LLMs Will Replace 8-Track Duplication Engineers
9 hours ago
- #NP-hard
- #audio engineering
- #human intuition
- The article discusses the historical practice of manually partitioning LP track listings into four programs for 8-track cartridges, an NP-hard balanced 4-way number partitioning problem solved by audio engineers without computers.
- Using data from Discogs and MusicBrainz (6,463 unique solves), analysis shows human engineers performed remarkably well: median solutions were only 5 seconds off optimal, with 27.5% achieving the exact optimum, often outperforming classical algorithms like LPT and Karmarkar-Karp (invented later).
- Different duplication houses had distinct 'styles'—some preferred splitting songs across programs, while others repeated short tracks to balance lengths, reflecting varied problem-solving approaches.
- LLMs (e.g., Fable 5, GPT-5.2) were benchmarked on the same partitioning task; models with reasoning capabilities matched or exceeded human performance, while those without reasoning performed poorly, highlighting the importance of reasoning in AI problem-solving.
- The project underscores how human intuition historically excelled at complex optimization tasks, and modern AI can now surpass it, though both remain somewhat opaque in their methods, offering a unique benchmark for evaluating AI against historical human expertise.