Ankit Khandelwal

@paulg Bullet points are great. They naturally provide logical breaks. I don't think optimising for whitespace provides any benefits.

The fact that AIs tend to answer you in bulleted lists tells us something important, though somewhat depressing: people can't read. They don't do this by accident. What you're seeing is an implicit portrait of the median user.

@HarveenChadha No human can write in one shot either. I would bet on divide and conquer.

all models suck at long form writing, how tf are people writing books with LLMs to get a first draft, 10k-15k words is the ideal limit

@lossfunk Whenever I see any post starting with the word 'Shocking', my mind says it's a clickbait.

🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵

@luke_metro Most of them have no clue how that data is used 🤣

Nvidia GTC 2026:

@eddybuild x.com/ankk98/status/…

today, we’re open sourcing the largest egocentric dataset in history. - 10,000 hours - 2,153 factory workers - 1,080,000,000 frames the era of data scaling in robotics is here. (thread)

@rohanpaul_ai @huggingface x.com/ankk98/status/…

Egocentric-10K, the largest egocentric dataset from Build AI just dropped on @huggingface a real factory first-person video dataset with 10,000 hours, 1.08B frames, and 2,138 workers, licensed Apache 2.0 and hosted on Hugging Face. Egocentric here means the camera is on a worker’s head so the model sees hands, tools, and objects from the doer’s view, which is the right signal for learning manipulation and step-by-step tasks. prior large egocentric sets like Ego4D had around 3,000+ hours in daily life scenes rather than production lines, so this is a big jump in both scale and domain specificity for factories. The dataset emphasizes high hand visibility and dense manipulation, which are exactly the pixels imitation learning and visuomotor policies need to map observations to actions. Clips are 1080p at 30 fps, organized as WebDataset shards with paired JSON metadata, totaling 16.4 TB, and they can be streamed directly via Hugging Face for training without full downloads. An evaluation split with 30,000 frames is also provided to benchmark hand and object interaction tasks in this domain.

@eddybuild x.com/ankk98/status/…

today, build ai is open sourcing the largest egocentric dataset in the world - 100,000 hours - 14,228 workers - 10,800,000,000 frames each pixel is a unique video (try below)

@ariG23498 @PyTorch This is really promising and will significantly reduce learning curve and entry barrier.

When you run a @PyTorch model on a GPU, the acutal work is executed through kernels. These are low-level, hardware-specific functions designed for GPUs (or other accelerators). If you profile a model, you'll see a sequence of kernel launches. Between these launches, the GPU can sit idle, waiting for the next operation. A key optimization goal is therefore to minimize gaps between kernel execution and keep the GPU fully utilized. One common approach is `torch.compile`, which fuses multiple operations into fewer kernels, reducing overhead and improving utilization. Another approach is to write custom kernels tailored to specific workfloads (e.g., optimized attention or fused ops). However, this comes with significant challenges: > requires deep expertise in kernels writing > installation hell > integration with the model is non-trivial To address this,@huggingface introduces the `kernels` library. With this one can: > build custom kernels (with the help of a template) > upload them to the Hub (like models or datasets) > integrate them to models with ease Let's take a look at how the transformers team use the kernels library to integrate it into the already existing models. (more in the thread)

More Info: dev.to/ankk98/kriya-e…

Introducing **Kriya-Egocentric-100K**: Action100M-style, fully automatic video action annotations for a 5-video preview of Build AI’s Egocentric-100K dataset on Hugging Face. First-person manual-labor videos (head-mounted fisheye camera) with hierarchical temporal segments + LLM-generated captions & GPT summaries. Super promising for video understanding, robotics, world models & physical AI! huggingface.co/datasets/ankk9… #EgocentricVision #ActionRecognition #VideoAI #LLM #HuggingFace #PhysicalAI #Robotics

More Info: dev.to/ankk98/kriya-t…

Introducing Kriya-EPIC-KITCHENS 🎥 – Action100M-style, automatic video action annotations for a preview subset of EPIC-KITCHENS-100 on Hugging Face. Results on egocentric kitchen videos look very promising for downstream video understanding. huggingface.co/datasets/ankk9…

@VraserX x.com/ankk98/status/…

V-JEPA 2 is Meta’s new world model. Trained on 1M+ hours of video + robot data, it doesn’t predict pixels but abstract states. That means it learns physics, causality, anticipation. Less guessing frames, more understanding reality.

@AIatMeta x.com/ankk98/status/…

Introducing V-JEPA 2, a new world model with state-of-the-art performance in visual understanding and prediction. V-JEPA 2 can enable zero-shot planning in robots—allowing them to plan and execute tasks in unfamiliar environments. Download V-JEPA 2 and read our research paper ➡️ ai.meta.com/vjepa/

@TheTuringPost @AIatMeta x.com/ankk98/status/…

A new comer in the family of world models — V-JEPA-2 By combining 1M+ hours of internet videos and a little bit of robot interaction data, @AIatMeta built an AI that can: • Watch • Understand • Answer questions • Help robots plan and act in physical world V-JEPA 2 shows true success of self-supervised learning and efficient scaling of everything. Here is how it actually works:

@omarsar0 x.com/ankk98/status/…

NEW: Meta releases V-JEPA 2, their new world model! Foundation world models aim to accelerate physical AI, the next frontier. Why is this a big deal? Let's break it down:

It was an information-packed talk yesterday, full of insights. Worth watching multiple times. @lossfunk @viraataryabumi

@randal_olson x.com/ankk98/status/…

EPIC-Kitchens: A large dataset of first-person, multi-faceted, non-scripted recordings in native environments. Looks like a gold mine for #MachineLearning research. #DataScience epic-kitchens.github.io/2018

@_akhaliq x.com/ankk98/status/…

The EPIC-KITCHENS Dataset: Collection, Challenges and Baselines pdf: arxiv.org/pdf/2005.00343… abs: arxiv.org/abs/2005.00343 webpage: epic-kitchens.github.io/2020

This version of the API focuses on replicating the Action100M pipeline. An improved pipeline that builds on it while addressing its shortcomings is coming soon.

This is a small preview (6 EPIC videos) to show what large-scale automatic video action annotation can look like. Excited to see how this feeds into video world models, VLM/VLA pipelines, and physical/robotic AI systems. #EPICKITCHENS #ActionRecognition #ComputerVision #LLM #VideoAI #EgocentricVision #HuggingFace #WorldModels #VLM #VLA #PhysicalAI #RoboticsAI