Rhythm Garg
137 posts
@rhythmrg
Co-Founder, CTO @appliedcompute 🚂 prev: research @OpenAI @Stanford
Controlling staleness in an async RL stack has not been well understood, so we derived a closed-form formula that predicts staleness in advance. Our predictions match measured staleness from production RL runs within a fraction of a step. Building these simulations has led to key insights on how we maintain high GPU utilization during our RL runs without sacrificing ML performance.
We used RL to train models that create curated context from long documents for downstream use by agents. The models sometimes learn to invent their own abbreviations and shorthand. Optimizing with RL for downstream use produces very different artifacts from ordinary summaries: shorter, denser, creatively concise. We call these neural cheat-sheets.
Every enterprise will have its own model-harness-sandbox-eval flywheel with token value per watt optimization. This is the future. Simple reason: tacit knowledge about the domain and customers and their workflows that the company uniquely understands and has built trust around.
"There should be as many models in the world as firms in the world." Satya and I dig into when to own vs. rent your intelligence, why every company should be building and climbing its own private evals, and what makes for a stable frontier.
We partnered with @harvey to post-train the state-of-the-art legal agent on their LAB benchmark. It surpasses Opus 4.8 Max and GPT-5.5 xhigh.
Model strategy for @harvey: We are working on the first model in our legal foundation model series, inspired by @cursor_ai's Composer. Two goals: 1. Allow us to serve frontier intelligence across our product surface areas at an affordable price and a strong security posture. 2. Create the foundations for law firms to build their own specialized models and own their own intelligence. The model series will focus on complex client matters that span months and take dozens of associates. The agentic system will learn to control legal tech tools, sub agents and ask for help from frontier models or human partners, much like a senior associate. We’ve open sourced benchmarks for evaluating our initial post training work that represents work done by associates and in-house lawyers. We are scaling these significantly using synthetic and human pipelines as well as building private evals for firms. Open sourcing this data has allowed us to quickly validate the feasibility of post training open weight models for legal work. With our research partners we’ve already shown promising results post training open source models to approach frontier performance: 1. @baseten - novel compaction strategies for analyzing large data rooms. 2. @FireworksAI_HQ - matching frontier performance by using frontier as an advisor. 3. @appliedcompute - improving performance and reducing cost of large scale review tables. 4. @trajectorylabs & @nvidia - sovereign continual learning over client matters. We plan to continue to invest heavily in working with research partners and open sourcing our data, models and research as much as possible. We believe open research in legal will be important to building trust in the frontier ecosystem. We are also scaling our research team. Harvey Labs is our internal research group, responsible for pushing the frontier of legal intelligence and working closely with labs, research partners, and academia to bring the frontier of agent research into Harvey. Labs is run by @nikogrupen and @ItsJulioPereyra - Niko worked on multi-agent RL at Google Brain and Julio clerked and worked in BigLaw. We believe this pairing is crucial for building frontier legal AI systems. Together they have already made significant progress in scaling our data and training efforts. The long term goal of Harvey Labs is to contribute to the research and infrastructure required for the legal industry to create a frontier ecosystem. We believe that the best version of legal super intelligence is one where each law firm, enterprise and government owns their own specialized version. We are hiring for Harvey Labs across the post training, agent and data stack and open to acquiring talented teams / neolabs in this space. If interested please DM me.
Preserving entropy is critical for continued training; in modern post-training recipes, entropy is often a fixed resource that gets exhausted over the course of a training run, making it difficult for the model to improve and learn on new tasks. Adaptive entropy control methods expose entropy as a controllable hyperparameter instead of a side effect. We find that in a continued training setting, GRPO leads to entropy collapse which stalls training performance on subsequent training phases. REPO-R (Petrenko et al., 2026) can hold entropy near a specific target, preventing this collapse, and pushing performance higher in follow-on tasks when the target entropy value is properly tuned.
