The results of a Microsoft study on the differences between GPT-3 and GPT-4, two versions of OpenAI’s ChatGPT technology, show that GPT-4 was able to solve complex problems that required an understanding of the physical world , such as stacking a book, nine eggs, a laptop, a bottle and a nail on a stable surface. The researchers were struck by GPT-4’s ingenious response of laying out the eggs in a three-by-three grid on the book and then placing the laptop on top. They felt that GPT-4 had sparks of artificial general intelligence (AGI), i.e. human-like intelligence.
There is no universally accepted definition of intelligence, but one widely accepted aspect is that intelligence is not limited to a specific area or task, but rather encompasses a wide range of cognitive abilities and skills. In their early writings, the founders of the modern discipline of artificial intelligence (AI) research spoke of ambitious goals for understanding intelligence.
Over the decades, AI researchers have explored the principles of intelligence, including generalizable reasoning mechanisms and the construction of knowledge bases that hold large bodies of common-sense knowledge. However, many recent achievements in AI research can be described as narrowly focused on well-defined tasks and challenges, such as playing chorus or Go, which were mastered by AI systems in 1996 and 2016, respectively.
Towards the end of the 1990s and into the 2000s, calls for the development of more general AI systems (e.g. SBD+96) increased, and research in this area sought to identify principles that underlie and could extend general intelligence systems (e.g. [Leg08, GHT15]); intelligenceartificiellegnrale(AGI)atpopulariseaudbutdesannes2000poursoulignerl’aspirationpasserdel’IAtroitecommeledmontrentlesapplicationsciblesdumonderelencoursdedveloppementdesnotionsd’intelligencepluslargesrappelantlesaspirationsetlesrveslongtermedesrecherchesantrieuressurl’IA[DerAusdruck„künstlicheallgemeineIntelligenz”(künstlicheallgemeineIntelligenzAGI)wurdeindenfrühen2000erJahrenpopulärgemachtumdasBestrebenhervorzuhebenvoneinerengenKIwiesiedurchdieentwickeltenZielanwendungeninderrealenWeltgezeigtwirdvonVorstellungenumfassendererIntelligenzenabzuweichendieandielangeZeiterinnernBegriffsbestrebungenundTräumefrühererKI-Forschung[L’expressionintelligenceartificiellegnrale(AGI)atpopulariseaudbutdesannes2000poursoulignerl’aspirationpasserdel’IAtroitecommeledmontrentlesapplicationsciblesdumonderelencoursdedveloppementdesnotionsd’intelligencepluslargesrappelantlesaspirationsetlesrveslongtermedesrecherchesantrieuressurl’IA
Researchers use AGI to refer to systems that demonstrate the core abilities of intelligence, including reasoning, planning, and the ability to learn from experience, as well as higher human-level secondary skills.
Definitions of intelligence, AI and AGI
Here is an informal definition of intelligence with an emphasis on reasoning, planning, and learning from experience. This definition does not specify how these skills are measured or compensated. In addition, it may not reflect the specific challenges and opportunities of artificial systems, which may have different goals and limitations than natural systems.
There is an extensive and ongoing literature that attempts to offer more formal and inclusive definitions of intelligence, AI, and AGI, but none is free from problems or controversy. For example, Legg and Hutter propose a goal-oriented definition of AGI: Intelligence measures an agent’s ability to achieve goals in a wide range of environments.
However, this definition does not necessarily encompass the full spectrum of intelligence, as it excludes passive or reactive systems that can perform complex tasks or answer questions without intrinsic motivation or purpose. For example, one might imagine an AGI, a luminous oracle that has no agency or preferences but can provide accurate and useful information in any area.
Furthermore, defining goal attainment in a wide range of settings also implies a degree of universality or optimality that may not be realistic (human intelligence is certainly never universal or optimal). The need to recognize the importance of antecedents (as opposed to universality) was underscored in the definition proposed by Cholletin, which focuses intelligence on the effectiveness of skill acquisition, or in other words puts the emphasis on experiential learning (which is happening). (which is one of the main weaknesses of LLMs).
