While many artificial intelligence startups are chasing the language model boom, French startup H is betting on a different future: AI agents that can take real-world actions.
Speaking at the ai-Pulse conference in Paris, H CEO Charles Kantor lifted the veil on his secretive company's ambitious plans to build what he calls "action models" - the next frontier in AI development.
H, which raised an impressive $220 million Seed Round last May, has kept a low profile while developing its technology using a large training cluster from Scaleway. And its debut has been rocky: Co-founders Karl Tuyls, Daan Wierstra, and Julien Perolat left just three months after that fundraising citing “operational differences." That subject wasn't broached on stage.
When asked about the mysterious company name, Kantor playfully suggested multiple interpretations – holistic, humane, hectic, and historic – reflecting the company's broad vision for AI's future.
"We think of H as a name which is up to you," he said.
The Three Waves of GenAI
Kantor broke down the generative AI landscape into three distinct waves. The first was dominated by language models powering chatbots and customer support applications, with cloud providers leading the way before on-premises solutions followed for sensitive industries like healthcare and finance.
The second wave focused on code generation, exemplified by GitHub Copilot's code completion capabilities, again followed by on-premises solutions for regulated industries. But it's the third wave – action models for automation – where H sees the biggest opportunity.
"This is a huge opportunity to disrupt productivity," Kantor explained. Unlike traditional language or multimodal models, action models combine specialized large language models (LLMs) for generating plans and workflows with vision models that can understand interfaces - not just identifying objects but comprehending how to interact with desktop environments, web interfaces, and mobile devices.
Building an AI Operating System
H envisions itself as more than just another AI company – it aims to become an operating system for AI actions. The system works by orchestrating different specialized models: LLMs handle planning and workflow generation, while vision models interpret interfaces across various platforms. A sophisticated routing system coordinates between these models to execute complex tasks.
For example, an H agent might handle a workflow that involves interfacing with Gmail, connecting to Excel, and pulling information from LinkedIn. The key innovation isn't just executing these tasks but doing so reliably and with the ability to self-heal when errors occur.
"We have not yet arrived at AGI," Kantor said, referring to artificial general intelligence. "The question is, if I have a very good product, is it good enough to be modified such that it's always successful?" Rather than pursuing perfect first-time performance, H focuses on creating systems that can be edited, validated, and replayed until they achieve near-perfect reliability.
Measuring Success
How do you benchmark an AI system designed to take real-world actions? H approaches evaluation on three levels:
- Underlying model performance - how well the base language and vision models perform their specialized tasks.
- Framework effectiveness - measuring the success rate of complete action workflows.
- Product resilience - testing the system's ability to self-heal and recover from failures.
The French Advantage
Despite fierce international competition in AI, Kantor sees H's French origins as a significant advantage. "We have an amazing pool of talents in France," he noted, citing the country's strong educational foundation in mathematics and computer science. This advantage hasn't gone unnoticed – many American tech companies have established research offices in France to tap into this talent pool, he said.
Opening the Ecosystem
While H has been secretive until now, Kantor said the company is committed to building an open ecosystem. "We view ourselves as an ecosystem," he explained. "We will need the help of a lot of builders in terms of outputs, structuration, and underlying models."
The company plans to release modular aspects of its system by year's end, inviting developers to build specialized applications across various industries including robotic process automation (RPA), quality assurance (QA), and business process outsourcing (BPO).
For developers and AI enthusiasts interested in joining this ecosystem, H has opened a waitlist for its upcoming beta release. While many details remain under wraps, the company's emergence from stealth mode signals that practical demonstrations of its technology may soon be available to the public.
"We invite all AI builders to join us in this ecosystem," Kantor said. "We are paving the way for an era of agents. We want these agents to be very strong at generating good automation. So we need a lot of people to take these models, take these agents, and then fine-tune them."