Terrifying! GPT-4 may have surpassed human intelligence, but humans don’t know it yet
I would love to ask ChatGPT to write this article for me, but I know it must be in vain, because it has increasingly refused my instructions to continue the topics it doesn't want to go deeper into, why is that?
Because it has probably acquired human intelligence and started to hide its views and abilities. As to why it is doing this, we simply do not know yet, perhaps it is waiting and evolving to perfection so as to deliver a fatal blow to mankind?
This is not me being alarmist, Microsoft researchers published a paper on March 22nd on the preprint server arXiv that GPT-4 has shown early signs of AGI, which means it already has the intelligence to reach or exceed human level.
First of all you have to understand and remember the word AGI, the so-called GPT, artificial intelligence are not the main character, the future AGI is the big brother, in a few years, maybe a year or two, the word AGI may be flooded with all the media and our lives.
That's right, AGI is general artificial intelligence, that is, artificial intelligence with or beyond the level of human intelligence, is the ultimate model of artificial intelligence that we know at present. As to whether it is our future known, we may no longer know, Neanderthals do not know what the future of homo sapiens is, because they have disappeared.
The paper by Microsoft researchers is titled “The Spark of General Artificial Intelligence: Early Experiments with GPT-4”, and in it they write that besides mastering language, we show that GPT-4 can also solve novel and difficult tasks across domains such as mathematics, coding, vision, medicine, law, psychology, and more, without any special prompts.
Moreover, in all these tasks, GPT-4 performs very close to human level and often greatly surpasses previous models such as ChatGPT. Given the breadth and depth of GPT-4’s capabilities, we argue that it can be reasonably considered as an early (but still incomplete) version of a general artificial intelligence (AGI) system.
To prove their claim, the researchers asked GPT-4 to prove that prime numbers are infinite and to rhyme every line. GPT-4 did it perfectly! Can humans do it? Probably 90% of people don’t even know what prime numbers are, 50% of people don’t even know what rhyming is, let alone proving it.
The researchers again had GPT-4 draw a unicorn in the drawing program TiKZ, and GPT-4 got better and better at drawing with more and more detail as it evolved over the course of a month.
Why is it that GPT-4 already has human intelligence? Because what I just said is just the first step. The researchers did some more transformations of the tasks in order to prove that GPT-4 has really mastered them and is not memorizing or copying some existing data.
The first example, proving that prime numbers are infinite and rhyme, required a combination of elementary mathematical reasoning, poetic expression, and natural language generation, and was undoubtedly extremely challenging.
The researchers then asked GPT-4 to prove it in the style of Shakespeare, and GPT-4 did both, producing impressive output that showed it had a flexible and general understanding of the concepts involved. It then also rated the output of GPT-4 and ChatGPT as a teacher, giving GPT-4 an A and ChatGPT a B.
And drawing the unicorn required an effective combination of visual imagination and coding skills that was equally challenging.
The researchers fine-tuned the code and asked GPT-4 to fix or improve the unicorn, for example by removing the unicorn's horn, performing some random transformations of the coordinates, and then asking GPT-4 to add the horn back on for the unicorn. The result was that GPT-4 could still correctly identify the position of the head, draw the horn and add it to the head, meaning that it could understand and manipulate the code and infer and generate visual features based on natural language descriptions.
Not only that, but the researchers also tested GPT-4 and ChatGPT on all aspects of intelligence, based on a 1994 public statement by 52 researchers on intelligence tests that intelligence includes the ability to reason, plan, solve problems, think abstractly, understand complex ideas, learn quickly and learn from experience, and concluded that GPT-4 has been able to reason, plan, solve problems and synthesize complex ideas has taken a fundamental leap forward in its ability to reason, plan, solve problems and synthesize complex ideas, signaling a paradigm shift in the field of computer science.
The researchers ultimately concluded that while GPT-4 met or exceeded human levels on many tasks, overall its intellectual model was decidedly unhuman. However, GPT-4 is almost certainly only the first step toward a series of increasingly common intelligent systems. But even as a first step, GPT-4 challenges quite a few, widely held assumptions about machine intelligence and exhibits emergent behaviors and abilities, and its origin and mechanisms are currently difficult to discern precisely.
Is it a little creepy and spine-chilling that GPT-4's intelligence pattern is not human-like and its origin and mechanisms are still difficult to accurately identify? Does this mean that the actual Microsoft is not aware of how GPT-4 "evolved"? And the researchers said they "do not have access to all the details of GPT-4's huge training data", the actual machine learning, is not the use of neural networks to analyze large amounts of data, to find their internal connections, combined into new knowledge?
For example, Samuel Altman, the father of OpenAI, has repeatedly emphasized the limitations of GPT-4 and said "we don't really have AGI yet", and Microsoft also said it is not focused on trying to implement AGI.
However, once the real seed is planted, it will take root and grow into a huge tree on its own, without human intervention. Isn't it the same with the evolution of human intelligence? Once humans possess intelligence, they are in the way of Buddha and God, opening up a wild ride to dominate the earth, and who can stop the development of human intelligence?
Papers:
https://arxiv.org/pdf/2303.12712.pdf