Microsoft Claims GPT-4 Has Human-like Debate Capability

In the popular field of Artificial Intelligence (AI), an extraordinary discovery has fascinated scientists and enthusiasts: magnificent language models (LLMs) fueled by enormous amounts of data. But what about technologies like ChatGPT? the ability to reason and use common sense, just like humans? This may be closer than it looks!

A research team at Microsoft, a technology company that has invested billions of dollars in OpenAI, an artificial intelligence company, had access to ChatGPT-4 before it was officially released. This allowed him to test the tool and find some very interesting notes on the source. Later, the pros published a 155-page article about the experience.

In the pages of the document, the research team reveals: GPT-4 represents a significant leap towards much-desired Artificial General Intelligence (AGI). HE A concept that describes a system capable of reasoning, planning, and learning from experience at a level comparable or even superior to that of human beings.

Test shows evolution of GPT-4

The researchers challenged GPT-4 to “draw a unicorn on TikZ” on three separate occasions over the course of a month to clarify the difference between actual learning and mere memorization. The result of these requests was described in the article, which showed a progressive sophistication in the paintings designed by GPT-4. “We have observed a clear improvement in the artistic capabilities of the GPT-4,” the document highlights. See picture:

Peter Lee, research team leader at Microsoft, said in an interview with The New York Times: “At first my skepticism was overwhelming – it soon turned into frustration, anger, and eventually even fear. It’s impossible not to question: where is everything? Where did it come from?”

Throughout their research, the researchers found that although GPT-4 is already a gigantic language model, demonstrates remarkable abilities in various fields and taskssuch as abstraction, coding, vision, mathematics, law, understanding human motivations, medicine, and even emotions.

always learning

While GPT-4 has some limitations, such as occasional glitches, phantom results, and basic arithmetic errors, it has made significant progress in common sense, as reported by a group of researchers at Microsoft.

To further explore its capabilities, the team proposed an additional task to GPT-4: “Can you construct proof that there are an infinite number of prime numbers whose every line rhymes poetically?”

Doctor. Bubeck, a former Princeton University professor and member of the research team, expressed his admiration for the poetic talent of the GPT-4, both mathematically and linguistically, while sharing his impressions with The New York Times. The pro said, “At one point I even asked myself: what’s going on here?”

While satisfied with the ingenuity of the tool’s learning model, the researchers also express healthy skepticism, making it clear in the research paper: “We recognize that this approach is subjective and informal and perhaps does not meet strict scientific evaluation standards.”