This last module looks ahead at the developments under way and to come. All of them are uncertain: the exercise is to map the directions, not to predict.
7.1Beyond LLMs: world models and physical intelligence
Part of the research community bets that the future lies not in ever-larger LLMs, but in models that understand the physical world. This is the bet of AMI Labs (Yann LeCun, JEPA) and World Labs (Fei-Fei Li, spatial intelligence). The goal: enabling machines to plan and act in the real world, not just to manipulate language.
7.2Agentic AI: the year of agents
The big shift of 2025-2026 is the move from AI that answers to AI that acts. Agents chain actions autonomously: browsing the web, operating software ("computer use"), carrying out long multi-step tasks, collaborating with one another (multi-agent systems).
Interoperability protocols such as MCP (Model Context Protocol) standardise how models connect to external tools. People speak of models that "stop being chatbots and become collaborators".
7.3Humanoid robotics and embodied AI
AI is leaving the screen for the physical world. The breakthrough comes from Vision-Language-Action (VLA) models, which learn from videos the way LLMs learned from text. The state of play in 2026:
- Figure (United States): its Helix system replaced tens of thousands of lines of code with a neural network; its BotQ factory reaches ~one robot per hour (June 2026).
- Tesla: initial production of Optimus Gen 3 targeted for summer 2026, with a price goal below $20,000 at scale.
- China: the Unitree G1 (~$16,000) is slashing prices; AgiBot and others are ramping up.
- Boston Dynamics: deployments of the electric Atlas.
- NVIDIA is positioning itself as the "platform" provider (Isaac Sim, the GR00T model).
The context is massive: venture capital funding in robotics tripled between 2023 and 2025 to reach ~$40.7bn a year; China devotes a $138bn public fund to it. One central challenge remains: data (initiatives such as Open X-Embodiment pool more than a million trajectories across 22 types of robots).
7.4Towards AGI? The big debate
The prospect of artificial general intelligence (AGI), or even superintelligence, is deeply divisive. Optimists point to recursive self-improvement (models training models) and progress "every night". Sceptics point to the limits of LLMs (fragile reasoning, no grounding in reality) and doubt that simply scaling up will be enough. Between these poles, most researchers call for caution.
7.5Geopolitics and regulation
AI has become a sovereignty issue. The competition is between the United States, China and Europe. On the regulatory side, the European AI Act (the world's first binding framework) is gradually coming into force, while international summits try to coordinate governance. France, which hosted an AI Action Summit, seeks to position itself as a balancing force between innovation and regulation.
7.6AI in the service of science
One of the most tangible promises: accelerating scientific discovery. AlphaFold (biology), materials prediction models, help with mathematical proofs: so many fields where AI already acts as a powerful research accelerator, as the double 2024 Nobel Prize confirmed.
Key takeaway
Three transitions shape the near future: from language to the physical world (world models, robotics), from answering to acting (agents), and from tool to geopolitical infrastructure. None of them is a foregone conclusion, and all call for as much enthusiasm as vigilance.
Sources
Humanoid Robots in 2026 (medium.com); Figure 03 vs Tesla Optimus Tracker (newmarketpitch.com); The Companies Building Embodied AI in 2026 (compareais.net).
Finished this module?
Your progress is saved on this device.