Chuck Tang

Samir Menon @blintzbase and I are thrilled to announce Sail @sailresearchco ! We build infrastructure for long-horizon agents: inference served at unbeatable prices-per-token for open models, plus sandboxes designed to run for days, weeks, or longer. We've raised $80M, w/ our seed led by @Sequoia and series A led by @KleinerPerkins. We're using this capital to build the most efficient infrastructure for long-horizon agents. What makes agents so different? Unlike a human waiting at a keyboard (top priority: speed), agents need scale, reliability, and sustainable cost. Sail finds this efficiency everywhere in the stack: we carefully choose our chips, write custom inference engines, and run a global controller that fully utilizes every computer in our fleet. Tight integration from silicon to API lets Sail open up the cost / latency frontier to our customers - the most patient agents can now access 10x more intelligence per dollar. We're excited to be working with great companies like @parallelweb, @detaildotdev,@Jackandjillai, and @quadrillion_ai to deploy long-horizon agents with trillions of tokens. Our team is thoughtful in our engineering craft and relentlessly ambitious in our pursuit of peak performance. We previously trained at companies like NVIDIA, OpenAI, Google, and so many trading firms. Now we're ready to do the work that will define our careers, in the most compute intensive market of all time. Welcome to the era of abundant intelligence. We can't wait to build with you!

Today, we’re excited to announce our $50M Series B, led by @GreenfieldVC, with participation from @lightspeedvp and @notablecap. 🚀 At Patronus AI, we develop simulations and evals to train and improve AI. The first phase of AI was built on static benchmarks, but that era is over. As agents are used to solve longer and longer tasks, they need to practice in dynamic, living worlds to get better. Simulations are the critical infrastructure powering this next phase. As a company, we’re behind the most influential research and products in AI evaluation, like FinanceBench, Lynx, and Percival. And things have moved at the speed of light since.⚡ We partner with the world's leading frontier AI labs and enterprises, and our revenue has grown more than 15x over the past year. Additionally, today, we’re introducing a preview of the first Digital World Model for AI agent training and simulation: Patronus-DWM. Digital World Models are language diffusion world models that predict realistic environment behaviors and steer agent actions across digital workflows. Just as physical world models predict how objects move through space, we’re developing the equivalent for the digital world: predicting how agents act in digital workflows, then using that to scale the creation of high-quality training data for LLMs. Digital World Models help us push the frontier of ultra long horizon workflows, and unlock a new class of self-improving RL environments. This is our scalable approach to simulating all of the world’s intelligence. The round was also joined by @datadoghq, @SamsungVentures, @gokulr, @factorialcap, and a large cohort of amazing AI leaders across @AnthropicAI, @OpenAI, @GoogleDeepMind, @nvidia, @Recursive_SI, and more.✨ It has been the ride of a lifetime. But we’re just getting started. The best is yet to come. "Do not go gentle into that good night, Rage, rage against the dying of the light" - Dylan Thomas (1954)

Now that Tacit Labs and Engram are out of stealth, a casual plug for interning on the research team at @databricks: * @ninklefitz → Founder @tacitlabsco * @lindensli → Founder @appliedcompute * @dan_biderman → Founder @EngramLab * @ChrisRinard → Founder @Standard_Kernel

@trajectorylabs @nvidia @harvey cool work @o__jerry__o

1/ We post-trained @nvidia Nemotron 3 Ultra on @harvey Legal Agent Bench in under 24 hours. The result: an open model reaching the same band as leading closed models on legal work, at a fraction of the cost. The correlating story: when a new open model ships, Trajectory can turn it into a specialized agent almost immediately.

@jjanezhang maybe barry’s fidi 7:10am?

Claude is always “excising” things from my writing. Where in the earth of training data did it uncover this word?

5 Days of Trajectory 🏹Day 5: Scaling SDPO to Agentic Tasks Continual learning means you must train on data from production. But production gives you one example per task. A user makes a request once. You get one trajectory, not a batch. However, current RL algorithms don't work that way, They need groups of tasks. By definition, that means you need some artificial environment to perform those rollouts in. But what if you don't? SDPO is a promising route. It learns from a single trajectory, with no group required and failures still producing signal. The shape of the method matches the shape of production data. But one fundamental problem remained. Every published SDPO work assumed fresh, on-policy rollouts. Agentic work cannot give you that. Trajectories run for an hour or more and arrive stale. On true agentic tasks, naive SDPO collapses. We fixed it. We're the first to make SDPO work on agentic tasks. On Mercor's APEX-Agents, with hour-long trajectories and near-zero base pass rates: 25% average reward, 5x over zero-shot. More importantly, it trains stably and the curve is still climbing. Read more below.

@trajectorylabs @neilkale questin for you: irl are you the teacher or student model?

@GoogleStartups @trajectorylabs Shoutout to the GCP team, Firestore and Kubernetes have been such incredible foundations to build on top off for our continual learning stack.

