【New Intelligence Report】
The AGI five-year countdown has begun! Hassabis predicts that perhaps just one or two AlphaGo-level breakthroughs could lead to the arrival of AGI within 5 years, and its speed and impact will be 10 times that of the Industrial Revolution.
Humanity is only 1-2 key technological breakthroughs away from AGI!
Just now, Nobel laureate and Google DeepMind CEO Demis Hassabis unveiled the ultimate timeline for AGI!
He believes that within 5 years, perhaps still requiring 1-2 major technological breakthroughs, we might cross the barrier to AGI.
While making optimistic predictions, Hassabis also didn't forget to pour some cold water on us:
He believes that simply scaling up data and computing power may not be enough to achieve AGI.
For example, Hassabis thinks that while large models are powerful, they lack a true understanding of the physical world, logical reasoning, and long-term planning.
Therefore, to achieve AGI, the missing piece for large models is the "World Model".
In addition, Hassabis also believes that AI will be the ultimate tool for scientific discovery.
AlphaFold is just the beginning. AI will usher in a golden age of scientific discovery within the next decade, especially in fields such as drug development, disease cure, new material discovery, and clean energy (fusion).
This will undoubtedly accelerate the arrival of AGI.
If Hassabis's prophecy comes true, this will be a momentous transformation, with its speed and impact being 10 times that of the Industrial Revolution.
And each of us will be caught in the shock of this epic transformation.
Why aren't large models AGI yet?
Taking the large models we are most familiar with, such as ChatGPT and Gemini, as examples.
Perhaps you feel that although they sometimes perform outstandingly on some difficult tasks, they also often make mistakes on some simple questions.
Hassabis used a very precise and vivid term to describe this state: "Jagged Intelligence."
This is like those students in the class who are extremely partial to certain subjects.
They may be geniuses in liberal arts and programming, but in physical common sense, logical reasoning, and long-term planning, they may not even reach the level of ordinary students.
Why is this?
Hassabis pointed out the essential limitations of large language models (LLMs) incisively: they are just top-tier "probability prediction machines."
They do not truly "understand" this world; they are only predicting the probability of the next word appearing, thus lacking awareness of the physical laws of the real world and not possessing a coherent, self-correcting thinking model like humans.
Therefore, they are extremely good at some things, but completely incapable in others.
This is like asking someone who only memorizes chess moves but doesn't understand the rules of Go to play. The first few moves might look decent, but once the situation becomes complex and requires thinking about strategies across dozens of moves, he will immediately collapse.
Therefore, to evolve from the current "partial student" to an omnipotent AGI, simply making the model bigger (Scaling) is no longer enough.
We need a qualitative leap to fill in the key piece of the puzzle leading to AGI.
The Key Piece of the Puzzle to AGI
Hassabis specifically pointed out the direction of these one or two key technological breakthroughs.
Key Breakthrough 1
"World Model"
If large models are "reading ten thousand books," then the "World Model" is like "traveling ten thousand miles."
The so-called World Model refers to a model that can predict and simulate changes in environmental states with actions. Its core logic is to truly "understand" the operating rules of the physical world.
For current large models, if you ask it "what happens if a cup falls off the table," it tells you "it might break" based on text probabilities.
But an AI with a World Model truly simulates gravity, friction, and the fragility of glass in its "mind," and it "sees" the process of the cup falling.
Currently, DeepMind is developing video/interaction models like Genie and Veo as the embryonic form for building a World Model.
This is also the prerequisite for AI to move from the "digital world" to the "physical world."
Only by understanding the physical laws can AI drive robots to serve tea and water, tighten screws, and handle complex causal relationships in reality, rather than just being able to chat with people.
Key Breakthrough 2
"Agentic Systems"
Having the ability to understand the world is not enough; AI also needs to be able to "act" in the world.
This is the second breakthrough: Agentic Systems.
Current AI is passive: you ask a question, it answers a question.
Future Agentic AI will be proactive.
You give it a vague goal, such as "help me plan and book a trip to a certain place."
It can break it down into dozens of steps: checking airfare, comparing prices, booking hotels, planning routes, adjusting the itinerary based on the weather...
