ChatGPT has taken the world by storm. Never before has an app gained so many users in such a short time. In this blog, we examine the current state of generative AI, highlight its challenges, and offer a glimpse into the future.
Generative AI today
Hardly a week goes by without a new announcement regarding Generative AI. At the beginning of November, OpenAI, the company behind ChatGPT, released its latest version of ChatGPT. With the ChatGPT Plus subscription, it's now possible to set up your own chatbots without any programming knowledge.
Infometis seized the opportunity and created a small Infometis Support GPT tool that can answer questions about TACC & TAMI . You can test it with a ChatGPT Plus account at the following link: https://chat.openai.com/g/g-hW9CfmCwf-infometis-support-gpt
Challenges of Generative AI
Generative AI has become an integral part of our everyday lives. However, it faces several challenges, including the following (but not limited to):
Hallucinations of answers: Generative AIs tend to invent answers when no precise answer is possible.
Transparency and trust: The process by which generative AI generates answers is often not transparent. How can we then trust the technology?
Copyright issues: Generated material is often based on copyrighted material, which can lead to legal problems.
High resource requirements: The deployment of Generative AI requires a lot of computing power and causes high hardware and electricity costs, which also has an impact on the environment.
Static data set: The data used to train Generative AI is statically limited. Even with the pioneering ChatGPT, the data set still lags behind the present by several months.
Retrieval Augmented Generation (RAG) as a solution approach
In 2020, Meta published a research paper on Retrieval Augmented Generation (RAG). Since then, this approach has gained increasing traction. It represents an innovative and pragmatic solution to the current challenges of generative AI. RAG combines information retrieval with generative AI to generate more precise and informed answers. By incorporating current information from predefined sources, GenAI models can reduce the likelihood of "hallucinatory" responses. RAG also improves transparency in the decision-making process of generative AI, as the information sources used are traceable.
How does RAG work?
RAG works in two steps: First, the 'retriever' identifies relevant information, usually from a vector database. A vector database stores information in the form of vectors, i.e., multidimensional arrays that represent data points in a vector space. These vectors contain a condensed representation of information, which can originate from text, images, or audio, for example. Vector databases enable the fast and efficient retrieval of information by comparing vectors.
The generator then uses generative AI, based on this information, to create a coherent and informative answer. This method makes it possible to generate answers based on the latest and most relevant information.
How can RAG solve the problems of Generative AI?
Reduction of hallucinations: By using current and relevant information from self-compiled sources, RAG can reduce Generative AI's tendency to 'hallucinate' answers. This leads to more precise and reliable responses.
Improved transparency and trust building: The traceability of the information sources used by RAG enables a more transparent presentation of the response process. This strengthens user trust in Generative AI.
Minimizing copyright problems: RAG makes it possible to more effectively identify and avoid copyrighted material by drawing on a wider range of information sources that are free from copyright restrictions.
More efficient resource utilization: RAG eliminates the need to retrain an entire LLM. Only the vector database needs to be kept up-to-date, allowing for more efficient use of computing power and thus reducing costs and environmental impact.
Handling static training data: By incorporating easily updatable data sources into the generation process, RAG can always take current information into account and thus provide more relevant answers.
The next steps in Generative AI
One of the co-inventors of RAG, Douwe Kiela, highlighted three major developments in his presentation at the Web Summit :
- Integration into systems: AI is increasingly being integrated into more comprehensive systems, rather than being used in isolation.
- Enhanced multimodality: We will see increased development of AI systems that go beyond text and are capable of working with various media such as images, videos, and audio files. These multimodal AI systems could, for example, generate realistic virtual worlds or create complex multimedia content for education and entertainment.
- Specialized AI: We will likely see a shift from general, versatile AI models to specialized AI solutions tailored to specific industries or tasks. This makes them faster and more efficient.
The future with Generative AI looks promising. Developments such as RAG, multimodality, and specialized AI have the potential to significantly improve the efficiency, accuracy, and versatility of AI solutions.
Interested in using AI in business, process automation, or testing? Feel free to contact us for more information.
