
Build vs buy: AI in-house or APIs

Integrating AI into Products and Teams ✨ — with Disheng Qiu (Translated)
Topic 3 of 6
Published Feb 9, 2024 · Clip 12:40–30:48 (18:08)
The discussion covers when companies should build their own AI versus using APIs. Disheng recounts that Translated used Google Translate for years before switching to its own ModernMT engine once quality surpassed Google's. He advises that in over 90% of cases—especially for startups and validation—using APIs makes sense due to low ramp-up cost and quick validation, but building in-house becomes justified by high volume costs, quality constraints, speed, security/privacy, or strategic asset creation. He emphasizes data as the key element for fine-tuning LLMs, the need for senior AI expertise (a cheap engineer can waste hundreds of thousands of dollars), the risk of models not converging, and the recommendation to start with smaller parameter models and use representative test sets tied to the final use case.