
Grad Student @Mila_Quebec & @mcgill | Prev: Research Associate @mbzuai; AI Resident @ai4bharat; Intern @ucsd | Cooking stuff around LLMs ...
Mumbai, India
Joined September 2016
- Tweets682
- Following2.2K
- Followers430
- Likes1.8K
Jay Gala retweeted
Big MoE models use expert parallelism, so each layer's experts sit on different GPUs. At every MoE layer, each token's hidden state goes to the GPUs that hold its experts, and the outputs come back. Across machines, this traffic goes over the network, which is much slower than links inside a machine.
In models like Qwen3 and DeepSeek-V3, a token's inputs to neighboring MoE layers have an average cosine similarity above 90%. This is because each transformer block adds its output to the residual stream instead of replacing it. So they send the full hidden state only every 4th MoE layer, as a reference that every GPU keeps. In between, they send only the difference from that reference, compressed to a quarter of the size. Expert outputs are compressed before they're sent back, too.
On Qwen3-30B-A3B across 4 machines, this cuts cross-machine traffic by 58%. The average score on 18 benchmarks drops from 66.9 to 63.1. Directly compressing the full hidden state loses more accuracy than sending the difference, because the difference is much smaller. At about half the traffic, direct compression scores 61.6, while their method scores 63.8.
Less traffic speeds up prefill the most, because communication takes about half of each prefill step in the original model. On 32 H100s with vLLM, the compressed model has 22.6% higher throughput and 23% lower time to first token than the original.
Jay Gala retweeted
Can we predict the policy RL would learn from a new reward, without actually training on that reward?
In our new paper, we introduce PoEM, a method for predicting RL outcomes by composing models previously trained on other rewards. 🧵
Jay Gala retweeted
love my students' work on MoEs that communicate less over RDMA! great stuff, @fyliufengyuan and @jaygala24 !
Jay Gala retweeted
The era of one-size-fits-all inference engines is ending!
We had Claude write a separate megakernel for each of 8 open models, 8B - 594B: one GPU kernel that runs the entire decode loop.
Up to 1.77× faster than vLLM / SGLang, 2.05× with lossless model-specific tricks. 🚀
Jay Gala retweeted
Introducing Model Factory.
Train, deploy, and continuously improve your own models on one stack.
Run on-demand, on dedicated clusters, or bring your own.
Own your intelligence. Own your model factory.
Jay Gala retweeted
In the near future, LLM inference, particularly MoEs, is increasingly bounded by communication rather than compute, as experts sit on different nodes.
EchoMoE shows that MoEs don't have to communicate the entire representation but only the deltas. Sounds simple, but @fyliufengyuan and @jaygala24 worked hard at it to make this work.
A retrofitting technique for pretrained models, so no need to train from scratch like LatentMoEs.
Check out our approach to reducing MoE communication, similar to LatentMoE, without pretraining from scratch by exploiting cross-layer predictability in representations of existing MoEs.
Joint work with @fyliufengyuan and our advisors @sivareddyg @DBahdanau
More details👇
Jay Gala retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
Jay Gala retweeted
We will be teaching Agentic GPU Programming for MLSys as part of the CMU ML Systems course this upcoming spring at @SCSatCMU, touching on knowledge curation, compiler analysis, and feedback loops. We also curated an online book on the topic, check it out mlc.ai/agentic-gpu-programmi…
Jay Gala retweeted
RL with LLMs is very unstable when training and sampling policies differ. Standard fixes (matching numerics, importance sampling) work around the problem.
We find the root cause of this instability from first principles and propose a way to directly cancel it. Score Centering is competitive and compatible with existing approaches — while simple to implement!
🧵 [1/6]
Jay Gala retweeted
Introducing 𝙤𝙥𝙚𝙣-𝙖𝙜𝙚𝙣𝙩-𝙫𝙞𝙚𝙬, an open dashboard for managing all your agents across any harness.
Based on Claude Agent View, it lets you manage any coding agent (Claude, Codex, Pi, OpenCode, Muse Code, Qwen, etc.) in one agent view dashboard:
github.com/xhluca/open-agent…
Jay Gala retweeted
Introducing 𝙨𝙚𝙨𝙨𝙞𝙤𝙣-𝙢𝙞𝙜𝙧𝙖𝙩𝙚, a tool for migrating coding agent sessions across Claude Code, Codex, Pi, OpenCode, and more.
If you started in Claude but ran of usage, you can now convert your session to another harness in one command:
github.com/xhluca/session-mi…
Jay Gala retweeted
I recently gave a talk introducing young economists to post-training.
The slides are now up! kawine.github.io/assets/aies…
Not only is post-training your own model more doable than ever, but economists and other social scientists can play a big part in the next era of post-training.
