Member of Technical Staff at @modal
San Francisco, CA
Joined September 2025
- Tweets41
- Following112
- Followers138
- Likes750
Modal magic, now for training!
Performant reinforcement learning is a hard infrastructure and research problem. We’ve recently made on-demand multi-node clusters available and showed how to take advantageous of heterogeneous sampling compute. But there still exists the need for high-fidelity environments, clear and robust reward functions, and tuned training recipes that actually work.
Most teams are actually nearly there! With an agent harness deployed in production and evals/agent traces collected from that deployment, you just need an easy way to start training your data. So we built a training library that gives you the building blocks, tuned defaults, and observability you need for easy, production-grade LLM training.
It’s available now. All it takes is a prompt.
𓆝 𓆟 𓆞 𓆝 𓆟 Raymond Tana retweeted
Performant reinforcement learning is a hard infrastructure and research problem. We’ve recently made on-demand multi-node clusters available and showed how to take advantageous of heterogeneous sampling compute. But there still exists the need for high-fidelity environments, clear and robust reward functions, and tuned training recipes that actually work.
Most teams are actually nearly there! With an agent harness deployed in production and evals/agent traces collected from that deployment, you just need an easy way to start training your data. So we built a training library that gives you the building blocks, tuned defaults, and observability you need for easy, production-grade LLM training.
It’s available now. All it takes is a prompt.
𓆝 𓆟 𓆞 𓆝 𓆟 Raymond Tana retweeted
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.
𓆝 𓆟 𓆞 𓆝 𓆟 Raymond Tana retweeted
Today we're announcing that Modal has crossed $1B in annualized run rate, as well as raised a $750M Series D at a $15.75B valuation, led by @Accel.
Since our Series C, usage has tripled and we’ve grown more than 15x year-over-year.
We are deeply grateful to the developers and companies building on Modal. There's still so much more to build, and we can't think of a better time to be building software.
𓆝 𓆟 𓆞 𓆝 𓆟 Raymond Tana retweeted
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.
Finally digitized notes on *algorithmic information theory* (raymondtana.github.io/ait-no…) from my PhD dissertation... I think of them like the backdrop to what I really want to write about:
the connection between information theory <-> ML <-> computability theory ... more soon(ish!)
𓆝 𓆟 𓆞 𓆝 𓆟 Raymond Tana retweeted
Been incubating a new project at modal, called Latte. We just got it working end-to-end Friday.
Off the name alone, take a guess what it is.
365 days ago i moved to sf, and i thank myself every day for making the leap in this city of contradictions
Should be a good time! I'll talk about the time-travel filesystem @erikdunteman and I have been working on
Our second meetup in SF is next week and it's going to be sick:
@glcst from @tursodatabase
@haakamaujla from @agentmail
@pretermath from @modal
Join if you're around!
𓆝 𓆟 𓆞 𓆝 𓆟 Raymond Tana retweeted
Come join me today at the @databricks agentic AI night, where I'll be presenting my work on agent filesystems at Modal.
You'll see a live demo of our filesystem Overeasy (pray to demo gods 🤞) and our vision for durability on Modal backed by the almighty WAL-on-S3
durable (and rewindable) state is an obvious inevitability, but it's undergoing active development and requires feedback from the people who are going to most benefit from it
so many SWE projects involve implementing a protocol. but some protocols are better-specified than others. and better specs => easier verification => better for agents. i think i want a protocol-protocol for declaring what makes a good protocol
𓆝 𓆟 𓆞 𓆝 𓆟 Raymond Tana retweeted
Something new we've been cooking on:
🍳 Overeasy: A durable filesystem overlay, backed by immutable logs on S3
Overeasy makes your FS:
- forkable
- resumable across machines (demo 👇)
- revertible to any prior state