Developing generative AI solutions
Selecting an FM
Pre-trained models give a head start and converge faster when fine-tuned, but they can carry bias or miss domain nuances. Choose based on the use case.
Selection criteria
| Criterion | What to check |
|---|---|
| Cost | Licensing, inference compute, and customization costs against the benefit |
| Modality | Text, image, audio, or multimodal output |
| Latency | Real-time apps (such as live translation) need fast inference |
| Multilingual support | The languages you need, or adaptability through transfer learning |
| Model size | Bigger models do better on complex tasks but need more compute |
| Model complexity | Complex models handle advanced tasks but are harder to deploy and optimize |
| Customization | Whether you can fine-tune it, and the data and compute that takes |
| Input/output length | Maximum sequence length for long documents or long outputs |
| Responsibility | Bias, misinformation risk, misuse, and training data sources |
| Deployment and integration | Fit with your infrastructure, tools, and libraries |
Amazon Titan
Amazon's own FMs come in three types: embeddings, text generation, and image generation. Use them as is or fine-tune them with your data.
Keep revisiting your criteria, because the model landscape changes fast.