Why Reinforcement Learning Is the Hot New Frontier—and Why You Shouldn't Start a Startup Around It

Today we discuss reinforcement learning (RL).

In AI, few technologies evoke as mixed emotions as RL, which first entered the public eye when AlphaGo defeated Lee Sedol, but truly went mainstream thanks to ChatGPT.

ChatGPT uses Reinforcement Learning from Human Feedback (RLHF) to align AI responses with human preferences, making it a more useful assistant than earlier GPT‑3 models.

Consequently, RL has transformed from a niche technique into a technology coveted by every model‑building company.

Before RLHF became popular, large models followed a pipeline of self‑supervised pre‑training (MLE) → supervised fine‑tuning (SFT/multi‑task/instruction tuning) → inference‑time constraints and safety filters, with little use of reinforcement learning.

RL essentially lets AI learn through "trial‑and‑error"—like a child learning to ride a bike by falling and balancing—without being told every step.

Technically, RL consists of four core elements:

Agent : the decision‑making AI system

Environment : the world the agent interacts with

Actions : the operations the agent can perform

Rewards : feedback signals from the environment

RLHF aligns AI with human "taste" through three steps:

1. Collect human preference data : annotators compare AI answers and select the better one.

2. Train a reward model : use the preference data to build a scoring system.

3. Optimize the language model : fine‑tune the model to produce high‑scoring responses.

OpenAI’s o1 and DeepSeek’s R1 push this pattern to new heights.

RL environments have become a hot commodity, with companies buying custom simulators for tasks like ordering food or managing a CRM.

However, these environments are consumable; once a model masters a task, the environment’s commercial value drops to zero.

Developers risk becoming “scaffolding workers” for AI, building temporary tools that models will soon replace.

Even large‑scale RL environment platforms are emerging as open‑source collections, turning the market into a free‑for‑all.

Experts such as Andrej Karpathy and Yann LeCun caution that RL’s impact may be limited compared to self‑supervised learning, and that the data requirements for RL are massive.

In the pre‑training era, the internet text mattered; in the supervised fine‑tuning era, high‑quality Q&A pairs mattered; today, we need "environments".

Karpathy argues that while environments are crucial, building a business around them is not advisable—just because a market is promising doesn’t mean you should start a venture serving it.

The article concludes that the true value lies in creating environments that enable models to acquire transformative abilities, not in merely selling disposable simulators.