Chuck Tang (@j316chuck) profile picture

scaling compute @trajectorylabs

San Francisco
Joined November 2016
Photo by Chuck Tang (@j316chuck)Photo by Chuck Tang (@j316chuck)Photo by Chuck Tang (@j316chuck)Photo by Chuck Tang (@j316chuck)Photo by Chuck Tang (@j316chuck)Photo by Chuck Tang (@j316chuck)
I used to have 50 PRs a week, now I have 1,900. Three weeks ago, I built an AI bot called SuperJenny that works with ~20 AI teammates. They ship Monster Walk every week: builds, QA, store reviews, release notes, analytics. I 6x'd my mobile game studio's revenue by automating myself. Today I'm sharing my exact setup, free: SuperJennyStack 👇 At all times, I have my team working on: 🧭 SuperJenny: delegates and compiles asks under her - also trained with my personality and quirks 📋 SuperPM: ranked daily fix list from metrics + reviews + trends 🚀 SuperRelease: weekly build, Friday ready to submit new version 🧪 SuperQA: real-device tests with video proof 💬 SuperSupport: drafts every store review reply for my OK and helps players out when I sleep I check in 4x a day from my phone and chat with SuperJenny about what I like and what I want to see next. The rest is handled in the cloud. Super honored to say it's in the Template Library today if you try @cognition 's devin bot. I'd love to see what you create with SuperJennyStack!! If you share something you've done with SuperJennyStack, I'll RT the ones I see and like :D
26
19
6
180
49,599
Chuck Tang retweeted
As one of our earliest customers, it's been super exciting to support Trajectory with Tinker. In some sense they've been an ideal Tinker customer - they were influenced by Connectionism's original Lora blogpost, they've tried all sorts of wacky posttraining stuff that have informed our roadmap (as well as ... boring stuff that have informed our roadmap 😅), and they've converted it into real value for their users!
Imagine if everyone had the tools to shape intelligence itself. That day is today. Today we launch the Trajectory platform - the easiest and most intuitive tool to post-train models of your own. The foundation for continual learning. With just a few agent prompts, you get access to a full creative suite to train, eval, and deploy trillion parameter models. On top of that, we believe in a future where you don’t just shape your model once, it also keeps learning on its own from the world around it. We’re releasing two research previews. 1. Continual grading scales your taste, so your model learns what good looks like from far more than you could ever grade by hand. 2. GS1, a new algorithm that lets models learn directly from what they do in production. We’ve now raised 60M, with a new round led by @Sequoia at a 300M valuation to democratize the most powerful technology of today - training models - so that the lawyers, doctors, and engineers of the world can own their own intelligence. You know what good looks like, now go build it. Own your own value function with Trajectory.
9
7
160
16,314
trajectory.ai/waitlist (◕‿◕✿)
Imagine if everyone had the tools to shape intelligence itself. That day is today. Today we launch the Trajectory platform - the easiest and most intuitive tool to post-train models of your own. The foundation for continual learning. With just a few agent prompts, you get access to a full creative suite to train, eval, and deploy trillion parameter models. On top of that, we believe in a future where you don’t just shape your model once, it also keeps learning on its own from the world around it. We’re releasing two research previews. 1. Continual grading scales your taste, so your model learns what good looks like from far more than you could ever grade by hand. 2. GS1, a new algorithm that lets models learn directly from what they do in production. We’ve now raised 60M, with a new round led by @Sequoia at a 300M valuation to democratize the most powerful technology of today - training models - so that the lawyers, doctors, and engineers of the world can own their own intelligence. You know what good looks like, now go build it. Own your own value function with Trajectory.
2
2
13
885
Chuck Tang retweeted
Super excited to share some of the work I’ve been doing on supporting 1M token context training on SkyRL for GLM 5.3 Flash, 64x higher than the out of the box Megatron baseline! 🚀 This work was done in collaboration with the amazing folks at @LilaSciences! 🤩 Getting there took key algorithmic and systems optimizations in four main areas: 1) DeepSeek Sparse Attention: run DSA on a sparse kernel; chunk and shard its quadratic indexer. 2) Activation recompute and offloading: recompute activations in backward and offload saved inputs to CPU. 3) Context parallelism: split each sequence across GPUs, with new CP support for KDA and a fix for DSA. 4) Avoiding unnecessary precision upcasts: keep FP32 copies out of memory and off the network. For more details on our implementations and our results, check out our full technical blog here: blog.skyrl.ai/blog/1m-contex…! This work benefited from the contributions of many people along the way. Big thanks to @pcmoritz, @erictang000, @sumanthrh, @CyrusHakha, @richliaw from @anyscalecompute and Keith Truongcao, Tushar Khot, Vivek Subramanian, Semih Kara from Lila Sciences for all the help and guidance!
5
26
6
174
17,278
Chuck Tang retweeted
Awesome seeing @JensenHuang talking about @harvey and @trajectorylabs work with Nemotron at GTC
7
33
4
127
11,875
Chuck Tang retweeted
7pm tonight @ 680 Folsom St, SF. @trajectorylabs x VESSL AI, for #sftechweek RSVP: partiful.com/e/UdGWc3yV9ffPB…
3
3
455
Chuck Tang retweeted
Excited to announce the release of SkyRL v0.4.0! Some highlights: - Support for GLM 5.3 Flash, GLM 5.3, and Kimi K2.7 - Top-P sampler replay and optimized R3 data transfer - Full FP8/MXFP8 RL - Sharded P2P weight transfer Blog: blog.skyrl.ai/blog/skyrl-v0-… Details in 🧵
2
32
3
191
32,501
The M in GLM stands for mountain ⛰️ (joke) Here is a blog on how we climbed it 🧗 (not joke)
Replying to @trajectorylabs
The first obstacle: silent numerical mismatches. In RL, a sampler writes answers and a trainer learns from them. Both run the same model, but with different kernels and math, they can quietly drift apart. Nothing crashes, all you just get a flat reward curve hours later. Very helpful.
2
17
1,819
Chuck Tang retweeted
The first obstacle: silent numerical mismatches. In RL, a sampler writes answers and a trainer learns from them. Both run the same model, but with different kernels and math, they can quietly drift apart. Nothing crashes, all you just get a flat reward curve hours later. Very helpful.
6
1
2
26
4,704
was very cool seeing chef @yapdianang cook up this training recipe producing: - less model yaps + more model intelligence or in other words more intelligently dense, open source models :)
There’s a good chance your open source model is costing more than the frontier. Cheap tokens ≠ cheap tasks. Here, we introduce Intelligence Density, and Density Aware Training, our post-training technique to achieve less wasted compute, better learning, all with no knobs to tune. Enabled by default in every Trajectory model.
3
331
Chuck Tang retweeted
All seven Databricks co-founders began our journey together at UC Berkeley. We started the company in a room in Soda Hall, moved to a small office on Addison Street, and spent our first few years in Berkeley before eventually heading to SF. Two of our co-founders are still on the faculty. Today we announced Databricks Field at California Memorial Stadium, our first collegiate athletics sponsorship. Berkeley shaped everything about how this company thinks, and this is our way of investing in the next generation of students and builders who'll do their best work on that campus. Go Bears! databricks.com/company/newsr…
123
220
53
2,351
234,243
goal this weekend: - eat more pizza - train more pissa nice work @neilkale :)
At Trajectory, we're constantly implementing and building upon the latest research ideas on the path to continual learning. We wish we had the time to share all of them, but here's a quick glimpse on our explorations with PiSSA, and choosing the right trainable geometries for agentic RL.
3
22
1,834