Applied Compute CEO on the Limits of RL, the New AI Hyperscaler & Why Post-Training Wins Inference

Summary · 5 min read instead of 57 min

Yash Patil is a former OpenAI researcher whose main work there was on Codex. He now runs Applied Compute, a 16-month-old company that post-trains and serves custom models for AI application companies, working with customers such as Harvey and a chip company.

Patil argues that owning your intelligence is about flexibility and control rather than fear of malicious labs, and that post-training lets companies push the cost-performance Pareto curve out, not just slide along it. He says the largest inference workloads are the most valuable to post-train and to serve, which is why he believes post-training wins inference and why Applied Compute is building toward a new AI hyperscaler on GPUs. He calls RL a hill climbing machine whose hardest part is defining the hill, so companies should guard their private evals. He also says RL generalizes far less than pre-training and that continual learning is blocked by the unsolved problem of data-efficient training from sparse rewards. Looking back on the past year, he says he underestimated how much cost, through Jevons paradox, would drive decisions over capabilities.

Minute by minute

1. 1:40

   Owning intelligence means flexibility and control. Patil says the case for owning models does not rest on labs being malicious; he credits their people, contracts and data protections. The point is control over where models run, how they are deployed and what they are optimized for, which requires weights plus training, inference and routing infrastructure. He notes models have been pulled from products more than once, especially in coding, and labs may stop serving companies they compete with.

2. 4:32

   The market is moving to many models. Patil reads OpenAI's partnership with Baseten, which lets customers spend OpenAI credits on open models, as a signal of a multi-model future and of OpenAI wanting to be a platform. He says the narrative shifted from the biggest model to a cost-performance Pareto curve. Open models let buyers slide along that curve, and Applied Compute's thesis is that training moves the curve out, giving better models at lower cost.

3. 7:18

   Differentiated data drives capability gains. Asked whether post-training can beat frontier models on capabilities, Patil says the more out of distribution a company's data is, the more likely post-training yields a capability lift. Most customers, though, post-train to optimize price performance, for example bringing a cheaper model to parity with the frontier so they can serve more users.

4. 8:58

   Judgment traces will stay proprietary. Patil expects training to move from offline RL on vendor datasets to online RL that improves models from production use. Data that stays out of distribution is produced inside a company, such as judgment traces shaped by its risk thresholds and operating model. He argues continual learning is blocked by data-efficient training from sparse rewards, an unsolved problem, and lists pre-training, supervised fine-tuning and RL with verifiable rewards as still not data efficient.

5. 14:29

   Update weights for judgment and changing tasks. Patil says the future involves a non-static set of weights with feedback from the real world, mainly for tasks that change over time or depend on judgment. One customer's assistant product uses direct user feedback with online training, improving satisfaction and completion rates and cutting tool call failures. He still advises optimizing context and harness first, then the policy, citing a chip company whose custom software general models used poorly.

6. 18:14

   Post-training costs can fall sharply. Patil says GPUs are more underutilized than people expect and credits cofounder Lyndon, the chief architect, with work on this. He argues most cost reduction will come from systems and from more data-efficient algorithms that train for less time, more than from synthetic data generation.

7. 19:44

   Companies differ enough to need custom models. Responding to Satya Nadella's view that there should be as many models as firms, Patil says most companies need not pre-train or mid-train, but closed feedback loops inside each are worth capturing. Fragmented work like accounting will consolidate, while banks make different trade-offs. He adds that amid a compute supply crunch, doing something 10x cheaper than a competitor is itself differentiation.

8. 24:16

   Rubrics make RL work beyond math. Patil says early RL focused on math and coding, but non-verifiable domains are surprisingly easy to hill climb by grading against expert answers with rubrics. Because RL is a "hill climbing machine" and defining the hill is hardest, he argues companies should build evals and keep them private, since public benchmarks get optimized against.

9. 27:18

   Applied Compute picks high-value workloads. Patil says custom training makes the business sticky, but most value capture is in inference, since most model spending goes to serving. Applied Compute looks for two kinds of customers: very large inference workloads, where small gains spread over billions or trillions of tokens, and very out-of-distribution data, as at pharma, cybersecurity or chip companies.

10. 28:59

    Post-training wins inference. Patil says the most "post-training actually wins inference" because the largest workloads are post-trained first, following a power law. Training choices shape serving, such as disaggregated prefill and decode setups for tool-heavy agents, and training kernel work carries over. He offers making models 10% more token efficient as an alternative to inference optimization, and calls serving the easier half, citing Dylan Patel on GPUs, vLLM and OpenRouter.

11. 35:02

    A new AI hyperscaler built on GPUs. Patil compares the opportunity to AWS, GCP and Azure building on CPUs. He describes an inverted pyramid: training at the bottom, where few can compete, then inference, routing, and context and harness on top, plus areas like security and agent sandboxing. Applied Compute, 16 months old, moved from training into inference and works only with mature AI products rather than zero-to-one agent builds.

12. 38:23

    RL needs task, environment and verifier. Patil says building training data mirrors building evals: the right task, environment and verifier. For a Harvey Review Table model, the team gathered expert rubrics and answers, added synthetic data, gave the model tools over a document base and graded with rubric-based RL. He says synthetic work like back-translating documents into question-answer pairs is possible, but expert input requires working directly with customer teams.

13. 41:40

    RL transfers weakly, cost matters more. Patil says RL generalizes to skills like error recovery and tool use, but training on math transfers little to legal work, far less than pre-training. His changed view is how real Jevons paradox is: OpenRouter data shows usage spikes after price cuts. He says cheap, fast models like Jev enable new uses, such as Applied Compute's massive trace analysis for online training.

14. 47:15

    Hire people who can explain their work. Patil, who worked mainly on Codex at OpenAI, supports AI coding but warns against delegating thinking, which produces "massive slop code bases" no one can explain. He recommends learning system design. Applied Compute's interviews let candidates use any tools, then probe trade-offs, including for finance roles. He says data vendors like Mercor are strong companies but must keep pivoting as research paradigms change.

Published Oct 6, 2026.