AI yesterday, today, tomorrow: our convictions
The topic of artificial intelligence is on everyone’s lips today. Bpifrance Digital Venture was one of the pioneers among VCs who believed in AI. So, this is the opportunity to share our experience of investing in AI start-ups and our convictions for the future.
Our success stories in AI
The Bpifrance Digital Venture team fell in love with AI with the advent of deep learning and supervised learning. As early as 2015, we recognized the full potential of AI in the field of health, particularly in diagnostics, thanks to its image recognition capabilities. This led us to invest in Cardiologs (AI-assisted diagnosis and monitoring of heart diseases), which was later acquired by Philips in 2021. We continued our AI investments in healthcare with Gleamer (AI applied to radiology), Primaa (AI applied to cancer detection), Kiro (AI applied to medical biology), Callyope (AI applied to psychiatric disorder detection), and Louise Life (AI to support patients undergoing medically assisted procreation).
The use of AI has also proven successful in the industrial domain. That’s why we invested in Extrality (an AI-powered numerical simulation software company) in 2020. In 2023, it was acquired by a large publicly traded company.
We also supported Mindee in revolutionizing automated document processing. Additionally, we have become convinced that synthetic data generation would become crucial for training AI models, and that’s why we’ve invested in AI Verse.
Our experience with the limits of AI
Our journey as investors in AI has also led us to experience the limits of AI. Surprisingly, these limits were less technical and more related to its applications. We invested in Vekia (AI for inventory optimization and supply chain automation) as early as 2017, and the commercial deployment did not go as fast as we expected. For AI to be perfectly efficient, users must have had to follow the AI’s recommendations to the letter, but it’s difficult for a purchasing manager who knows their job to rely entirely on AI. The purchasing manager prefered using Vekia as a copilot but it defeated the purpose of the solution and its intented value proposition.
We also faced the ethical limits of AI. Our former portfolio company Easyrecrue is an example. They developed a video screening and interview software designed to streamline candidate assessment processes by providing pre-recorded and live video interviews. They also developed a highly efficient AI capable of analyzing videos and scoring candidates to help recruiters prioritize which videos to watch. This enters the realm of AI now considered as high-risk under the AI Act. This kind of AI could even be prohibited as the AI act prohibits emotion recognition systems in the workplace. At that time, there was no regulation for AI but the ethical risk associated with the use of this AI for scoring candidates was ultimately deemed too significant compared to the value it provided for employers. The feature was not released, which did not prevent us from making a very successful exit when Easyrecrue joined forces with iCIMS.
The revolution of Large Language Models
When the wave of Large Language Models (LLMs) arrived after the release of ChatGPT in late 2022, we quickly seized the opportunity by investing in Mistral (foundation model) and Poolside (Ai for software development).
We observed that generative AI could also amplify the capabilities of an existing solution. This led our portfolio company iAdvize to enhance its conversational platform by creating an Agent Copilot, which increased the productivity of teams handling customer responses, as well as a Shopper Copilot, automating certain conversations without compromising the customer experience. Our portfolio company OpenClassrooms, which is an online school that offers more than 600 online courses and 50 complete training programs, has launched an AI assistant to provide real-time support to its students between tutoring sessions.
AI tomorrow : an AI capable of planning and reasoning?
AI continues to advance, and we already glimpse the future milestones it could achieve. Yann Le Cun, Chief AI Scientist at Meta, demonstrated the future of AI by introducing the V-JEPA model in February 2024. His vision is that AIs must learn a world model to be capable of reasoning. LLMs do not yet possess this ability; they are merely trained to predict words in a sentence. Future AIs will need to be objective-driven and able to predict sequences of actions. To achieve this, they must learn to predict sequences of images. However, not pixel by pixel, as that would be resource-intensive and inefficient due to irrelevant details often present in videos. It is within this context that Yann Le Cun developed V-JEPA, a non-generative and self-supervised model that learns by predicting missing or masked parts of a video in an abstract representation space. V-JEPA has the flexibility to discard unpredictable information, which leads to improved training and sample efficiency by a factor between 1.5x and 6x.
What are the next steps? Developing a multimodal AI that can predict not only images but also audio along with the visuals. And then demonstrating that one can use this kind of a predictor or world model for planning or sequential decision-making.
As for us, we are more active than ever in investing in AI startups, and we will participate in the upcoming revolutions of AI.
