Applied Compute

Read the full research report on our blog: Predicting and Controlling Staleness in Fully Asynchronous RL Training appliedcompute.com/research/stale…

Our key finding is that staleness is maximized when the RL pipeline is most efficient. Lower staleness requires slower steps and a different resource distribution between rollout generation and training. This distribution is just one input into a broader trade-off between staleness and performance. Our formula lets you sweep out the Pareto frontier and choose where you land.

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.

Join Applied Compute and @modal for an evening in Seoul! We’re hosting an ICML rooftop happy hour for top researchers, PhDs, AI lab scientists, and the people building the future of ML infrastructure and custom models. Come unwind and enjoy one of Korea’s hidden gems with us - RSVP in comments. #ICML2026

Read the full research report on our blog: appliedcompute.com/research/neura…

At deployment time, the trained model takes in a new knowledge artifact: an enterprise SOP, a meeting transcript, a coding agent trace, or even a novella, and outputs a dense set of notes that captures fine-grained detail and amortizes reasoning. This approach is used to build memory for agents, instantiated at Applied Compute as a ContextBase.

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.

5 takeaways from Satya: Why every company needs its own model: “There should be as many models in the world as firms in the world. Because after all, what is a firm? A firm is a learning system that today is mostly about human capital with digital tools. Every day compounding the tacit knowledge that the firm has and how it produces a good or a service that's valuable in the economy." Avoiding vendor lock-in: “We want to build every layer of the stack as something that is ecosystem extensible and model diverse. At the end of the day, we want to make sure we are making each layer competitive and at the same time making sure that the customers or third party developers have choice." Why firms should build a learning loop:  "You can always buy a tool, you can even outsource a task or even a job, but you can't outsource your learning. If you outsource your learning, then why exist?"  On Microsoft's AI strategy:  With Microsoft 365, "all the communications, data and information, their projects, the timelines, the calendars, the documents, all of that -- it's the unstructured database. We describe it as 'Work IQ.' We want that to be the context layer along with any model." How companies are transforming in the AI era:  "This refounding as a concept is empowering. The opportunity for anybody, whether you're a CEO or part of the management team or even an IC in this company, is to be part of the learning culture of this place and be part of that refounding team... We have to now go from our old to the new, fast. We have to work like an AI native company because an AI native company is born today."

@ypatil125 Great conversation @ypatil125! Human capital and token capital compounding together is the entire game. This is the positive-sum future we need to build to benefit everyone.

"You can always buy a tool, you can even outsource a task or even a job. But you can't outsource your learning. If you outsource your learning, then why exist?" Thank you @Microsoft for your partnership.

@sigalitperelson so back

sf is so back 🐎 @appliedcompute

@_gracehua Thanks for coming!

new challenge for your next startup party: start conversations with strangers - but u have to use openers in alphabetical order “applied compute.. so what do they do?” “bet this company is going to the moon” “compute, but applied amirite” happy office warming @appliedcompute!

“A modelless company is sitting on shifting sand.” Our CEO @ypatil125 sat down with @mariogabriele to talk about why owning your model is the difference between building on bedrock vs on someone else’s roadmap. It’s the core of what we do at Applied Compute. We train better, faster, cheaper custom models on your data, serve them in production, and continuously improve them as your definition of “good” evolves. Your model becomes your moat instead of a dependency that can change underneath you. Full conversation below!

Harvey partnered with @appliedcompute to train a legal agent. We optimized each part of the agent stack: - eval loop - agent harness and compaction - post-trained GLM-5.1 using reward signal from our Legal Agent Benchmark (LAB) More in our agent-training deep dive:

Harvey partnered with @appliedcompute to train a legal agent. We optimized each part of the agent stack, including the eval loop, agent harness and compaction, and post-trained the underlying GLM-5.1 model using reward signal from Harvey's Legal Agent Benchmark (LAB). Check out more in the agent-training deep dive below. Kudos to @nikogrupen, @ItsJulioPereyra, @rhythmrg, @jacob_dphillips, and @raymondmfeng for leading this effort - more to come, with lots of opportunity to push the frontier with GLM-5.2.

It was great collaborating with @nikogrupen, @ItsJulioPereyra, and @gabepereyra on a custom post-trained model for LAB. The rigorous work Harvey is doing to map out and build representative evals that reflect how real legal work gets done will pay massive dividends over time and help them continue to build unique research and model IP.

More evidence that the frontier is attainable with (1) high quality environments and domain expertise, of which @harvey has in abundance; (2) post training infrastructure to execute big runs. This ran on our Blackwell cluster without any issues, thanks to infrastructure that allows us to elastically scale inference compute. Was great to collaborate with @gabepereyra @nikogrupen and team!