73% of Artificial Intelligence startups are just prompt engineering

A widely shared post claims the author reverse-engineered 200 Artificial Intelligence startups and found 73 percent were effectively thin wrappers around provider models. Hacker News commenters debated the methodology, the value of prompt and context engineering, and whether those companies have defensible moats.

Polish outperforms English and Chinese in long-context large language model tests

A new OneRuler benchmark presented at COLM 2025 finds Polish achieves the highest accuracy when large language models process very long documents. The study links the advantage to tokenization and script differences rather than dataset volume.