More importantly, it has the ability to "cognitive correction."
If it finds during execution that airfare prices have increased or the hotel is fully booked, it can stop like a human, rethink, and adjust the plan, instead of directly reporting an error or getting stuck in a loop.
Hassabis also specifically mentioned DeepMind's "secret weapon": AlphaGo.
The reason why AlphaGo could defeat human champions back then was because it possessed this "planning" ability; it could deduce changes in the game dozens of moves into the future.
The current goal is to generalize this "planning" ability on the chessboard to specific scenarios in the real world.
When the extensive knowledge of large models meets the physical cognition of World Models, plus the action capabilities of Agentic Systems, it may fill the key piece of the puzzle leading to AGI, ushering in the moment of AGI's arrival.
A Future 10 Times Faster than the Industrial Revolution
Hassabis is so persistent about AGI not to create a more chatty Siri or to make ad recommendations more precise.
His ambition is written in DeepMind's core mission, which has never changed:
AI for Science (Using AI to advance science).
In an official blog post written by Hassabis and others, it was stated that DeepMind will establish its first automated laboratory in the UK in 2026, focusing on materials science research.
The laboratory will be built from scratch, fully integrating the Gemini system, and will command world-class robots to synthesize and characterize hundreds of materials daily, greatly shortening the time required to discover transformative new materials.
Imagine such a scenario:
AI is responsible for reading massive amounts of papers and proposing new scientific hypotheses;
Agentic Systems are responsible for designing experimental plans;
Robots connected to World Models are responsible for operating precision experimental instruments;
Finally, AI analyzes the experimental results, self-iterates, and starts the next round of experiments.
AI's involvement in scientific research is expected to reduce costs and spawn new technologies, with the efficiency of scientific research being improved by hundreds or even thousands of times.
Perhaps in the near future, superconductors that work at room temperature and pressure can enable low-cost medical imaging and reduce power loss in the grid.
Other new materials can help us address key energy challenges by promoting the development of advanced batteries, next-generation solar cells, and more efficient computer chips.
Therefore, Hassabis said that the scale of this transformation will be "10 times that of the Industrial Revolution," while the speed will be "10 times that of the Industrial Revolution."
The Industrial Revolution took more than 100 years to reshape human civilization, while AGI may only need 10 years.
This will be an era of great abundance, but also an era of extreme turbulence.
Old jobs will disappear, old economic structures will collapse, but the boundaries of human cognition will be infinitely extended.
China's AI Models Only Lag Behind the US by "A Few Months"
In this ultimate race to the future, where will China stand?
When Hassabis was interviewed by CNBC, he said that the gap between China's AI models and the capabilities of the US and the West may have narrowed to "only a few months":
China's AI models may be much closer than we imagined a year or two ago. So far, they may only be lagging by a few months.
The sudden emergence of DeepSeek and the strong performance of Alibaba's Qwen model have proven the amazing engineering capabilities of Chinese technology companies.
Chinese AI companies have trained powerful models using relatively backward chips and at lower costs.
Nevertheless, Hassabis believes that although China has proven its ability to catch up, it remains to be seen whether it can achieve a real AI breakthrough.
He thus raised a deeper question, which may be an objective review and a "wake-up call":
Hassabis compared DeepMind to a "modern version of Bell Labs," a holy land that gave birth to source innovations such as transistors and information theory.
He believes that China has currently proven itself to be a world-class "engineer" capable of quickly replicating and optimizing cutting-edge technologies (Copy and Improve).
However, the real test lies in whether it can be the "inventor":
The key question is whether they can achieve original innovation beyond the cutting edge? Can they truly create something new, such as a new Transformer, to achieve a breakthrough beyond the cutting edge?
This is DeepMind's moat, and it will also be the next key battleground in the US-China AI competition.
Regardless, the judgment of this global AI leader is already very clear:
The AGI countdown has begun, with only one or two key technological breakthroughs left.
Within five years, we are likely to witness the historic moment of AGI's arrival.
References:
https://x.com/Ric_RTP/status/2012523232998334577?s=20%20
https://www.cnbc.com/amp/2026/01/16/google-deepmind-china-ai-demis-hassabis.html