Marble also points out that for many applications, the superiority of a GPT-like model is not what drives value, and smaller models can be cost-effective and competent. He emphasizes that responsible use of AI requires access to deep knowledge of the technology, not just the API reference. While APIs can be useful for certain products, for a new and rapidly evolving technology like AI, deep access to models and code is essential for real participation.
Key takeaways:
- Despite the cost and performance advantages of using APIs from companies like OpenAI or Anthropic for language models, the author argues that self-hosting models can be beneficial, especially for product or internal capability development.
- Self-hosting models offer control over model architecture and weights, removing uncertainty about future changes and allowing customization. This approach allows for a long-term relationship with the AI model and deeper integration into products.
- While GPT-4 and similar models are superior in many applications, smaller models can be cost-effective and competent for many tasks. They can be run on local systems, offering more flexibility and control.
- Working with self-hosted models provides valuable experience in the rapidly evolving field of AI. The author suggests that organizations making significant use of AI should have deep knowledge of the technology, beyond just API references, to fully understand its capabilities and potential applications.