In a recent interview, the CEO of OpenAI shared his fundraising and development ambitions. Its ambition: to make AI as intelligent as humans. The giant, thanks to a personalized GPT function, is also tackling the creation of chatbots dedicated to specific tasks.
OpenAI thus encroaches on the terrain of open source models, such as Mistral 7B, the first generative AI model from Mistral AI, or those of LightOn, another French player. The resistance is organized and LightOn announces a new version of its LLM Alfred.
Fewer contextual hallucinations
The Frenchman today unveiled Alfred-40B-1023, the latest iteration of his open source language model. LightOn’s positioning consists of developing solutions dedicated to businesses and public services.
The new model was trained on Amazon SageMaker. Alfred-40B-1023 will soon be available on HuggingFace and AWS Jumpstart, which he believes will facilitate its integration into other work environments.
For the general director and co-founder of LightOn, Laurent Daudet, the continuous improvement of ongoing LLMs “contributes to increasing the adoption of generative AI by strengthening the efficiency and quality of the tasks that can be carried out.”
In this context, Alfred would therefore bring his share of improvements. LightOn places particular emphasis on reducing contextual hallucinations, the main problems from which general models such as those of OpenAI suffer.
Take care of the use case of documentary research
The French publisher also claims better self-knowledge for its technology. More concretely, this means that the model, when it does not have a satisfactory response, is programmed not to respond.
“I don’t know”, Alfred V2 can indicate, “which improves its transparency and reliability” for its designers. LightOn also takes care of a very popular use case in companies: searching in documentary databases.
This feature called “Chat with docs” has been improved. The open source LLM is specially trained for these tasks consisting of a user asking a question in natural language on a database of internal documents.
“The model thus simplifies interaction with documents and optimizes the search for information,” says LightOn. Finally, the editor highlights an expanded context (8192 tokens), giving the model the possibility of understanding and generating longer and more complex content.
Additional functions on Paradigm
Still to respond to the concerns of corporate customers, LightOn would also attach more importance to data security.
“When deployed in a customer’s infrastructure (on-premises or in a private cloud), all data interactions are confined to that environment, ensuring data confidentiality and security,” explains he.
But LightOn is a dual approach in the field of generative AI: autonomous open source models and also a complete platform called Paradigm. By opting for the version of Alfred v2 integrated into Paradigm, users have a “more advanced version”.
The publisher cites “additional enhancements and features, ensuring they always benefit from the highest performance capabilities.” Through its platform, LightOn also intends to simplify the use of generative AI.
“We also focused on creating intuitive interfaces and workflows, so that everyone, regardless of technical expertise, can harness the power of generative AI. Features like “Task Factory” thus allow non-experts to integrate AI into their daily tasks in a transparent manner,” declares Laurent Daudet.