@modal / ex @citadel @snowflake @shopify @splunk / waterloo se 23 / spain traveler @BTRVALORANT
New York, NY
- Tweets 1,454
- Following 712
- Followers 675
- Likes 15,162
Pinned Tweet
upset because I really think we could have won the whole thing, but damn just being half way across the world seeing all my friends and family back home cheering and supporting us is a feeling i will never forget
3
2
96
damian retweeted
from the school that brought you 30% median inside out midterm
21
195
5,789
203,303
The secret to Modal's speed is our custom lazily loaded container filesystem, ImageFS. ImageFS is built on a massive tiered distributed filesystem, layering local NVMe disks, regional and p2p caches, and global blob stores.
When we launched VM Sandboxes this month, we wanted to make sure they could take advantage of ImageFS in the same way our gVisor containers do. To do that, we built out a new filesystem stack that tricks a VM's ext4 driver into thinking it's talking to a normal disk.
14
30
543
52,737
1. wake up
2. work on cool systems problems with a bunch of geniuses
3. ???
4. profit
modal.jobs
1
58
4,945
damian retweeted
Spent the past three weeks on this idea.
A sampler on a different precision or accelerator is just off-policy, like stale samples in fully async RL, and off-policy correction doesn't care why the sampler differs from the trainer. Meanwhile, inference keeps diversifying across precisions(MXFP8, MXFP4, NVFP4) and accelerators outside RL. So why not use all of that compute for RL rollout?
The sampler and the trainer are two systems built for different purposes. Hope this opens up new possibilities for RL algorithms and systems, and leads to cleaner system designs. Enjoy the blog, and stay tuned for the next one on our infra :)
Training large models with RL typically runs on a fixed homogeneous cluster of the latest accelerators for numerical stability and faster throughput. This throws away inference optimization work done for other vendors and generations, limits the capacity you can utilize, and is unnecessarily expensive.
So we trained a model on agentic coding tasks using fully asynchronous RL and mixed-precision, mixed-accelerator samplers. With off-policy correction, training stayed stable through convergence, on par with homogeneous samplers, at a lower cost.
24
32
385
32,326
in the past few years i’ve seen us invent forms of gambling i didn’t even know were possible
We created a card that sometimes doesn't charge you.
Buy Now, Pay Maybe.
Community note
The Tuyo card is a Visa debit card for spending USDC stablecoin. The "Buy Now, Pay Maybe" feature randomly covers some purchase costs, but probabilities are not disclosed. It is self-custodial, with no traditional banking protections like FDIC insurance.
apps.apple.com/us/app/tuyo-bu…
tuyo.co
1
8
2,687
damian retweeted
I made Wrapped but for your taxes. See what the government spent with your money
86
485
6,897
433,427
damian retweeted
i think my smart bed cover won’t get colder bc us-east-1 is down
24
150
5,865
181,732
