ThinkBase
AI generated art on the theme of "Connections and lines suggesting knowledge."
ThinkBase - Safe and Trustworthy AI
Many of the world’s governments are regulating to make AI safe and trustworthy UK,EU,US
The problem is that the most popular techniques used, Deep Learning and Generative AI, are intrinsically unsafe and untrustworthy, as many users are discovering. Much effort is required to make them safe for some applications.
These are not the only forms of AI. There are other strong threads of research with the same pedigree that are intrinsically safe and trustworthy. Our product, ThinkBase, uses the best of these and has dramatically simplified the process of using them.
It depends very much on the application: Generative AI is great for creating images, writing an essay, or having a purely speculative conversation. However, such tools hallucinate – they invent things - and don’t give the same answers twice. For any kind of regulated process, such as finance, medicine, HR, law and science, these flaws make Generative AI very problematic, and liable to an outright ban or heavy regulation in some jurisdictions. ThinkBase cannot write an essay for you, but it can make decisions and predictions in a way that is completely transparent, repeatable and explainable. These properties make ThinkBase safe and trustworthy.
Knowledge graphs can lower development time dramatically
Enormous effort has been put into creating Large Language Models. Many applications just add RAG post processing to create functionality, which is a relatively simple process. For regulated applications, however, this is not enough. The developer must rigorously test the application. Creating a substantial test set is in some cases impossible, and in all but trivial cases very expensive. In most cases creating an application requires an analysis of the domain, creation of a training data set, training of the model and post training analysis and testing.
ThinkBase models can be learned from data or constructed by hand. In the latter case the domain analysis can be turned directly into a knowledge graph. ThinkBase can automatically test for lacunae – sets of inputs which don’t create an output – and the intrinsic explainability of the model makes testing for biases a process of inspection.
For regulated applications, therefore, ThinkBase can be dramatically cheaper in development time compared to Deep Learning and Generative AI.
Editing Knowledge graphs
Using the GraphQL API
ThinkBase versus popular AI technologies
| Feature | ThinkBase | Deep Learning | Generative AI |
|---|---|---|---|
| Prompt based generation of text and images | No | No | Yes |
| Can perform image processing | No | Yes | Yes |
| Supervised, unsupervised and reinforcement learning | Yes | Yes | Yes |
| Transparent and explainable models | Yes | No | No |
| Small model size | Yes | Only for trivial models | No |
| Manual construction of models | Yes | No | No |
| Completeness checking | Yes | No | No |
| Repeatable results | Yes | Yes | No |
| Low computation overhead | Yes | Only for trivial models | No |
| Interaction via API | Yes | Yes | Yes |
| Interaction via Chat interface | Yes | No | Yes |
| Low training/development costs | Yes | No | No |