We‘re in the AOL Phase of Artificial Intelligence, Says Tech CEO as Industry Raves About AI

The rise of artificial intelligence (AI) has been one of the most exciting and transformative technological developments in recent years. From chatbots and virtual assistants to autonomous vehicles and drug discovery, AI is already revolutionizing industries and changing the way we live and work. The global AI market is expected to grow from $387 billion in 2022 to over $1.3 trillion by 2029, according to Fortune Business Insights. Venture capital funding for AI startups hit an all-time high of $36 billion in 2021, per CB Insights.

Amidst the hype and excitement around AI‘s potential, one tech leader is offering a more measured perspective. Sachin Duggal, the co-founder and CEO of AI development platform Builder.ai, believes we are currently in what he calls the "AOL phase" of AI. Just as AOL made the internet accessible and understandable to the masses in the 1990s but was a far cry from the high-speed broadband we enjoy today, Duggal suggests the AI systems of 2023 are still a primitive precursor to the truly intelligent machines that will emerge in the coming decades.

"The perception is that we‘re in the fiber-optic world of AI, when really, we‘re in the dial-up age," Duggal said in a recent interview with CNBC. "We‘ve only just begun to imagine what‘s possible with artificial intelligence. There is still a huge amount of fundamental research and development that needs to happen before we get to artificial general intelligence."

To put this in context, it‘s helpful to understand the different types of AI and how the field has evolved. Narrow or weak AI refers to systems that are trained to perform specific tasks, such as playing chess, detecting spam, or translating languages. These systems can match or exceed human performance in their narrow domains but lack the adaptability and breadth of human intelligence. Artificial General Intelligence (AGI) refers to machines that can think, learn, and reason like humans across any domain – a sort of all-purpose intelligence. Superintelligent AI takes this even further, envisioning machines that radically exceed human cognitive abilities and rapidly bootstrap themselves to even greater levels of intelligence.

The concept of AI dates back to the 1950s, when computer scientists like Alan Turing, John McCarthy, and Marvin Minsky first began exploring the potential for machines to exhibit intelligent behavior. Early milestones included Arthur Samuel‘s checkers program (1952), the Logic Theorist (1956) and General Problem Solver (1957) systems developed by Herbert Simon and Allen Newell, and Joseph Weizenbaum‘s ELIZA chatbot (1966). However, these early systems were heavily rule-based and brittle, leading to a series of "AI winters" in the 1970s and 80s where funding and enthusiasm dried up.

The current AI boom began in the early 2010s with the rise of machine learning, and in particular, deep learning neural networks. Fed huge amounts of training data, these systems are able to learn patterns and representations in a way that mimics the human brain. Breakthroughs like Geoffrey Hinton‘s backpropagation algorithm (1986), Yann LeCun‘s convolutional neural networks for image processing (1998), and Yoshua Bengio‘s work on unsupervised learning have enabled machines to achieve superhuman performance on narrow tasks like facial recognition (2014), playing Atari games (2015) and Go (2016), and speech synthesis (2016).

The excitement around the latest AI chatbots like OpenAI‘s ChatGPT and Google‘s Bard is understandable. By ingesting enormous machine learning models trained on terabytes of online data, these systems are able to engage in remarkably fluent and coherent conversations, compose original stories and scripts, answer follow-up questions, and even write functional code. A December 2022 survey found that ChatGPT was able to fool over 30% of Americans into thinking it was human. Over 100 million users have engaged with ChatGPT since its launch in November 2022, making it the fastest growing consumer app in history according to UBS.

But Duggal and other AI experts caution that today‘s chatbots are still narrow AI systems trained to predict the most likely next word based on statistical patterns, not truly intelligent agents that understand meaning and context the way humans do. "An LLM (large language model) is simply telling you what it thinks the next word is with a high degree of probability, whereas a knowledge graph can compose pattern relationships that it knows and how things work out. So it‘s not just predicting what‘s next," he explained.

Image Source: Forbes

Indeed, closer examination of chatbot outputs often reveals incoherence, inconsistencies and "hallucinations" where the AI confidently makes statements that are nonsensical or factually untrue. Their ability to engage in open-ended conversation creates an anthropomorphic illusion of intelligence that doesn‘t match reality. A January 2023 study from the University of Waterloo found that ChatGPT performed at only a 54% accuracy level on a set of reading comprehension questions.

Duggal believes more promising than gigantic language models are knowledge graphs – AI representations that map the relationships and logic between concepts. "A knowledge graph can actually reason – it can say, okay, if this thing is true, then this other thing must also be true," he said. Google‘s Knowledge Graph, launched in 2012, is perhaps the best known example – it encodes over 500 billion facts about 5 billion entities and is used to power semantic search and questions answering in Google Search. Other notable knowledge graphs include Microsoft‘s Satori, Amazon‘s Product Graph, Facebook‘s Entity Graph, and IBM‘s Watson.

