What is AGI, and what does it mean for the future?

Step into a future shaped by the power of AI, and you’ll find yourself in a world populated with the darkest sci-fi nightmares‌ – ‌think Terminator, HAL, and a host of others. It’s a future that often evokes a deep-seated fear.

This apprehension is not unfounded. Legitimate concerns about the trustworthiness of AI, the threat of job displacement, the potential for the spread of fake information, and the fear that we might lose control is something we all need to be aware of and take seriously.

We believe a better future is possible if we make the right choices, especially as AI grows more powerful. We can already see glimpses of this extraordinary potential in artificial general intelligence (AGI), which will be like nothing that came before it.

Artificial general intelligence (AGI) is a hypothetical type of artificial intelligence (AI) system that possesses almost human-level intelligence – the ability to learn, converse, plan, and perform tasks across a range of domains, from the scientific to the creative to the philosophical to the personal – combined with the speed, memory, reliability, and scalability of high-performance computing.

Unlike today’s narrow AI, which handles specific tasks but can’t step outside the bounds of its training data, AGI can apply knowledge across fields. Think of the difference between AI systems that can be trained to play chess at grandmaster level versus a machine that can learn chess, write symphonies, and talk philosophy with minimal guidance, all while learning autonomously.

While we haven’t reached AGI yet, today’s large language models (LLMs) provide a glimpse of what’s possible, with the ability to carry on open-ended conversations with natural language, coupled with their access to vast stores of knowledge in just about every domain. Already, they are beating human experts on a wide range of tasks across various disciplines.

In the best-case scenario, with the appropriate guardrails, AGI promises to turn everything we value about intelligence (the ability to understand context, to solve novel problems, and to plan for the future) into intelligence on demand, a resource anyone can use at any time, whether that be for business operations or as personal assistants in daily life.

The current AI systems we rely on today are known as narrow AI, meaning they are narrow in scope and only excel at one specific task. For instance, you might see an AI model that performs a task like image recognition, recommending products or, as we touched on earlier, beating the world’s best players at chess.

These systems are exceptional at what they do, but that’s where their scope ends. You wouldn’t ask the world’s strongest chess engine to assist with doctor diagnoses in a hospital, for instance. This is because narrow AI lacks general intelligence. It can’t reason like we can, meaning it can’t learn new skills unless we train it.

This is the fundamental difference that defines AGI. It has that ability to learn outside of its scope, apply general reasoning, and transfer knowledge to every area – even those it has no familiarity with.

We’ve talked about AGI in theoretical terms, but let’s translate that into real-world scenarios. Here are some ways AGI might be used in the future:

• Acting as an everyday copilot, providing guidance and assistance in various tasks from shopping to party planning.
• Offering expert-level understanding of diet, fitness, medication, and disease mechanisms, leading to effective diagnosis and therapies.
• Assisting healthcare workers by handling administrative and logistical tasks, reducing their workload and minimising medical errors.
• Entertaining by providing recommendations for movies, music, and books.
• Serving as a career collaborator, mentor, and problem-solver, and even contributing to the development of new products and industries.

Despite the excitement, AGI doesn’t actually exist yet. Even the most advanced models remain a significant distance from the possibilities of the future. Human learning sets a high bar. LLMs may appear to know a little bit about everything, but they’re still limited to the contents of their training data and they require costly training sessions to maintain and update them.

AGI must be able to learn autonomously and know when and how to seek out information of its own accord. LLMs simply aren’t capable of performing in this capacity. Instead, they require human developers to bombard them with info like a student at the mercy of a teacher. AGI will be more like an ambitious adult, embracing curiosity, anticipating problems and reacting to mistakes with a continual desire to improve.

Building AI systems capable of this kind of abstract thought and autonomous learning is, as you can imagine, no small task. The sheer technological challenge of creating a machine that can rationalise, adapt and learn without task-specific programming still requires breakthroughs in areas like machine learning and cognitive science. The need to navigate ethics and ensure safety poses another challenge; we need to be able to hit the big red abort button at any time.

Terms like ‘strong AI’ and ‘full AI’ are often used interchangeably when picking apart the definition of AGI. Strong AI, in particular, implies something more than general intelligence – sentience rather than competence. To reach that benchmark, AGI will need to engage with the world in a way that mirrors a human’s experiential understanding: the way we see, smell, and hear, and how we link these senses together to form our worldview. This is a further chasm that AGI will need to cross to achieve the cognitive capabilities we envision.

