Landan Seguin

@maxisawesome538 👋❤️👋

I'm starting a surfing group chat. Any surfers on my TL? reply to be included

today we're announcing our @DbrxMosaicAI x @Shutterstock partnership, and a new text-to-image diffusion model: ✨ImageAI!!✨ this model is geared towards enterprise use cases and is trained exclusively on shutterstock's trusted data catalog! databricks.com/company/newsro…

@ilyas121_real @summerlinARK @Replit @amasad @MosaicML Our SD training estimates are in the blog post below: $160k in 2 weeks. Next, we will train the thing :) watch out for it! twitter.com/MosaicML/statu…

@summerlinARK @Replit @amasad @MosaicML When you say train SD for 160k is that from scratch? Has it really come down from 600k in just 5ish months?

.@Replit & @amasad are writing the playbook on building an AI-first company. With @MosaicML, the cost to train your own model is de minimis. I.e., you can train Stable Diffusion on MosaicML for $160k. Other reasons to train your own model: inference cost/speed, data privacy, etc.

Just in time for Thanksgiving - we're dropping a new batch of recipes for training image segmentation models. Reduce time-to-train by up to 5.4x, improve quality by up to +4.6 mIoU, and impress everyone at your #efficientML potluck! mosaicml.com/blog/mosaic-im…

New blog post: mosaicml.com/blog/behind-th… We're setting an up-to-date baseline for semantic segmentation model training: 45.56 mIoU on the ADE20k benchmark in 3.5 hours using 8x NVIDIA A100 GPUs. Next step: develop and release #EfficientML recipes to speed it up!

Having trouble keeping up with arXiv? 🎉 Announcing "Davis Summarizes Papers" 🎉 tl;dr: People kept telling me I should make the ~15 paper summaries I do each week into a newsletter, so I did: dblalock.substack.com Free forever, and you can also read all past posts as a blog

We've shared great research before, but reproducing methods from papers is hard. Announcing Composer, our library of ML speedups: github.com/mosaicml/compo…. Train CV models ~4x faster and NLP models ~2x faster at the same accuracy -- with minimal tuning. (1/5)

TLDR: Announcing 🌟COMPOSER🌟, a PyTorch trainer for efficient training *algorithmically*. Train 2x-4x faster on standard ML tasks, a taste of what's coming from @MosaicML. Star it, 𝚙𝚒𝚙 𝚒𝚗𝚜𝚝𝚊𝚕𝚕 𝚖𝚘𝚜𝚊𝚒𝚌𝚖𝚕, contribute, be efficient! github.com/mosaicml/compo… Thread:

@borisdayma Nice! At @MosaicML, we’ve seen significant differences in perplexity when training a GPT-2 model on OpenWebText with vs without dropout! Look forward to seeing your results 😁

Do you still use regularization even when your dataset is huge? I was still using dropout while my dataset won't get processed more than 2-3 epochs (I also have a large batch size). Let's see what happens🤞

@johncarlosbaez I found this talk (and many others) by Alan Watts enlightening youtu.be/mMRrCYPxD0I (start at 1:26, ignore the cheesy music and video) "So after you are dead, the only thing that can happen is the same experience, or the same sort of experience as before you were born"

Left: MIT computer scientist Katie Bouman w/stacks of hard drives of black hole image data. Right: MIT computer scientist Margaret Hamilton w/the code she wrote that helped put a man on the moon. (image credit @floragraham) #EHTblackhole #BlackHoleDay #BlackHole

Kinda crazy how much more advice there is on persuading people than on being persuaded. Not just when to buy an argument, but also how to get it to percolate through your models of the world & actually behave as though you'd bought it.

@michael_nielsen @TheAtlantic It's frustrating how thinking feels like exploring a large idea cave serially with a tiny flashlight. It's a bit shocking how underdeveloped our tooling is in surpassing limitations of thinking / short term memory. Pen & paper was a good first step, haven't taken too many since.