Hacker News debate: OpenAI and Anthropic's AI edge comes from data and training, not fundamental breakthroughs

A hobbyist asked whether leading AI labs are doing genuine research or just fine-tuning existing models. The consensus emerging from the Hacker News thread: OpenAI and Anthropic invest heavily in pre-training data sources (like Mercor for competitive programming examples) and synthetic data generation, rather than inventing new AI architectures.

The practical difference users see is narrow and task-specific — Opus and Codex feel smarter at coding because they were trained on more coding examples, while open-source and Chinese models perform similarly on general reasoning. This suggests that 'intelligence' improvements are largely about feeding models better examples of what you want them to do, not breakthrough algorithms.

For business professionals and engineers: if frontier AI companies are competing mainly on training data and fine-tuning rather than fundamental research, it means their advantages can erode faster as competitors gather similar data. Open-source models may catch up sooner than expected, lowering barriers for companies building internal AI tools and reducing lock-in to OpenAI or Anthropic products.