Is GPT-5 a "pocket doctor"? Nobel laureate Demis Hassabis slams Sam Altman: The idea of a doctor-level AI is pure nonsense.

The so - called doctoral - level artificial intelligence at present is sheer nonsense!

Surprisingly, Demis Hassabis, a Nobel laureate and the CEO of Google DeepMind, publicly lashed out at Altman.

In a recent interview, Hassabis openly stated that referring to today's large language models (LLMs) as "doctoral - level intelligence" is pure nonsense!

They are not truly of doctoral - level intelligence. Although they possess some abilities at the doctoral level, they lack comprehensiveness overall.

True general intelligence should have comprehensive abilities at the doctoral level in all fields.

True general artificial intelligence would not make elementary mistakes. Current AIs do not have the ability to conduct continuous reasoning, adapt, and learn.

Hassabis believes that there are probably still one or two key breakthroughs missing, and it will still take 5 to 10 years to achieve true "doctoral - level intelligence".

Hassabis's criticism of "doctoral - level AI" and his insightful views on the essential capabilities of AGI are quite well - received:

However, his prediction about the arrival time of AGI may not be accurate.

Besides discussing the path to AGI, at the All - In Summit, Hassabis first recalled his Nobel Prize moment and then systematically elaborated on his latest judgments regarding world models, robots, scientific research acceleration, energy consumption, and efficiency:

Genie 3 can transform a piece of text into a real - time interactive "world". Gemini is becoming Alphabet's "AI engine". However, a truly creative and consistent AGI still requires key breakthroughs and time to mature.

An AI Genius at the Helm of DeepMind

AlphaFold Helps Win the Nobel Prize

Hassabis became a chess prodigy at the age of 4. In 2023, he was knighted by the British royal family for his contributions to AI, and in 2024, he won the Nobel Prize in Chemistry.

Hassabis and John M. Jumper of Google DeepMind shared half of the Nobel Prize for protein structure prediction.

However, he was notified of the award only ten minutes before the official announcement, and he didn't have time to process the news. He was completely stunned.

Subsequently, his week - long participation in the award ceremony in Sweden was truly amazing. Every arrangement, including interactions with royal family members, left him in awe.

In this 120 - year - old honor tradition, the most shocking part was a special arrangement by the organizing committee - they took out the historical signature book of the Nobel Prize from the vault.

Hassabis experienced a once - in - a - lifetime highlight moment:

He signed his name in the same register as all the previous Nobel laureates in history, including Marie Curie and Albert Einstein.

As the CEO of DeepMind, Hassabis is at the helm of Google's AI.

To develop AI, Google and the AI teams under Alphabet (including the original DeepMind) were integrated to form the current Google DeepMind.

Hassabis described the new DeepMind as the "engine" of Google and Alphabet.

DeepMind is responsible for the development of generative AI models such as Gemini, Gemma, and Veo, and also conducts scientific research projects represented by AlphaFold.

Gemini is Google's core AI model, which is applied to multiple products such as Google Search and Gmail.

The total number of people he leads is about 5,000, and more than 80% of them are engineers or doctoral researchers.

When Google started developing Gemini, it adhered to the multi - modal approach, which means it can recognize images, listen to audio, watch videos, and output in various forms.

To achieve general artificial intelligence, the system should not only understand language and abstraction but also the physical world around us. This is why robots are difficult to develop and intelligent glasses - type assistants are crucial.

He introduced the newly launched world model Genie 3, Google's "new Android" Gemini Robotics, and the popular "NanoBanana".

The first two are aimed at a common goal: enabling AI to truly understand and manipulate the physical world.

DeepMind is working on making Gemini Robotics a "quasi - operating system layer" for cross - robot platforms, which can be understood as the "Android" for robots.

Hassabis believes that the development of robots is still in its early stage, but there will probably be an "Aha moment" in the next one or two years.

