Microsoft Research Finds that GPT-4 Shows ‘Sparks’ of Common Sense, Human-Like Reasoning

• OpenAI’s more powerful version of ChatGPT?
• GPT-4 was able to solve complex math problems, such as finding the area of a triangle?
• GPT-4 was able to identify objects in images, such as a cat or a dog?
• GPT-4 was able to diagnose diseases, such as pneumonia or cancer?
• GPT-4 was able to provide legal advice, such as drafting a contract or arguing a case in court?
• GPT-4 was able to offer psychological counseling, such as helping someone with anxiety or depression?

According to a recent study conducted by Microsoft, OpenAI's enhanced iteration of ChatGPT, known as GPT-4, exhibits the capacity to engage in human-like reasoning and employ common sense. The study highlights the potential of training GPT-4 to emulate the cognitive abilities inherent in human intelligence.

The analysis revealed that GPT-4 represents a significant milestone in the advancement towards artificial general intelligence (AGI), showcasing its ability to engage in reasoning, planning, and experiential learning on par with, or potentially surpassing, human capabilities. This underscores the remarkable progress made by GPT-4 in achieving a level of cognitive proficiency comparable to that of human intelligence.

The artificial intelligence (AI) in question belongs to a novel group of expansive language models (LLMs), which includes ChatGPT and Google's PaLM. These LLMs possess the capability to undergo extensive training with vast volumes of data, encompassing both textual information and visual content. Through this comprehensive input, they demonstrate the ability to generate insightful responses and solutions.

Prior to its public launch, OpenAI received substantial investment from Microsoft, amounting to billions of dollars, granting the company privileged access. Microsoft has recently released a comprehensive 155-page analysis titled "Sparks of Artificial General Intelligence: Early Experiments with GPT-4," delving into the preliminary trials and observations surrounding this advanced language model.

GPT-4 is also used to power Microsoft’s Bing Chat feature.

The research team made a groundbreaking revelation by uncovering the profound capabilities of large language models (LLMs) in emulating human reasoning and leveraging common sense. Through their extensive experiments, they demonstrated the remarkable prowess of GPT-4, a leading LLM, in successfully tackling intricate tasks across various domains without the need for explicit guidance. GPT-4 exhibited exceptional problem-solving abilities in mathematics, vision, medicine, law, and psychology, signifying a significant advancement in the potential of LLMs to simulate human-like cognitive processes.

Microsoft clarified that the publicly available version of the system is not on par with the robustness and capabilities exhibited in the version used during their testing phase.

The research paper provided numerous instances where the AI demonstrated a comprehension of various concepts, including the definition of a unicorn. Notably, GPT-4 successfully generated a representation of a unicorn using the TiKZ sub-programming language. Despite the simplistic nature of the visual renderings, GPT-4 exhibited an accurate understanding of the fundamental essence of a unicorn.

According to OpenAI, GPT-4 showcased a higher level of common sense compared to previous models such as ChatGPT. In a particular task, both GPT-4 and ChatGPT were instructed to stack various items, including a book, nine eggs, a laptop, a bottle, and a nail. GPT-4 exhibited a greater ability to understand the physical limitations and practical constraints involved in the task, resulting in a more sensible arrangement of the items.

While ChatGPT provided a recommendation to place the eggs on top of the nail, the advanced model displayed a higher level of sophistication by arranging the items in a manner that would prevent the eggs from breaking. This demonstrates the improved understanding and practical reasoning capabilities of the more sophisticated model in handling complex tasks.

Original Article publish in DTE