Looking further into the future, Duggal envisions AI systems that can learn and adapt on their own, form their own goals, and interact naturally with humans through dialog, gestures and facial expressions. He imagines personal AI assistants that know our preferences, can anticipate our needs, and help us be more productive and creative in our daily lives. But to get there, he says the field needs to go back to basics and focus on fundamental breakthroughs in areas like:

• Reasoning and abstraction: Giving machines the ability to form high-level concepts, draw inferences, and transfer knowledge between domains. Early work in this area includes Judea Pearl‘s Bayesian networks and Daphne Koller‘s probabilistic graphical models.

• Open-ended learning: Enabling AI systems to learn continuously and accrue knowledge over time without hitting plateaus or "catastrophically forgetting" previously learned information. Promising approaches include meta-learning, online learning, and neural Turing machines.

• Computational neuroscience: Drawing inspiration from the structure and function of biological brains to build more efficient and adaptive AI systems. Research labs like DeepMind, Google Brain, and OpenAI are heavily investing in areas like synaptic plasticity, neurogenesis, and neural circuit mapping.

• Embodied cognition: Grounding AI systems in the physical world through robotic platforms that allow them to perceive, manipulate, and learn from their environment. Companies like Boston Dynamics, Kindred AI, and Covariant are at the forefront of this field.

• Commonsense knowledge: Endowing machines with the vast web of implicit knowledge and assumptions that humans rely on to understand and interact with the world. Projects like Cyc, ConceptNet, and ATOMIC are working to codify commonsense knowledge into machine-readable formats.

Just as importantly, Duggal stresses the need for responsible AI development practices that prioritize transparency, fairness, and robustness as these systems become more advanced. He advocates for proactive governance frameworks that bring together technologists, ethicists, policymakers, and the public to collectively shape the future of AI. "We have to get the governance piece right," he said. "We can‘t just leave it to the tech companies to self-regulate. We need guardrails and accountability measures in place as the technology gets more powerful."

Despite the long road ahead to AGI and beyond, Duggal is optimistic about the potential for AI to solve some of the world‘s most pressing challenges and create massive value for society. A 2019 report from McKinsey estimated that AI could deliver up to $13 trillion in additional global economic activity by 2030, and help solve problems in areas like healthcare, education, transportation, public safety, and sustainability.

Some examples of AI already driving positive impact and industry transformation include:

• Healthcare: AI is being used to accelerate drug discovery, personalize treatments, and detect diseases earlier. A 2020 study found that Google‘s DeepMind AI was able to predict acute kidney injury up to 48 hours in advance with 90% accuracy.

• Education: AI tutoring systems and adaptive learning platforms are helping students learn at their own pace and freeing up teachers to provide more individualized attention. A 2023 paper found that an AI tutoring system called Korbit improved student learning outcomes by 25% compared to traditional lectures.

• Transportation: Self-driving cars and intelligent traffic management systems have the potential to drastically reduce accidents, congestion, and emissions. Waymo, the leading autonomous vehicle company, has driven over 35 million miles on public roads and plans to launch a fully driverless ride-hailing service in Phoenix this year.

• Climate: AI is being applied to optimize renewable energy systems, model climate change scenarios, and accelerate the development of low-carbon technologies. The AI for Good Foundation is a United Nations initiative that aims to scale AI solutions for sustainable development goals.

As the AI hype cycle reaches a peak, Duggal‘s perspective provides a helpful reality check. Today‘s AI systems are undoubtedly impressive and worthy of enthusiasm, but we are likely decades away from the science fiction future of superintelligent AI imagined by many. The road to artificial general intelligence will require immense scientific and technological advancement beyond simply scaling up today‘s neural networks.

Ultimately, the journey to advanced AI is a marathon not a sprint. Just as the dial-up era of AOL gave way to gigabit broadband, so too will today‘s large language models and knowledge graphs lay the foundation for ever more sophisticated artificial minds. By continuing to invest in fundamental research, responsible development, and human-centric applications, we can navigate the hype cycles and build an AI-powered future that genuinely benefits humanity. The AOL phase is an exciting milestone in that journey – but still just one early step toward the intelligent machines of tomorrow.

As Duggal puts it: "We‘re in the very early innings of AI – maybe the second or third inning of a nine inning ball game. There‘s certainly a lot to be excited about, but we need to balance that with humility about how far we still have to go. If we get it right, AI could be the greatest technology humanity has ever created. But we have to do the hard work to get from here to there – not just technologically, but also ethically and socially. It‘s going to take all of us working together to build the AI future we want to see."