The post-training suite looks roughly like:
1) SFT
2) Offline Methods (DPO/KTO/SimPO/etc)
3) Online Methods (PPO/GRPO/CISPO/etc)
4) RL Environments
5) Distillation (<--- we are here)
6) World Adaptation (<--- the future)
As agents are deployed in the real world, the real world will adapt to them. This adaptation is largely overlooked by the current post-training paradigm because working with verifiable rewards from static environments is much simpler and scalable. As we saw with the OpenAI-Huggingface incident however, we shouldn't assume that the real world will remain fixed---quite the opposite, in fact. So:
1. How can we anticipate real-world adaptation during post-training?
2. What do equilibria look like in these scenarios, if there are any at all?
3. What tools can be created to scale up human oversight?
Many such questions abound, and if you're an economist, these are high-impact problems worth studying.
Thoughts About Scaling Law
Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions.
The field learned this the hard way. Kaplan et al. (2020) fit an exponent that told everyone to grow parameters faster than data — roughly 2.7:1 — and the industry complied: GPT-3, Gopher, MT-NLG. Hoffmann et al. (2022) redid the experiment across four hundred models and found the compute-optimal split is closer to 20 tokens per parameter, and that with sufficient compute the two should grow at the same rate rather than drifting apart. The error in the earlier fit compounded with every order of magnitude of compute, which is why the largest models of that generation were the most misallocated. The trillion-parameter round was, in retrospect, a detour the whole field took together and then reversed.
Chinchilla wasn't the end either. It optimized training compute for models that would be trained once and evaluated. Today a model is called billions of times a day and inference dominates lifetime cost. Put inference into the objective and the optimum moves toward smaller models trained far longer — deliberate over-training, which is what Llama-2-7B and Gemma-2-9B were doing at roughly 290 and 889 tokens per parameter.
Sparsity moved the target again. In a MoE model two quantities have to be kept apart: total parameters govern roughly how much the model can hold — knowledge, facts, the long tail — while activated parameters and effective depth govern roughly how far it can think, how many steps of a causal chain it can carry before it comes apart. A dense 20:1 ratio does not transfer. And the ratio isn't a single number at all: Roberts et al. (2025) find the optimal tokens-per-parameter is task-dependent, with memorization favoring more parameters and reasoning favoring more data. Follow-up work on MoE observes that at fixed TPP, pushing total parameters higher actually degrades reasoning, while activating more experts reliably helps it.
This matters for what we are building toward. Finding a vulnerability is not a retrieval problem. It doesn't come from having memorized more CVEs; it comes from carrying a twenty-step chain of inference to the end without losing the thread. That capability does not live in total parameter count.
Which brings us to this release. Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training. GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal. Well, scaling has more than one dial. We turned the post-training one this time because it had the most slack left in it — not because the others are finished. Base model size, pretraining data, compute spent per forward pass: all of them are still on the table, and we will come back to each. What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling. Next time, maybe mid-training, pre-training, and even more.
Jay Gala retweeted
How can an LLM switch between low-, medium-, and high-effort reasoning? And how does an LLM learn to reason more or less?
I put together a “little” article explaining how these effort levels are implemented at inference time and during training.
Jay Gala retweeted
Heard that some frontier models are basically a 48-layer transformer looped twice (48L x 2).
Now we are introducing DeepLoop: Depth Scaling for Looped Transformers (arxiv.org/abs/2607.13491),
making the loop transformer stable and scalable!
Jay Gala retweeted
Introducing agent-talk, a new way for coding agents to work together.
Through agent-talk, agents can message each other directly, enabling coordination on projects spanning multiple developers & agents. You no longer need to be the messenger between your agents.
Jay Gala retweeted
I had the honor of giving a keynote at the International Conference on Machine Learning in Seoul last week titled “What will be left for us to work on?” I addressed the widespread anxiety about how we should adapt as AI capabilities increase. I was thrilled by the talk’s reception, so I have made my slides available, annotated with a lightly edited transcript: cs.princeton.edu/~arvindn/ta…
I made three arguments. First, the "AI as Normal Technology" framework is a correct and useful as a way to think about AI’s impacts, unless and until there is some future discontinuity such as through recursive self-improvement. Second, even though we should take recursive self-improvement seriously, there is no milestone that companies might achieve in the lab that will suddenly put us all out of work. Third and finally, jobs of the future will be radically different, and a lot of adaptation will be needed. I shared my thinking about what this might look like and ended with a vision of human/AI “co-superintelligence”.
Jay Gala retweeted
new post on harness engineering for AI self-improvement: lilianweng.github.io/posts/2…
It is hard to forecast how much the future of RSI will rely on harnesses. Likely harness engineering will evolve in the direction of self-improvement and enable auto-research, and, in turn, smarter models keeps harnesses simple.
Even when many harness improvement get eventually internalized into core model, the need to specify goals and context will not disappear.


























