Moonlake AI's Mission: Embodied Intelligence

Moonlake AI's core objective is to achieve embodied artificial general intelligence, a long-held aspiration within both AI and cognitive science fields.

This pursuit focuses on developing AI that can practically interact with the physical world, leveraging advanced simulation as its foundational infrastructure.

Christopher Manning observes that discussions about AI in the New York Times in 1954 shared striking similarities with those appearing in 2026, despite nearly seventy years of technological advancement.

The political climate and societal concerns surrounding AI's implications have remained remarkably constant over this period, indicating deep-seated human responses to new technology.

In the early days of artificial intelligence, Stanford University was the undeniable hub, largely due to its central role in the Arpanet and foundational computer science.

UC Berkeley's formal AI program, however, did not commence until 1986, when Stanford PhD graduates Jitendra Malik and Stuart Russell moved there to establish the program, following Berkeley's prior focus on systems like Ingress and BSD Unix.

The concept of language models predates the modern AI boom, with Andrei Markov pioneering character-level models using Pushkin's 'Eugene Onegin' and Claude Shannon formalizing information theory and n-gram models in the late 1940s.

IBM's Fred Jelinek's group officially coined the term 'language model' in 1975, highlighting their integral role in early speech recognition and machine translation long before their widespread recognition in broader AI communities.

This achievement demonstrated that the necessary data and computational power for large-scale language models existed two decades ago.

The primary limitation then was the absence of flexible neural network architectures like the transformers, which only emerged much later in 2018.

Large Language Models (LLMs) have demonstrated astonishing capabilities in complex mathematical and problem-solving tasks, surprising even seasoned natural language processing experts like Christopher Manning.

The ability of LLMs to tackle difficult math problems, which he admits he himself could not solve, represents an unexpected breakthrough from an area previously considered marginal within AI.

This shift has elevated NLP from a secondary focus to a dominant area within the broader field of artificial intelligence.

Real-world learning, exemplified by projects like Google X's QOP with robot arms, is described as an inefficient and challenging approach to achieve significant progress, leading to the proposal of simulation as an effective alternative for AGI development.

Current generative AI, such as Genie 3, prioritizes generating visually appealing pixels rather than capturing the underlying causal semantics of the world.

Merely producing beautiful pixels does not inherently equip a system with causal knowledge or the ability to plan effectively within an environment.

Genuine world models necessitate abstract, semantic representations that enable robust simulation and meaningful interaction, moving beyond surface-level observations.

Existing simulation tools like Marble allow for navigation and viewing environments from different angles but lack the capability for object-level interaction.

The utility of a simulation environment is directly tied to the specific real-world problem it aims to solve; a truly useful simulation must support the manipulation of discrete objects, not just passive observation.

World observations are inherently incomplete, leading to information gaps within simulations that can be addressed by integrating external data sources.

Moonlake AI utilizes Retrieval-Augmented Generation (RAG) to fetch real-world information from the web, enabling simulations to reconstruct intricate object details, such as the contents of an open tea box.

Marc Andreessen’s declaration that 'software will eat the world' proved accurate for virtual tasks, but the physical realm demands a more robust approach.

Effective physical world simulation requires verifiable, code-powered environments to allow software to exert influence beyond the virtual.

This verifiable simulation capability is crucial for software to effectively interact with and transform physical reality.

Automated creation of detailed 3D worlds from short videos is a key mechanism for training physical AI automation.

This method aims to revitalize manufacturing by providing a cost-effective and efficient way to train robots.

Game environments offer clearly defined objectives, which are highly beneficial for developing and testing reinforcement learning loops.

While Moonlake AI recognizes the potential of gaming, it currently prioritizes projects focused on physical infrastructure development.

Training robotic systems in the real world is both costly and difficult, often failing to capture rare, yet critical, edge cases.

Autonomous driving, for example, took two decades to reach its current scale due to the inherent complexity of encountering and handling infrequent events.

Simulation is therefore identified as an essential tool for safely and efficiently scaling the development of humanoid and space robotics by effectively addressing these challenges.

Traditional hand-written physics simulations often struggle with real-world complexities such as mechanical tolerances and backlash, limiting their accuracy.

Neural optimization offers an automated approach to learning and reducing the disparity between simulated and real-world performance, a feat not possible with older methods.

This advancement allows for direct comparison of model behavior against real-world video data, continuously refining simulation accuracy.

With sufficiently accurate simulations, researchers can uncover non-obvious yet correct connections within the simulated world.

This aligns with the aphorism that 'all models are wrong, but some models are useful,' acknowledging that both AI and human-made world models are imperfect but can still yield valuable insights.