prev founding ML @Lancedb co-created yolov8/ @ultralytics framework , led oss channel @weights_baises
India
Joined August 2016
- Tweets1.2K
- Following465
- Followers1K
- Likes4.3K
Pinned Tweet
Quick announcement- I no longer maintain yolo or work on adding features & answering issues. I appreciate all the lovely DMs and emails and I try not to ignore any. Also, I haven't kept up with latest developments from past months so I can't really answer those anyways. Thanks!
Ayush Chaurasia retweeted
Feature of the Week: Column Reads Merged Into One Request
Mostly null columns compress to a few hundred KB, but each paid its own round trip. Now @lancedb sorts reads from every column into 1 request, merging neighbors into fewer GETs.
Fewer round trips to S3 on full scans ⚡
Ayush Chaurasia retweeted
Some nice benchmarking of Jev on reranking for IR problems!
This post compares Jev as a reranker on a few common IR datasets.
TLDR:
1. Its a good default reranker (based on cost and speed)
2. Its definitely not a silver bullet. Its not a drop-in replacement for rerankers across all workloads
3. Sensitive to prompt (as expected). Customising prompt for a dataset yeilds better perf.
lancedb.com/blog/how-jev-com…
This post compares Jev as a reranker on a few common IR datasets.
TLDR:
1. Its a good default reranker (based on cost and speed)
2. Its definitely not a silver bullet. Its not a drop-in replacement for rerankers across all workloads
3. Sensitive to prompt (as expected). Customising prompt for a dataset yeilds better perf.
lancedb.com/blog/how-jev-com…
Ayush Chaurasia retweeted
Feature of the Week: Batch Vector Search Now Shares IVF Partition Scans
In a batch, many queries want the same IVF partitions, but Lance was re-reading them per query. Now it loads each partition once and scores every query that wants it.
Same answers, ~2.5x faster 🚀
@lancedb
stable-worldmodel got into neurips 2026!
it's basically everything we keep rebuilding for world model research (data loading, baselines, planning, evals) in one place so nobody has to do it again
arxiv.org/abs/2605.21800
Ayush Chaurasia retweeted
Love seeing @lancedb and @huggingface working together. Opensource ftw!
LeRobot now natively supports @lancedb datasets, with fast training and global shuffling directly from HF Storage Buckets - no need to download the dataset first 🚀
From vector & full-text search to curation & mining, discover what you can unlock with LanceDB + LeRobot 👇
Really excited about this one with @huggingface @LeRobotHF
the interesting part of LeRobot + LanceDB is that you can keep the same LeRobot workflow as datasets scale, while training directly from object storage + using vector/FTS to curate and mine the same data.
Thankyou @Caro_Nahkriin and @lhoestq for collaborating on this!
LeRobot now natively supports @lancedb datasets, with fast training and global shuffling directly from HF Storage Buckets - no need to download the dataset first 🚀
From vector & full-text search to curation & mining, discover what you can unlock with LanceDB + LeRobot 👇
Ayush Chaurasia retweeted
LeRobot now natively supports @lancedb datasets, with fast training and global shuffling directly from HF Storage Buckets - no need to download the dataset first 🚀
From vector & full-text search to curation & mining, discover what you can unlock with LanceDB + LeRobot 👇
Ayush Chaurasia retweeted
The best kind of work almost always comes from collaborations!
This time @loldedxd helped me figure out the internal workings on funes. While I knew how to make funes work, he made sure I understand how things run on my system.
It is a fun 30 mins read.
huggingface.co/blog/ariG2349…
Ayush Chaurasia retweeted
Raw text → curated data → 2.43B tokens → trained GPT-2 in 25 minutes on 8 H100s.
@loldedxd built the entire pretraining pipeline on one LanceDB table, streaming + globally shuffling directly from S3 at 3.16M tokens/sec.
Full global shuffle, not a pre-shuffled copy.
na2.hubs.ly/H07Ssr40
Ayush Chaurasia retweeted
Lance is now natively supported in 🤗 @huggingface LeRobot.
Ingest through train directly from S3, GCS, or HF Storage Buckets with the same LeRobotDataset API and true global shuffle. No need to download the full robotics dataset first.
Big thanks to @Caro_Nahkriin and @lhoestq for working with us to get this merged! na2.hubs.ly/H07PyKN0
Ayush Chaurasia retweeted
@LeRobotHF is opening up to new dataset formats 🤩
You can now train directly from a @lancedb dataset, no conversion needed.
Pair it with HF Storage Buckets and go straight from storage to training 😎
Lance is now natively supported in 🤗 @huggingface LeRobot.
Ingest through train directly from S3, GCS, or HF Storage Buckets with the same LeRobotDataset API and true global shuffle. No need to download the full robotics dataset first.
Big thanks to @Caro_Nahkriin and @lhoestq for working with us to get this merged! na2.hubs.ly/H07PyKN0
Ayush Chaurasia retweeted
Exciting news from Jasper Research! 🚀
We built an internal 4B text-to-image model from scratch that turned out to be one of the fastest models out there!
1k image generation in less than a second!
Our technical report shares everything we learned along the way.
👇
Ayush Chaurasia retweeted
Growth doesn't just solve a startup's problems in the sense of making them go away. It solves them like you solve a math problem. For example, it can design your product for you: start with something minimal, then add whatever will make you grow.
Ayush Chaurasia retweeted
Length is easy when you ignore quality. I can generate random pixels for 90 minutes and claim to be a movie director too and call it H3-ultra-lightning-long-gen. But this is not how people should innovate on length in the open.
After the first live stream AI created by @rehan_shei and @alexgmatt , here the first movie of 1h30 !!! fully generated by itself with consistent style and consistent character, much more to come on @fal !!!
would need a lot of edit and cut, but I wanted to show the raw result. I am working much more now on consistent environment as well to make this much better.
To give your more insight this is based on minimax h3 with a train lora that was only train on 15s of data ! So a single movie could allow you to generate a full movie or much more. Working on that and if I see people interested, I could make it available !
Ayush Chaurasia retweeted
oh, thanks for evaluting this for us. We wish we had more compute to distill with more data (like you would have), but it is happy to see that, with our very limited dozens of GPUs, we are already generating good videos.
Also, I hope you could be less mean to us next time, as I believe fastH3 generates sufficient decent videos that many users can already use at much lower cost, definitely not a "random pixel generator"
Speed is easy when you ignore quality. I can generate random pixels under 1 nanosecond and claim to be ♾x faster than H3 and call it H3-ultra-lightning-fast. But this is not how people should innovate on speed in open.
Image and video models are notoriously hard to optimize and we @ fal spend 2-3x more time on evals and making sure optimizations work than actual kernel engineering work.
But when you see people claiming they are X times faster than original while messing the quality to a degree that the outputs has nothing to do with original model, it is annoying.
Numbers below are for this case, but this is not the first/only one doing it. I'd recommend the authors below, and everyone on the field to run their own evals.
Without preference and quality evals in the loop, ai-slop-performance agents can just lead you into a way you generate random pixels and it is just dishonest to call it anything but a random-pixel-generator.
Ayush Chaurasia retweeted
Every generative model is a kind of dream, rendered from what it has seen.
Introducing Reverie, LanceDB’s summit on the data systems those dreams are made on.
Nov 5 in SF, with speakers from @nvidia, @runwayml, @LumaLabsAI, @AppliedInt, Cruise by @GM, @Adobe, @ExaAILabs + more.
na2.hubs.ly/H076c7p0
