How Post‑Training Turns General AI into Enterprise‑Specific Intelligence

Bottleneck Shifts and Data Efficiency

AI progress follows a bottleneck‑shift pattern: 2012‑2016 data & compute (ImageNet/AlexNet), 2017 Transformers, 2018‑2021 compute scale, 2022‑2024 alignment (RLHF/ChatGPT). Each shift reduces the data required for the next breakthrough. The 2024 release of o1 introduced Test‑Time Compute —model size stays constant while inference uses more compute to “think”. RLVR (Reinforcement Learning with Verifiable Rewards) places the model in a constrained RL environment (e.g., math, coding) and trains reasoning ability with a verifiable reward function, without step‑by‑step annotations. DeepSeek R1 reproduced this path using GRPO (Group Relative Policy Optimization), applying RL directly to the base model (R1‑Zero) and observing emergent chain‑of‑thought. Open‑source reproductions confirmed RLVR’s broad applicability.

The data‑efficiency curve shows:

Pre‑training: trillions of tokens.

SFT: tens of thousands of labeled examples.

RLHF: thousands of preference pairs.

RLVR: only a verifiable reward function plus compute.

Thus, later scaling directions achieve larger performance gains with dramatically less data.

Business Knowledge Gap

“They’re like smart geniuses that know nothing about your business.”

Pre‑training compresses public internet knowledge; enterprise‑specific decision logic never appears in that data. Concrete examples cited: Goldman Sachs risk‑control rules, Kirkland & Ellis contract‑review criteria, Cognition’s internal bug definition. Even larger future models would not close this gap because business logic is fragmented across organizations.

Eval Definition and RL Optimization

Enterprises need only define a quantitative evaluation function (Eval) that captures “what is good”. Reinforcement learning then maximizes this Eval.

Case studies:

DoorDash menu extraction : Convert unstructured restaurant menu images to a standardized format following a detailed style guide (modifiers, add‑ons, mix‑and‑match). Pipeline: model output → human correction → error‑rate calculation → RL optimization of error rate.

Cognition bug detection : Detect bugs within 2 seconds. A large model meets accuracy but exceeds latency; a small model meets latency but misses bugs. Post‑training a small model on the bug‑detection task achieves both high accuracy and low latency.

Both examples illustrate the pattern: define a concrete Eval (error rate, bug‑capture rate) and let RL drive the model toward optimal performance without the enterprise needing RL expertise.

Compute Economics of Post‑Training

DeepSeek V3 pre‑training consumed ~2.66 million H800 GPU‑hours. DeepSeek R1 RL training used 512 H800 blocks (~80 GPU‑hours), less than 2 % of the pre‑training compute, yet yielded performance gains far exceeding the compute ratio. Pre‑training optimizes next‑token prediction, producing gradients that rarely align with any specific downstream task. RL training repeatedly trials the target task, so each gradient directly improves the desired capability, giving orders‑of‑magnitude higher marginal returns. Scaling laws for post‑training indicate that larger batch sizes, more rollouts, and longer inference chains continue to improve performance, and current workloads are far from the diminishing‑returns inflection point.

Reward Extensions (RLVRR)

RLVRR (Reinforcement Learning with Verifiable Reference‑based Rewards, ICLR 2026) decomposes rewards into:

Content Reward : Extract key entities from a reference output, allowing flexible generation.

Style Reward : Verify generated Python scripts for syntactic correctness.

This extends RLVR from binary‑answer tasks to reference‑based tasks, enabling enterprise scenarios such as PPT generation where functional correctness and aesthetic quality are jointly optimized.

Continual Learning Paths

Cursor online weight updates : Composer, built on the open‑source Kimi K2.5 model, collects implicit feedback (accept/reject of code suggestions, rollback events). Large batches aggregate billions of tokens of interaction, compute directional gradients, run an eval suite to guard against regression, and deploy a new checkpoint. The full loop takes ~5 hours, allowing multiple daily iterations.

