Twitxy Saved

Rhythm Garg

Verified account Applied Compute

@rhythmrg

Co-Founder, CTO @appliedcompute 🚂
prev: research @OpenAI @Stanford
Joined January 2021
562 Following 5.1K Followers 195 Posts

Latest posts 5

Rhythm Garg Verified account @rhythmrg
We're excited to announce that Delegate is joining Applied Compute!

Felix and Jason have already made a big impact at AC since they joined. I'm excited for everything we will build together
Yash Patil Verified account @ypatil125 5d
Applied Compute has acquired Delegate!

We were one of Delegate’s first customers, and their product is how we kept up with hundreds of accounts before we had a sales team. Felix Su and Jason Yang’s attention to detail, pace of shipping, and product sense are unparalleled, and I’ve wanted to work with them ever since we met.

Before Delegate, they helped build Scale AI’s enterprise business from the ground up, getting AI into production at some of the largest companies in the world.

Much of our team is former founders like Felix and Jason. They bring a rare combination of ambition, agency, and the ability to run through walls.

Welcome, Felix and Jason! I couldn’t be more excited to build with you. 🚀

appliedcompute.com/company/delega…
View quoted post on X
3 replies 1 reposts 60 likes 2.4K views View on X
Rhythm Garg Verified account @rhythmrg
Great work on low-precision training with quantization-aware distillation from @danyulol. We are constantly working to make our training stack the most efficient, economical way to train the best model on your task
Dan DiPietro Verified account @danyulol Nov 9
Low-precision is the future. Frontier open models like 4.1 Flash and K3 already serve inference with low precision. I think the same thing will happen soon for training. Training actor forwards, backwards, and even optimizer state will move to 8-bit and 4-bit regimes.
View quoted post on X
1 replies 0 reposts 26 likes 1.1K views View on X
Rhythm Garg Verified account @rhythmrg
We're really excited by the idea of deploying a model for a customer and then continuously improving it through RL on real production traffic.

We've been doing a lot of research on this front (and deploying it with customers!) and are excited to share some of our learnings. A few interesting tidbits:

- Simply retokenizing production inference traces before training can cause RL runs to completely collapse, even with just one optimizer step per deployment. Preserving the exact token IDs from inference fixes this.

- Production RL introduces significant train-inference mismatch from different engines, quantization, and stale data. We found that score centering achieves faster and higher reward improvements than traditional importance sampling.

- When some runs collapsed, rolling back checkpoints, resetting optimizer states, and reshuffling data didn't recover them. This suggests gradual loss of plasticity rather than a single bad update, making long-horizon stability especially important for production RL.
Applied Compute @appliedcompute Nov 6
Production traffic is full of real experience from economically valuable work, but very little of it currently makes it back into training.

We built a recipe for production RL that steadily improves a model straight from its deployment’s data stream, then redeploys it to keep learning.

We are sharing some of our initial learnings from a recipe capable of continual learning in production.
View quoted post on X
6 replies 3 reposts 66 likes 4.4K views View on X
Rhythm Garg Verified account @rhythmrg
One thing we've learned building our autonomous post-training agent Ari is that identifying what's wrong with an RL run is very different from figuring out what to do about it.

Agents can often spot symptoms like truncation or failing tool calls, but then jump to the most obvious fix which might not be correct.

In one experiment, Ari observed that many RL rollouts were getting truncated. The obvious fix was to increase the response length limit, which our baseline agent spent several experiments exploring.

But Ari identified the underlying bottleneck: the model was frequently making illegal chess moves. So Ari modified the harness to only expose legal moves to the model.

We built a dataset of research decision points from past training runs and used it to hill-climb Ari's research skills. This improved root-cause identification from 38% to 54% on held-out tasks, without needing to run full post-training experiments for every iteration.
Applied Compute @appliedcompute Nov 4
Post-training a model usually takes a team of researchers running experiments for weeks. With Ari Autoresearch, now live on AC2, an agent can run that whole loop itself. In one 16 hour episode with no researcher intervention, it took Qwen3.5-9B from 0.118 to 0.629 held-out reward on chess puzzles.
View quoted post on X
3 replies 0 reposts 33 likes 1.5K views View on X
Rhythm Garg Verified account @rhythmrg
We crossed some exciting milestones at Applied Compute. But the best part about working here is the people.

We sit at the rare intersection of research engineering and practical deployment.

The people at Applied Compute are some of the best engineers I’ve ever worked with but also extremely pragmatic and driven by delivering real impact with customers. Many of them are former founders.

We took this picture in August, and the team has already grown more than 50% since then! We plan to keep growing quickly alongside customer demand, but never at the expense of the culture and the kind of people that make this place special
Yash Patil Verified account @ypatil125 Nov 2
Applied Compute update:

- crossed ⁠$100M annualized revenue run rate up from zero 18 months ago
- raised $350M at $3.25B led by Elad Gil
- 40% of customers use AC for both custom model training + production inference
- inference serving 1T tokens a day after launching 12 weeks ago
- 2X enterprise customers since last raise
- 3X team since January, 85% are researchers and engineers
⁠- Working with customers like Microsoft, Shopify, Harvey, DoorDash, Cognition, Clay, Decagon and others

To move even faster we need people who are extremely technical AND extremely practical. If this is you, please reach out.
View quoted post on X