Why did we invest (WDWI) in H, the France-based startup building Action Models, AI’s Next Wave

by Adrien Muller & Angélique Tézé

With €2.5 billion asset under management, Bpifrance’s Large Venture fund is on a mission to propel French tech scale-ups into their next growth phase with funding rounds from €20 million. This investment capacity allows us to back French AI gems with significant financing needs, such as H, which has just completed an initial fundraising of $220 million. H is not just playing the AI game; they’re redefining it with their Action Models, which will be able to generate series of actions beyond simple text and image processing.

A Team Gathering the World’s Best AI Experts…

The founders’ profile remains a critical factor in investment decisions. With H, the founding team takes center stage: it comprises world-class researchers in AI, specializing in a wide range of domains (scaling language models, reinforcement learning, multi-agent systems, planning, memory, causality, etc.). Only a few researchers worldwide can push the boundaries of AI, and we firmly believe that H’s founders are among them.

The founding team was trained in the leading scientific French institutions before joining Google DeepMind for most of them.

• Charles (Chief Executive Officer) co-founded H at the end of 2023, shaping its vision and strategy. He received his education at Paris-Dauphine (PSL University), Centrale Supélec (Paris-Saclay University), and Stanford.
• Laurent (Chief Technological Officer), a graduate of the École Polytechnique and the École Normale Supérieure, spent a decade at DeepMind, leading several key research projects in the field of GenAI and deep neural networks. These projects included AlphaGo, AlphaFold, AlphaStar, Chinchilla, Gemini, and Gemma.
• Karl (Chief Research Operations), a renowned scientist in the multi-agent systems community, led DeepMind teams focused on game theory and multi-agent systems starting in 2017. He initiated and directed well-known projects such as TacticAI, a football coach assistant.
• Daan (Chief Scientist) is a founding researcher at DeepMind and a pioneer in reinforcement learning techniques. At DeepMind, Daan led a team of over 100 people for many years, playing a crucial role in shaping DeepMind’s initial research direction and contributing to the invention of most AI paradigms (such as VAEs, etc.).
• Julien (Chief Multi-Agent), an expert in multi-agent reinforcement learning and game theory, co-led the scientific and technical development of DeepMind on the game of Stratego (DeepNash).

H has the ability to attract top talents to innovate in this cutting-edge sector. It’s one of the reasons why we chose to invest alongside leading international partners such as Accel, UiPath, Eric Schmidt, Xavier Niel, Amazon, FirstMark, Elaia Partners, Yuri Milner, Aglaé Ventures, Creandum, Eurazeo, Motier Ventures, Samsung, and Visionaries Club, among others.

…To Pioneer the Next Generation of AI…

The founders of H are not looking to create yet another generative AI startup. They are tackling a new subject — Action Models, fundamental models focused on “action” and “multi-modality” — through which they intend to contribute to the development of a general artificial intelligence (AGI).

Today, most large companies concentrate their resources on developing large-scale language models (LLM), which are often proprietary and monolithic. These models, such as ChatGPT, Gemini, or those offered by Anthropic, Inflection, Cohere, Character.ai, or Adept, are based on “transformer” architectures and massive data from the internet, often enriched by techniques like Reinforcement Learning from Human Feedback. However, this approach has its limits and is being caught up by open-source initiatives like Mistral.

We are firmly convinced that the real value lies in combining (i) multi-modal action models with (ii) the business processes of companies, closely connected to real-world data.

With multi-agent models, several agents can coordinate to automate complex tasks, unlike chatbots that generate only a single output. With these models, one agent can generate results while another provides constructive feedback by using tools, such as running tests or conducting web searches to verify the accuracy of the results. The multiplicity of agents and their coordination thus improve the performance of AI through feedback and iterative discussions.

By integrating directly into the business processes of companies, H aims to significantly improve their productivity. For example, while more than half of automation projects today fail to reach their full potential, combining AI with Robotic Process Automation (RPA) would address the two main limitations of this field: dependence on legacy tools and the need to work with structured data. Concrete application cases would include invoice processing and verification, customer claims management, or online order and appointment scheduling.

… and all of this, from France

H has chosen to establish itself in Paris, confirming the attractiveness of the French ecosystem in the field of AI. This attractiveness has been recognized by US tech giants like Meta or Google, who have been establishing their AI research centers in France for nearly a decade. The quality of our scientific talent as well as the financing and support mechanisms for R&D provided by Bpifrance (€5 billion over 5 years) explain this power of attraction. France has fully grasped its potential, as evidenced by President Macron’s announcement at the gathering of the greatest French AI talents on the eve of VivaTech.

This investment fully aligns with Bpifrance’s strategy, which aims to promote French AI startups. After investing in Mistral.ai, Poolside, or Bioptimus — Owkin’s spin-off, Large Venture’s investment in H’s first round reaffirms our dedication to nurturing this thriving ecosystem.

Now, more than ever, we firmly believe that France has a leading role to play in this cutting-edge sector. There are numerous opportunities, whether in the development of foundational models, the creation of application layers, or on the infrastructure side.

Can’t wait to see what our entrepreneurs will accomplish!