AGI would have been almost inconceivable at the turn of the century. Now it no longer feels impossible. Artificial general intelligence research is already underway thanks to emerging technologies. Here are some of the concepts driving the innovation.

• Deep learning - focuses on neural networks to recognise patterns and make decisions. Deep learning is the building block of modern AI solutions and powers everything from LLMs to computer vision models.
• Generative AI - lets AI systems like foundation models produce one-of-a-kind, relevant content based on the massive datasets they’re trained with. Gen AI can respond to queries, create images, and write in human language — all key building blocks of AGI.
• Embodied AI - gives artificial intelligences the opportunity to learn from their physical environment by grounding them in the real-world. This may be a precursor to the concept of AGIs mirroring human ability.
• Neuro-symbolic AI - combines the strengths of neural networks (great at learning from data) with symbolic reasoning systems (great at following rules) to create systems that can learn from experience while also using common sense, which is crucial to human thinking.
• Computer vision - refers to AI’s ability to extract and comprehend information from the world around it. This is the technology you’ll often find in self-driving cars, and it could one day enable AGI to navigate complex environments.

All of this reveals one thing: We aren’t as far away from AGI technology passing the Turing Test as we think.

The development of AGI will transform our lives in a limitless number of ways, performing tasks we already do with efficiency that is hard to imagine.

Imagine a kind of ‘everyday AGI copilot’, ready to provide whatever input or guidance its user might need and take friction out of life. They will take care of shopping, plan a party, make a financial decision, or mediate a disagreement with a friend – adapting to the user’s needs.

These are all small changes to improve daily life, but let’s go bigger. What if AGI could transform our healthcare by detecting early signs of disease and offering effective treatments? Imagine a system that could detect early cancer cells dividing and recommend genome-specific therapies to eliminate the disease.

For entertainment, AGI systems won’t just be able to recommend movies and books you’re likely to enjoy. They could create such works entirely from scratch based on your preferences and desires. This kind of capability is already underway with LLMs. But in the future, could we see entire storylines and translate them into screenplays and generate HD video to bring it to life?

Although it’s hard to conceive today, AGI agents would make for such great conversation partners and creative collaborators that, over time, we’d come to view them as peers, and maybe even friends. As AGI develops, its place in our world may soon be as social as it is practical. Considering that humans have developed unexpectedly deep bonds with machines as far back as the ELIZA chatbot in the 1960s‌ — ‌about as simple and limited a conversation agent as can be imagined — we shouldn’t be surprised by the possibility of superhuman intelligence becoming a captivating companion.

AGI promises to transform the way we work and redefine productivity. Its reasoning abilities, access to data, and conversational skills will make it a natural collaborator and even mentor for almost any kind of career, whether it’s simply gut-checking a decision or asking for a complete strategy to accomplish a task.

Just as AGI can boost individual health, it could also enhance organisational stability. Imagine bringing expert level machine learning and intelligence‌ to every moment of every day of a business’s life, ensuring that problems are anticipated as far in advance as possible, and addressed in the very first moment action can be taken.

In science, AGI could act as an AI alchemist — designing and producing novel materials such as proteins for biotechnology and crystal structures that lead to entirely new alloys and materials. In robotics, AGI may power maximally-efficient machines in warehouses, surpassing human-level performance by unleashing productivity while reducing errors.

Connecting the human brain to artificial superintelligence also has enormous potential for a variety of tasks. Neural prosthetics could restore memory or motor functions. AI-powered limbs could directly connect with our neural signals, mirroring human capabilities. The possibilities are, quite literally, endless.

The futures of AI and AGI systems depend on building safeguards to make sure that AI systems do not make wrong decisions, are competent at their assigned task, and are more immune to cyber attack, where bad actors can take control or hijack the AI systems and perform hostile actions. It’s also coming up with frameworks to account for mistakes. What happens if our efforts go wrong? Only when we can solve these problems will we have achieved AGI.

The way forward with AGI is for AI researchers to make trust the number one priority, which is a significant challenge given AGI will perform tasks with less human supervision (ideally none). This is a pivotal moment, the time that historians will look back on. If we can get it right, we step through perhaps the biggest door to opportunity that humanity has ever opened, an event even more life-changing and economically valuable than the Industrial Revolution that preceded it.

• Uncover original research, execute perspectives and the latest AI developments from the world’s #1 AI CRM on the AI Research page
• Find more stories on tech ethics on the Salesforce Newsroom. Looking for a place to start? Try this byline from Salesforce’s Principal Architect for Ethical AI, or this story about how Salesforce AI research is shaping society.