Algorithms like GRPO are powerful but require parallel rollouts in a reproducible environment. Building and maintaining these high-fidelity sandboxes, complete with realistic data and failure modes, is the hardest part of implementing RL today and a significant barrier for most companies.

A key surprise in AI development was the non-linear impact of scale. Sebastian Thrun noted that while AI trained on millions of documents is 'fine,' training it on hundreds of billions creates an 'unbelievably smart' system, shocking even its creators and demonstrating data volume as a primary driver of breakthroughs.

The history of AI, such as the 2012 AlexNet breakthrough, demonstrates that scaling compute and data on simpler, older algorithms often yields greater advances than designing intricate new ones. This "bitter lesson" suggests prioritizing scalability over algorithmic complexity for future progress.

The "bitter lesson" in AI research posits that methods leveraging massive computation scale better and ultimately win out over approaches that rely on human-designed domain knowledge or clever shortcuts, favoring scale over ingenuity.

When OpenAI started, the AI research community measured progress via peer-reviewed papers. OpenAI's contrarian move was to pour millions into GPUs and large-scale engineering aimed at tangible results, a strategy criticized by academics but which ultimately led to their breakthrough.

Contrary to the "bitter lesson" narrative that scale is all that matters, novel ideas remain a critical driver of AI progress. The field is not yet experiencing diminishing returns on new concepts; game-changing ideas are still being invented and are essential for making scaling effective in the first place.

Much RL research from 2015-2022 has not proven useful in practice because academia rewards complex, math-heavy ideas. These provide implicit "knobs" to overfit benchmarks, while ignoring simpler, more generalizable approaches that may lack intellectual novelty.

OpenPipe's 'Ruler' library leverages a key insight: GRPO only needs relative rankings, not absolute scores. By having an LLM judge stack-rank a group of agent runs, one can generate effective rewards. This approach works phenomenally well, even with weaker judge models, effectively solving the reward assignment problem.

Instead of pinpointing which specific action led to a good outcome, the GRPO algorithm solves the credit assignment problem with a simple heuristic: it assumes rare tokens in a high-scoring output were responsible and upweights all of them. This "unsatisfying" but practical approach works surprisingly well.