for Defense!?

@Chandan_Perla @primal_brainer I am interested. What's the idea

We’re working on a project that we’re very bullish on Building it with a lot of care We’re looking for operators People who get the vibe and get things done. If this interests you, feel free to drop a message - Let’s talk!

Big find. Altimeter of PAF F-16 Blk 15 (ex RJAF) that was shot down by S400, in May.

@idrwalerts @idrwalerts Dear IDRW team, this work is done by our teammates, please give them credit in place of Indian Reddit User, it took lot of time to get this information and to understand exact part of aircraft 🙏

Anyone who see this tweet of mine 👇 I request you to RT and tag the UP CMO, please I humbly request you .