What capabilities are necessary to realize AGI? What is intelligence in the first place?



AI technology is advancing rapidly, but artificial general intelligence (AGI) that can do anything has not yet appeared.

François Cholet , a former Google researcher who has led AI research, explained the requirements for AGI and the shortcomings of existing AI at the ' AI Startup School ' held in June 2025.

François Chollet: How We Get To AGI - YouTube


Cholet is a Google researcher who developed the deep learning library ' Keras ' and was employed at Google until November 2024. At the time of writing, he is conducting research at ' Ndea ', an AI research company he founded, and is developing an AI benchmark called 'ARC-AGI' to promote the development of AGI.

The ARC Prize is a contest that will award a total of $1 million to researchers who develop general artificial intelligence with human-level intelligence.



Many AI benchmarks focus on 'what knowledge the AI has.' The graph below shows the number of years since each benchmark was introduced on the horizontal axis and the score recorded by the AI on the vertical axis. Many benchmarks stand in the way of AI as 'difficult tests to solve' immediately after their introduction, but as AI performance improves, the scores rise and within a few years the AI exceeds the 'human score.' However, just because an AI exceeds a human score on these benchmarks does not mean that the AI has acquired the same level of intelligence as humans.



Cholet defines intelligence as 'the ability to deal with problems and new situations that you encounter for the first time,' and calls this ability 'fluid intelligence.' The benchmark test '

ARC-AGI-1 ' released by Cholet in 2019 is designed to measure AI's 'ability to combine existing knowledge to solve problems,' and in the case of OpenAI's GPT series, even GPT 4.5 , which was released in 2025, can only record low scores.



Since the introduction of the o series of inference models, the scores on the ARC-AGI-1 have increased rapidly, and it has become possible to record scores of over 80%. However, the ARC-AGI-1 only measures whether or not a person has fluid intelligence, and does not measure the degree of ability. For this reason, even if a person scores perfectly on the ARC-AGI-1, it cannot be said that AGI has been achieved.



The diagram below shows a simple example of AI's 'ability to adapt to new situations based on existing knowledge.' Low intelligence systems have a low ability to associate knowledge with each other, but high intelligence systems have improved association capabilities and can adapt to more situations.



ARC-AGI-2 , which was released in March 2025, is designed to measure levels of fluid intelligence.



In addition, a development preview version of ' ARC-AGI-3 ', which consists of highly difficult problems, was released on July 19, 2025, Japan time. However, according to Mr. Cholet, even if the ARC-AGI-3 problems could be perfectly solved, it cannot be said that AGI has been realized. Mr. Cholet says that we will approach AGI through future versions such as 'ARC-AGI-4'.




In addition, Cholet is also developing AI at Ndea, an AI company he founded. Ndea's research focuses on 'program synthesis,' which creates new problem-solving programs based on existing knowledge, and at the time of writing, they are aiming to create an AI that can solve ARC-AGI problems without knowing anything about ARC-AGI.

Ndea
https://ndea.com/



Related Posts:

in AI,   Video,   Software, Posted by log1o_hf