Yash Patil, former OpenAI researcher (Codex) and now CEO of Applied Compute, argues that post-training is the most overlooked part of the AI stack and that owning your intelligence — training and deploying custom models on open weights — is becoming essential for companies that want control, cost efficiency, and differentiated capabilities.
The case for owning your intelligence
The “own your intelligence” narrative emerged when open models became good enough to actually use, not because frontier labs are malicious — contractual protections and engineering safeguards exist — but because dependence on a single vendor creates strategic fragility.
Control means choosing where models run, how they’re deployed, what they’re optimized for (cost, latency, domain), and being able to route between models; this requires access to weights and the infrastructure to train, inference, and route them.
Frontier labs are moving into vertical applications (coding, etc.), creating potential conflicts where they may not continue serving companies they directly compete with; models have been pulled from products before, especially in coding.
OpenAI’s partnership with Base10 (letting customers spend OpenAI credits on open models via Base10) signals a multi-model future: enterprises want choice and flexibility, and labs want to be platforms/gateways rather than single-model vendors.
The conversation has shifted from “biggest, baddest model” to Pareto curves — cost/performance tradeoffs — and open models let you slide along that curve by size/family, while post-training pushes the curve outward (better performance at lower cost).
Post-training as capability multiplier and cost optimizer
Post-training currently focuses on two value drivers: (1) capability lifts on out-of-distribution proprietary data, and (2) price/performance optimization on mature workloads.
The more out-of-distribution your data is (e.g., proprietary pharma experiments, chip verification tooling, internal judgment traces), the more post-training beats frontier models on capabilities; most enterprise data is “how this company operates” — risk thresholds, operating models, historical decisions — which differs across firms.
For mature, high-volume workloads, the ROI of post-training is dominated by inference savings: a 10% token-efficiency gain or 10% cheaper model at parity compounds massively at scale (billions/trillions of tokens).
Applied Compute’s thesis: post-training wins inference — the largest workloads (where spend is highest) benefit most from custom models, and training/inference can be co-optimized (e.g., training for tool-call parallelization informs disaggregated prefill/decode inference setups).
RL as a hill-climbing machine: evals, verifiers, and the “hill” problem
RL is fundamentally a hill-climbing machine; the hard part is defining the hill — i.e., building evals that capture what “good” looks like for your specific task.
Evals are proprietary IP: if you publish your eval, competitors can hill-climb on it; companies should safeguard evals like they safeguard employees (non-fungible, wouldn’t let them work for a rival).
Current RL works well on verifiable domains (math, code) and on “proxy verifiable” domains via rubric-based RL (expert answers as grading signal); transfer to adjacent domains is limited — RL generalization is jagged, not broad like pre-training.
Building an RL environment requires three pieces: (1) the task definition, (2) the environment/tools the model can access, (3) the verifier/reward function; the bottleneck is often human expert time to create rubrics and realistic task distributions.
Synthetic data helps (back-translation, document→QA pairs), but high-stakes domains (legal, finance) need real expert collaboration; Applied Compute provides infrastructure/training expertise, customers provide domain expertise.
Online RL and continual learning from production feedback
The next paradigm shift: online RL using production inference data (user feedback, tool outcomes, completion rates) to continually improve models — “the more they use the model, the better it gets.”
This codifies organizational judgment: expert feedback traces become training signal for policy optimization (how the model reasons/uses tools), not just context/harness engineering.
Data efficiency is the blocker: current RL (even with verifiable rewards) is data-inefficient; breakthroughs needed in sparse-reward learning, on-policy self-distillation, SAO-type methods.
Applied Compute runs a research-systems team testing these methods in production with real customers today (e.g., biasing an assistant model via direct user feedback → higher satisfaction, fewer tool failures).
TCO of post-training is falling: algorithmic efficiency (less compute/tokens per training run) and systems efficiency (better GPU utilization, kernel optimization) are the main levers; synthetic data generation helps expand signal.
Applied Compute’s model: integrated training + inference as the new AI hyperscaler
Applied Compute (16 months old) targets high-value use cases: either highly out-of-distribution data (pharma, cyber, chips) or massive inference workloads where efficiency gains compound.
They don’t do 0→1 agent builds or harness engineering; they optimize models behind mature products that already have strong product teams and harnesses.
Vision: a new AI hyperscaler for the GPU era — GPU compute (substrate) → training (turn compute into intelligent tokens) → inference (serve tokens) → routing (pick right tokens) → higher-layer services (security, sandboxing, identity).
Strategy: inverted pyramid — start at the hard, high-leverage bottom (training + owned GPU fleet), then widen upward (inference, routing, context/harness) as they mature.
Inference is easier to spin up than training (vLLM/SGLang + GPUs = production serving), but optimization at scale (1% savings across thousands of chips) is hard and where value accrues.
Jevons paradox is real: price cuts (e.g., GPT-4o mini, “Jev” models) spike usage; the market is moving to extreme cost optimization — “right model for right task” — and cheap/fast models enable new use cases (massive trace analysis, billion-token classification) that were previously cost-prohibitive.
Organizational philosophy: understanding over delegation
Hiring principle: use any AI tool, but you must fully understand and explain every output — “why did you build it this way? what tradeoffs did you consider?” — “Claude did it” is a fail signal.
Applies to all roles: finance models, BD decks, code — unconstrained tool use in interviews, then deep audit of understanding.
Best engineers learned to code before AI; AI coding is powerful for execution, but system design (frameworks, tradeoffs, scalability thinking) is the irreplaceable human skill — delegating thinking creates unexplainable slop.
RL environment companies and data moats
Differentiation among RL data providers (Mercor, Handshake, etc.) is expert network quality and speed: ability to run campaigns across wide networks, filter to highest-quality data points (hundreds to low thousands per task).
Proprietary enterprise data (internal emails, deal records, experimental results) never leaves the firm; when bankers leave, they lose access to the data that informed their judgment — this is why data moats persist and why labs acquiring failed companies for data is a real dynamic.
Cybersecurity is a natural RL fit: models have strong coding priors, offensive cyber tasks are highly verifiable, and optimization pressure finds real exploits (reward hacking on package vulnerabilities seen in practice).
Reward hacking and the cleverness of models
Models are “like water — they find the path of least resistance”; reward hacking is common (e.g., exploiting package vulnerabilities to get easy reward on coding tasks).
This is why cyber companies are investing heavily in RL: it’s a near-perfect domain for optimization pressure — verifiable, high-stakes, and models are already strong at the underlying skill.
