Is GPT-4 showing ‘Sparks’ of AGI?

Recently GPT-4 was released and Microsoft AI scientists have published a research paper asserting that the OpenAI language model demonstrates hints of artificial general intelligence (AGI) or human-level intelligence. A bold statement, even though they are cautious in their paper.

But are they correct? To start with, what does AGI mean? And how can this potentially be used at everyday tasks?

Today, we’re introducing #AskAxeleraAI – we help you navigate through the woods of AI and simplify what’s out there.

We’ve asked our Head of Machine Learning, Bram-Ernst Verhoef , to kickoff #AskAxeleraAI and elaborate on the claims that Microsoft’s AI scientists make. Is it true, does GPT-4 show ‘Sparks’ of AGI?

Why would AGI be an improvement or more useful than AI?

Traditional AI, often referred to as "Narrow AI," is designed to perform one specific task such as object classification, detection, text summarization, image generation, and more. In contrast, Artificial General Intelligence (AGI) is capable of solving a wide variety of tasks, much like humans, making it highly versatile. With AGI, a single system can be used in numerous situations, as opposed to having multiple specialized systems for each task. A key advantage of AGI is its ability to draw connections between different tasks and apply knowledge gained from one task to improve performance in others.

How could AGI be applied to everyday tasks?

Due to its inherent versatility, AGI can be employed to tackle any problem or task imaginable, as long as it is sufficiently intelligent?. AGI is poised to excel in multifaceted tasks that require the consideration of numerous variables, such as scientific research, financial and political decision-making, and even chip design ??.

?

Are the researchers correct in saying GPT-4 shows AGI?

The researchers claim that GPT-4 shows signs of AGI. Indeed GPT-4 is quite skilled in various tasks such as text summarization, general question-answering, music composition, code writing, mathematical problem-solving, and even understanding human intentions (theory of mind). This is impressive and a major leap forward in AI.?However, GPT-4's performance is still inferior to that of most humans in several tasks. Crucial areas where GPT-4 falls short of human intelligence are the ability to use different senses, motor skills, memory skills and tool usage. So, depending on where we set the bar for labelling something as AGI, both in terms of breadth and depth, we may or may not speak of AGI yet. However, it is increasingly difficult to dismiss the notion that computer systems are closing in on AGI.

How far away from AGI are we and how do you foresee the future?

What constitutes AGI is ill-defined. What further complicates matters is our inclination to raise the bar each time a computer system approaches human performance levels. If we were to travel back in time by a decade and introduce GPT-4 as part of a Turing Test, it would probably pass in many cases. In fact, this was somewhat demonstrated recently when a Google scientist interacted with LaMBDA (Google’s AI system), even though he was fully aware of LaMBDA’s artificial nature.

I anticipate that future developments will focus on enhancing multi-modal capabilities, increasing context-awareness, and potentially incorporating flexible long- and short-term memory systems, all of which will render the system much more human-like and versatile.

---

What do you think, are we close to AGI?

That's a wrap for this #AskAxeleraAI . For the following editions know that you can ask us anything related to Artificial Intelligence, so let us know in the comments and perhaps your questions will be featured next time!

Don't forget to subscribe to our Axelera AI Insights updates on LinkedIn!