AI aims for its Einstein moment, but DeepMind reminds us why large chatbots alone won’t be enough by 2030

Chatbots are already capable of solving complex problems. The test proposed by Demis Hassabis aims even higher: to discover a new scientific idea without having read it. This threshold explains why Google DeepMind, OpenAI, Anthropic, and Yann LeCun are looking beyond just large language models.

The timeline for AGI is approaching, but 2030 remains a technical horizon

In an episode of Y Combinator published in late April 2026, Demis Hassabis, co-founder and CEO of Google DeepMind, places general artificial intelligence close to 2030. AGI refers to an AI capable of learning, reasoning, and acting across multiple domains without narrow specialization.

This timeline does not imply that current chatbots will merely grow larger. Hassabis distinguishes between the already useful foundations, such as massive training and human feedback, and the still-open challenges. For the industry, the real challenge seems less a race for size than a question of architecture.

Large language models have three weaknesses that limit their autonomy

Large language models, or LLMs, predict the likely continuation of a text. This method produces useful responses, but does not always provide a stable representation of the world. Like a GPS without up-to-date maps, the system can move quickly but then lose its reference point.

Hassabis identifies three obstacles: continuous learning, long reasoning, and memory. Continuous learning means integrating new information without retraining the entire model. Structured memory involves retaining useful, verified, and reusable information rather than accumulating a massive digital attic.

Reasoning also remains inconsistent. A model might successfully complete a competitive task, only to fail on a shorter variant. In chess, this flaw is evident when the AI identifies a poor move, searches for an alternative, and then returns to the same choice.

The Einstein test measures a capability rarer than simply finding the right answer

The test mentioned by Hassabis involves limiting the available knowledge before a scientific breakthrough and asking the system to identify the next idea. With a corpus cut off before 1905, the AI would need to reconstruct the theory of special relativity, rather than just reciting phrases about Einstein.

This challenge separates imitation from invention. Special relativity addressed tensions between observations and existing theories. For an LLM, the text alone often lacks a physical grounding: it may know that a ball falls, but it doesn’t necessarily simulate gravity as a phenomenon.

AlphaFold and world models outline a more modular AI

The 2024 Nobel Prize in Chemistry awarded to Demis Hassabis, John Jumper, and David Baker provides a useful example. AlphaFold is not a generalist chatbot. The system predicts protein structures with a specialized goal, and then AlphaFold 3 extends this work to molecular interactions.

Yann LeCun, former Chief Scientist of AI at Meta and co-founder of AMI in Paris, advocates a similar approach with world models. These models of the world seek to represent objects, actions, and their consequences, acting more like a mental simulator than merely a statistical parrot.

The emerging idea combines several components: LLMs for dialogue, specialized models for calculations, ordered memory for retention, and reinforcement learning for decision-making. In this framework, the AI of tomorrow resembles less a single brain and more a workshop with various tools on the workbench.

Scroll to Top