Big win for @Trajectorylabs! Continual learning just got a massive upgrade with C-LoRA—delivering 2.81x throughput for frontier models without sacrificing performance. Check out the open-source stack on SkyRL below!👇

@trajectorylabs code sample is in the getting started section btw :)

🏹5 Days of Trajectory. Day 3 - An Open Source Training Stack for Continual Learning Building the platform for continual learning requires both partnering with pioneering AI companies, as we showed on Day 2 with Harvey, and working toward frontier research, which we are highlighting today. Continual learning means models that improve hourly from real production use. But with the size of frontier models, this becomes quite difficult. A Qwen-397b would need to spin up and tear down repeatedly across six GPU nodes, and that's valuable time gone. Our contribution is Continual LoRA (C-LoRA): many lightweight adapters running at once on one shared base model. Our insight centers on where the parallelism lives: instead of splitting one giant job across nodes, we load-balance many small jobs over a single base. The result: 2.81x experiment throughput over single-tenant training, with no regression on rewards. We built this together, with @anyscalecompute, @NovaSkyAI, and generous support from @GoogleCloud and @GoogleStartups. We've open-sourced on SkyRL as one of the first multi-LoRA, RL training platforms, so that every team can get to continual learning faster. We’re very excited to see what you build, please reach out!

@sumanthrh @pcmoritz @tyler_griggs_ Kudos to the skyrl folks for seeing the promise in multi-LoRA so early 👀

This is a special release. @j316chuck and team have been working closely with us in designing multi-LoRA training with SkyRL. @pcmoritz @tyler_griggs_ realized the value of the Tinker API a while back and led the Tinker server effort in SkyRL few months back. However we only had multi-tenancy with the Jax backend. We now have an initial implementation for multi-tenancy with Megatron+vLLM thanks to @j316chuck @erictang000. Props to @j316chuck for all the hard work on the benchmarks!

@MichaelElabd @erictang000 @charlie_ruan @sumanthhegde @pcmoritz moar loras and moar gpus

@j316chuck @erictang000 @charlie_ruan @sumanthhegde @pcmoritz Fire analysis!!! Whats next for you to Chuck??

Had such a blast working with @erictang000 , @charlie_ruan, @sumanthhegde, and @pcmoritz on enabling multi-LoRA RL training in SkyRL! We observed ~3x higher experiments throughput in comparison to running experiments in the traditional single-tenant fashion. One of my favorite parts of this collaboration is that all this code is open source so you can play with it yourself :) Here's the technical deep dive 🧵

@rronak_ @erictang000 @charlie_ruan @sumanthhegde @pcmoritz i heard trajectory uses forkable to feed their chefs

@j316chuck @erictang000 @charlie_ruan @sumanthhegde @pcmoritz The implications for the open source community are massive, congrats on the release Chuck, and amazing analysis!! What are they feeding you at Trajectory?👨‍🍳

@yapdianang @erictang000 @charlie_ruan @sumanthhegde @pcmoritz answer for you 8 loras

@j316chuck @erictang000 @charlie_ruan @sumanthhegde @pcmoritz Question for you: what's a better than octopus recipe

@erictang000 @charlie_ruan @sumanthhegde @pcmoritz 6/ To check out multi-LoRA training yourself, please see the getting-started section here: trajectory.ai/field-notes/mu…! Also, If problems like this interest you feel free to reach out to hello@trajectory.ai :)

@erictang000 @charlie_ruan @sumanthhegde @pcmoritz 5/ Does quality still hold in multiLoRA-training? Yes Every run hit reward accuracy >90% by step 9 at every concurrency level, tracking the serial baseline within ±1σ across the final 4 steps. No regression.

@erictang000 @charlie_ruan @sumanthhegde @pcmoritz 4/ As usual, there is no free lunch, the first multi-tenant experiment lands 1.97× slower than serial, trading off experiment throughput for latency. Our mean experiment time to completion is still much faster by 1.63×!

@erictang000 @charlie_ruan @sumanthhegde @pcmoritz 3/ In practice, we tested on 8x GSM8K synchronous RL jobs recast as an agentic tool-learning task with Qwen3-4B as the base model. We saw 8 concurrent LoRAs outperform the serial baseline by 2.81x!

@erictang000 @charlie_ruan @sumanthhegde @pcmoritz 2/ At Trajectory, we built a training platform served upon a warm engine with multi-LoRA training. The core idea: parallelize across jobs rather than within! We call this continual LoRA or (c-LoRA) for short.

@erictang000 @charlie_ruan @sumanthhegde @pcmoritz 1/ There are 4 reasons RL training doesn't scale today - Cold starts: 30+ min per large job - Large models: Qwen3.5-397B takes 6 nodes for RL - Single-tenant: Only 1 experiment per set of GPUs - Low utilization: trainer and generator hard to optimize jointly