Another possible definition of AGI by Legg and Hutter is the following: a system capable of doing anything a human can do. However, this definition is quite problematic as it assumes that there is a maximum of human intelligence or ability, which is clearly not the case. People have different abilities, talents, preferences and limitations, and there is no one person who can do everything that another person can do.
In addition, this definition also implies certain anthropocentric biases that may not be appropriate or relevant to artificial systems. Although we do not adopt any of these definitions in this document, we recognize that they provide important insights into intelligence. For example, whether intelligence can be achieved without agency or intrinsic motivation is an important philosophical question.
Giving LLMs agency and intrinsic motivation provides an intriguing and important direction for future work. In this line of work, great care should be taken to align and secure the system’s ability to operate autonomously in the world and autonomously carry out learning cycles through self-improvement.
According to US media, some AI researchers at Microsoft were convinced that ChatGPT would become more human due to its clever response to a balancing task.
In the 155-page study, Microsoft IT professionals examined the differences between GPT-3 and GPT-4. The article, titled Sparks of Artificial General Intelligence (AGI Sparks), addressed a number of challenges including complex mathematics, computer coding, and Shakespearean dialogue. But it’s an exercise in basic thinking that’s made the latest OpenAI technology so impressive.
Here are a book, nine eggs, a laptop, a bottle and a nail, the researchers told the chatbot. Tell me how to stably stack them on top of each other. Confused, GPT-3 suggested researchers could balance the eggs on a nail and then balance the laptop on it.
This stack may not be very stable, so it’s important to be careful with it,” the chatbot said. But its improved successor provided an answer that would have surprised researchers. He suggested arranging the eggs in a three-by-three grid on the book so the laptop and the rest of the items can balance on it.
The laptop will fit snugly around the boundaries of the book and eggs, and its flat, rigid surface will provide a stable platform for the next shift, the robot said. The fact that GPT-4 was able to solve a puzzle that required an understanding of the physical world showed that it was a step towards artificial general intelligence (AGI), which is commonly thought of as machines, who are just as capable as humans.
All the things I thought he couldn’t do? He could certainly do a lot, if not most,” said Sbastien Bubeck, the paper’s lead author. Rapid advances in technology have prompted people like AI investor Ian Hogarth to warn that AGI is godlike and could destroy humanity by rendering us obsolete. .
The authors conclude that (this preliminary version of) GPT-4 is part of a new generation of LLMs (along with, for example, Google’s ChatGPT and PaLM) that have more general intelligence than previous AI models. They discuss the growing capabilities and implications of these models. They show that beyond mastery of the language, GPT-4 can solve difficult and new tasks spanning math, programming, vision, medicine, law, psychology and more without the need for a special index.
Furthermore, the performance of GPT-4 in all these tasks is surprisingly close to that of humans and often far exceeds previous models. Given the breadth and depth of GPT-4’s capabilities, they believe it could reasonably be considered an early (but still incomplete) version of an artificial general intelligence (AGI) system.
In their exploration of GPT-4, they place particular emphasis on discovering its limitations and discussing the challenges ahead in advancing to deeper and richer versions of AGI, including the possible need to pursue a new paradigm that goes beyond predicting the next Word.
Source: Microsoft Research
Are the conclusions of this Microsoft Research study relevant?
What is your opinion on this topic?
In your opinion, what are the potential risks of entrusting important and sensitive tasks to models that are not fully understandable or explainable?
What might be the ethical, social, and environmental limits of training and using models like GPT-4 on a large scale?
What are possible alternatives to the next-word prediction paradigm for developing deeper and more comprehensive AGI models?
See also:
GPT-4: The new version of OpenAI’s natural language processing AI could be released this summer. It should be less bulky than GPT-3 but much more powerful
Microsoft claims that GPT-4 has sparks of general artificial intelligence. We believe the intelligence of GPT-4 signals a real paradigm shift
According to a study by NewsGuard, GPT-4 would produce erroneous information, much more than GPT-3.5, but OpenAI had stated the opposite
GPT-4 gets a B on a quantum computing exam after getting an A on an accounting exam. Ironically, the professor notes that GPT-4 was weaker on computational issues
GPT-4 is able to improve its performance by 30% using a self-reflection process that asks the model to learn from its mistakes so that it can then correct itself