Applied Compute context‑based method : Offline agents analyze interaction traces to identify decision patterns that lead to good outcomes, then inject those learnings as context without changing model weights. This avoids catastrophic forgetting but is limited by context‑window size and the quality of the analysis agent.

Trade‑offs: weight updates provide lasting capability changes but risk forgetting and require sustained compute; context injection is lightweight and safe but cannot alter the underlying model.

ROI Multipliers for Post‑Training

Three multiplicative factors determine ROI:

Data exclusivity : Proprietary data and eval standards (e.g., DoorDash’s menu rules) create a moat.

Eval formalizability : Ability to encode “what is good” as a computable function (error rate, bug‑capture rate).

Task frequency : Number of task executions per unit time spreads the one‑time training cost.

Break‑even formula:

Training cost ÷ (token savings + consistency premium + latency constraint)

High‑frequency, high‑compliance, low‑latency tasks (e.g., DoorDash menu processing) fall on the post‑training side; low‑frequency or low‑exclusivity tasks favor prompting + RAG.

Capex vs Opex Boundary and Evolution

Post‑training is a capital expense (one‑time training) that eliminates recurring inference token costs. Prompting + RAG is an operational expense, incurring token cost on every request. When the number of task executions exceeds “training cost ÷ per‑inference token savings”, post‑training becomes cheaper. Additional structural advantages of post‑training include:

Consistency : RL compresses the policy distribution, reducing output variance compared with in‑context learning, which is sensitive to prompt variations.

Latency : Short prompts after post‑training lower both latency and cost; RAG‑based approaches cannot meet sub‑second latency for tasks like Cognition’s bug detection.

As enterprises accumulate enough proprietary data, post‑training can evolve into domain‑specific pre‑training. Examples:

Cursor announced a joint effort with SpaceX/xAI to pre‑train a 1.5 trillion‑parameter coding model on 10 × 10⁴ GPU blocks, using 4 million developer interaction signals.

Harvey AI’s roadmap describes moving from post‑training + eval to full domain pre‑training once data volume reaches a critical threshold.

Most enterprises lack such data scale, so the “general model + post‑training” paradigm remains optimal.

References

Stanford MS&E 435: Economics of the AI Supercycle, Spring 2026. https://www.youtube.com/watch?v=LRGX-gTegVA

DeepSeek‑AI, “DeepSeek‑V3 Technical Report,” arXiv:2412.19437 (2024). https://arxiv.org/abs/2412.19437

DeepSeek‑AI, “DeepSeek‑R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,” arXiv:2501.12948 (2025). https://arxiv.org/abs/2501.12948

Guo et al., “DeepSeek‑R1 incentivizes reasoning in LLMs through reinforcement learning,” Nature 645 (2025). https://www.nature.com/articles/s41586-025-09422-z

DeepSeek‑AI, “DeepSeek‑V4 Technical Report,” April 2026. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf

Jiang et al., “From Verifiable Dot to Reward Chain: Harnessing Verifiable Reference‑based Rewards for RL of Open‑ended Generation (RLVRR),” ICLR 2026. https://arxiv.org/abs/2601.18533

Applied Compute launch announcement, Oct 2025. https://siliconangle.com/2025/10/30/former-openai-researchers-launch-applied-compute-80m-funding/

Cursor (Anysphere), “Improving Composer through real‑time RL,” March 2026. https://cursor.com/blog/real-time-rl-for-composer

Reuters, “SpaceX locks in $60 billion Cursor deal,” June 16 2026. https://www.reuters.com/legal/transactional/spacex-buy-anysphere-60-billion-2026-06-16/

Ars Technica, “SpaceX to acquire AI coding platform Cursor,” June 2026. https://arstechnica.com/ai/2026/06/spacex-will-acquire-coding-tool-cursor-to-compete-with-anthropic-openai/

Harvey AI, “Harvey Raises at $11 B Valuation,” March 2026. https://www.harvey.ai/blog/harvey-raises-at-dollar11-billion-valuation-to-scale-agents-across-law-firms-and-enterprises

Pereyra, G., “Model strategy for Harvey,” X post, June 2026. https://x.com/gabepereyra/status/2067324200801452105