The Summer That Sparked AI: A Journey from Dartmouth to Deep Learning's Rise
In the summer of 1956, a select group of brilliant minds gathered at Dartmouth College in New Hampshire, among them Claude Shannon, the father of information theory, and Herb Simon, the only individual to ever win both the Nobel Memorial Prize in Economic Sciences and the Turing Award. The meeting was organized by a young researcher, John McCarthy, who sought to explore how machines could "use language, form abstractions and concepts," and solve problems typically reserved for humans.
This gathering marked the first academic meeting focused on what McCarthy would later call “artificial intelligence” (AI), setting the stage for a field that, despite its lofty ambitions, would struggle to achieve breakthroughs matching its early aspirations.
Though the Dartmouth meeting is often seen as the starting point for AI, the concept of thinking machines had already intrigued figures like Alan Turing and John von Neumann. By 1956, various approaches to creating such machines were already in development. McCarthy’s term "artificial intelligence" was broad enough to encompass these diverse efforts, from systems that used logical rules to infer conclusions to those that relied on probabilities.
In the years that followed, the AI field experienced intense debate and innovation, culminating in the 1980s with the rise of "expert systems." These systems, which captured human knowledge using symbolic logic, received substantial support, particularly from the Japanese government. However, they ultimately proved too rigid for the complexities of the real world, leading AI to fall out of favor by the late 1980s as it became synonymous with unfulfilled promises.
Yet, the seeds of today’s AI renaissance were sown by those who persisted. Inspired by the way neurons work in the human brain, researchers began experimenting with artificial neural networks. Initially modeled in hardware by Marvin Minsky, a participant at Dartmouth, these networks later evolved into software simulations. Unlike traditional AI systems, neural networks weren’t explicitly programmed; they learned by example, adjusting the connections between simulated neurons to produce desired outcomes.
A Heartfelt Letter to My Newborn Daughter Eva | Kamal H. Tamini
Despite initial setbacks, a breakthrough came in 2009 when researchers at Stanford University dramatically accelerated the speed of neural networks using a gaming PC. The key was the graphics processing unit (GPU), which was well-suited to running neural-network code. This advance, combined with more efficient training algorithms, allowed for the development of "deep learning," where networks with millions of connections could be trained to handle complex tasks like image recognition.
The power of deep learning was showcased in 2012 when a team led by Geoff Hinton at the University of Toronto achieved an 85% accuracy rate in the ImageNet Challenge, a significant leap forward. By 2015, deep learning had become the standard in image recognition, surpassing human accuracy, and expanding into other fields such as speech recognition, face recognition, and translation.
The explosion of data available on the internet and the vast potential markets it represented played a crucial role in deep learning’s success. As networks grew larger and were trained on more data, their performance continued to improve, leading to the development of new products and services, from Amazon’s Alexa to automatic translation tools.
A significant evolution came in 2017 with the introduction of transformers, a new architecture that allowed neural networks to better understand context by focusing on specific features in the data. This innovation led to the rise of large language models (LLMs) like OpenAI’s GPT-2 in 2019, which demonstrated "emergent" behaviors not explicitly programmed, such as the ability to perform simple arithmetic and write code. However, these models also mirrored the biases present in their training data, reflecting societal prejudices in their outputs.
The public unveiling of GPT-3.5 in November 2022, through the chatbot ChatGPT, marked another leap forward for AI. Within weeks, it was generating everything from college essays to computer code, capturing the world’s imagination. While early AI applications focused on recognition, this new wave is centered on generation, with models like Stable Diffusion and DALL-E creating images, videos, and even music from text prompts.
This series will explore how these advanced models work, their potential future applications, and the ethical considerations surrounding their use. AI has indeed come a long way since that fateful summer at Dartmouth, and its journey is far from over.
