Should Your Company Create Its Own LLM?

Should you train your own large language models from scratch? Likely not. OpenAI has poured hundreds of millions into their GPT models' training, an investment beyond the reach and necessity for most companies. However, unique circumstances might require such a move because OpenAI has unintentionally degraded GPT-4’s performance. While evaluating early releases of GPT-4, Microsoft researchers observed its performance degrading as OpenAI introduced safety guidelines. While these guidelines, aimed at preventing misuse for harmful activities like bomb creation, commendably protect public safety, they inadvertently impair performance.

These constraints can hinder the models' utility in specific scenarios. Imagine a business needing a custom model for competitive advantage. Although I lack concrete examples, it's easy to foresee such needs, given how often companies make high-stakes investments.

Certain organizations, like intelligence agencies, law enforcement, military units, and compliance departments, might find customized models a necessity. In situations where the goal is to outpace wrongdoers or detect illicit activities, OpenAI-like safety guidelines could reduce the model's effectiveness.

I personally encountered this limitation during an experiment. I aimed to create scenarios illustrating how GPT models could aid money laundering schemes. I found OpenAI had put in place guardrails that barred reference to specific cryptocurrencies and networks relevant to the task. OpenAI opted for generic models or cryptocurrency platforms. Understandable for a public model release, these guidelines offer little to a bank's compliance department needing detailed illicit behavior insights.

However, I wouldn't advise most companies to invest in customized model training at this time. The technology is evolving too rapidly, and OpenAI or competitors may soon offer more potent models without the restrictions of those designed for public use.