In the next few years, general models will be stronger, more robust, and more aware of the details of the physical world, which will be sufficient to support robots' manipulation ability in the physical world.

Regarding the future development of creative work, Hassabis said that top - notch creative people will still dominate engaging experiences and dynamic storylines. They may become "editors of worldviews", responsible for guiding and integrating the collective creativity of others.

Where Is the Path to AGI?

The scientific application of AI is the direction that Hassabis cares about the most.

He has staked his entire career on AI in order to use it to accelerate scientific discovery and improve human health.

If AGI is built in the right way, it will become the ultimate scientific tool.

In the past few years, DeepMind has demonstrated many paths. The most famous one is AlphaFold, but Google has also applied AI to material design, plasma control in controlled fusion devices, weather forecasting, and even mathematical problems at the Olympiad level.

An AI system with the same paradigm, with a little task - specific fine - tuning, can work in many complex fields.

Hassabis believes that the acceleration of scientific discovery by AI has just begun.

Of course, there is still a missing piece: true "creativity".

Given a proposition, today's AI can prove and solve problems, but it can't put forward new conjectures, hypotheses, or theories on its own. When it can independently pose good questions, that may be a key milestone test.

What is "creativity"?

Hassabis believes that it is the "intuitive leap" that we often applaud - the kind of leap that top - notch scientists and artists in history have made.

Perhaps creativity relies on analogy, on connecting seemingly unrelated things.

Psychology and neuroscience have their own theories on how humans achieve this, but an operational test is:

Truncate a modern AI's knowledge as of 1901 and see if it can "come up with" a theory like the special theory of relativity in 1905.

If it can, it means that humans have touched on something real, and perhaps AGI is just around the corner.

Another example: Ten years ago, AlphaGo not only defeated the world champion in Go but also made a "divine move" - the famous "37th move" in the second game.

But the question is: Can AI not only invent new strategies but also "create a game as elegant, playable, and aesthetically appealing as Go"?

The answer is currently no. This is exactly the shortcoming of "generality". A true AGI should also be able to achieve this level of creation.

So, what is specifically missing?

Dario of Anthropic and Altman of OpenAI believe that AGI will arrive soon.

Hassabis is more cautious. He believes that the core lies in whether we can reproduce the "intuitive leap" of the world's best scientists, rather than just making incremental improvements.

The difference between great scientists and good scientists lies not in basic skills but in creativity. They can capture patterns from other disciplines and apply them to current problems by analogy.

Hassabis believes that AI will eventually achieve this, but currently, in terms of reasoning thinking, AI still needs improvement to support such a breakthrough.

Another shortcoming is "consistency".

Altman and others claim that current AIs have reached "doctoral - level intelligence", but Hassabis doesn't think so.

In some subtasks, they have reached the "doctoral level", but it doesn't mean they are "fully doctoral - level".

And "general intelligence" means being able to reach that level stably in all dimensions. In fact, we've all seen that:

Just by changing the way of asking questions, current chatbots can make elementary mistakes in high - school math or even simple counting.

For a true AGI, this should not happen. Hassabis believes that it will still take about 5 to 10 years to achieve an AGI with the above - mentioned capabilities.

In addition, AI also lacks the ability of "continuous learning": being able to absorb new knowledge online and adjust behavior in a timely manner.

Perhaps the Scaling Law will continue to bring some improvements.

But if we have to make a bet, Hassabis believes that one or two key original breakthroughs are still needed, and these breakthroughs are likely to occur in the next five years.

AI4S Continues to Make Efforts in Solving Scientific Research Problems

Besides AlphaFold, which has achieved numerous significant results and won the Nobel Prize, AI will also help improve energy efficiency and solve the derivative problems caused by its huge energy consumption.

Hybrid models like AlphaFold are the future development direction of AI

AlphaFold is a hybrid model.

A hybrid model refers to the simultaneous use of probabilistic models and deterministic models.

Probabilistic models, which are commonly used in current large models, predict the next token based