Generally AI - Season 2 - Episode 3: Surviving the AI Winter

Early AI and Perceptrons

• In 1966, Seymour Papert initiated the Summer Vision Project, aiming to develop a visual system capable of pattern recognition by breaking it down into sub-problems, such as dividing images into regions like likely objects and backgrounds. (0s)
• The project aimed to handle complex patterns and backgrounds, exemplified by objects like cigarette packs, reflecting the social norms of the time. (54s)
• In 1980, Seymour Papert authored the book "Mindstorms: Children, Computers, and Powerful Ideas," proposing that children learn faster and become more social when using computers. This book inspired the name for LEGO's Mindstorms robotics kits. (1m26s)
• The discussion introduces the concept of AI winters, periods of reduced funding and interest in AI, and notes that many technologies from these times, such as neural networks, remain relevant today. (2m25s)
• The term "artificial intelligence" was coined in the 1950s, with early AI research exploring various technologies, including symbolic logic and search trees, alongside neural networks. (3m38s)
• Frank Rosenblatt developed the perceptron in 1957, marking an early advancement in neural network technology. (4m33s)
• A perceptron is a mathematical model of a single biological neuron, which has several numeric inputs, computes a weighted sum of those inputs, and thresholds that weighted sum to decide whether to fire its output (4m46s).
• The perceptron's weights can be taught using supervised learning, and the learning algorithm involves putting in inputs, knowing the expected output, getting the actual output, and updating the weights iteratively (5m29s).
• Rosenblatt wrote a software implementation of the perceptron, which could classify a 72x72 pixel image, and in one experiment, it could distinguish between the letter "e" and the letter "X" even if they were rotated, with about 95% accuracy (5m55s).
• Despite this achievement in 1960, AI research did not progress as expected, and one reason was the publication of a paper by Seymour Papert and Marvin Minsky in 1969, which showed a severe limitation of perceptrons: they can only learn linearly separable classes (6m54s).
• This limitation led many people to give up on perceptrons, and AI research funding was cut in the US and the UK, with the US government pulling funding for machine translation in 1966 and the British government publishing the Lighthill report in 1973, which criticized the lack of progress in AI (8m9s).
• The Lighthill report led to a cut in funding for AI research in British universities, and DARPA changed its policies to focus on concrete projects with obvious military applications, marking the end of the "AI summer" and the beginning of the "AI winter" (8m40s).

AI Winters and Reemergence of Neural Networks

• Despite the lack of funding, some AI research continued, and there were other projects in the field, including a machine translation project in the 1960s that did not meet expectations (7m33s).
• There was a period known as the AI winter when funding for AI research significantly decreased, although interest and study in AI continued in some institutions like Stanford AI Lab and Carnegie Mellon. (8m59s)
• During this time, some researchers continued working with perceptrons, which were rebranded as artificial neural networks. Minsky and Papert noted that neural networks could learn complex functions like exclusive or (XOR) with multiple layers of perceptrons, but the training algorithm for these networks was not yet known. (9m27s)
• In 1985, researchers including Jeff Hinton developed an algorithm for training multi-layer perceptron networks using gradient descent with backpropagation, which revitalized interest in neural networks. (10m19s)
• By 1989, Yann LeCun was using multi-layer neural networks to recognize handwritten digits with 95% accuracy, demonstrating practical applications such as sorting letters in post offices. (11m9s)
• The early 1980s saw a boom in AI with the popularity of expert systems and significant investments from Japan's fifth-generation computing project and the US's DARPA. However, this was followed by a second AI winter in the 1990s. (11m42s)
• The second AI winter was partly due to limitations in the backpropagation training algorithm, which struggled with very deep networks due to issues like vanishing and exploding gradients. Additionally, hardware limitations and insufficient data sets contributed to the decline. (12m22s)
• In 2012, Geoffrey Hinton and his colleagues at the University of Toronto published results using the MNIST dataset, achieving over 99% accuracy in recognizing handwritten digits. (13m21s)
• The ImageNet dataset, consisting of millions of images, became the gold standard for object recognition, and Hinton's student developed AlexNet, which achieved about 85% accuracy, significantly surpassing the previous record. This marked the beginning of the modern era of deep learning. (13m49s)
• The success of ImageNet and AlexNet initiated a period of rapid advancement in AI, with the understanding that increasing data, model size, and computational power leads to predictable improvements in training outcomes, as demonstrated by OpenAI's scaling laws. (14m40s)

Modern Deep Learning and Potential AI Winters

• The second AI winter is considered to have occurred during the 1990s and early 2000s, with discussions about a potential new AI winter emerging around late 2018 to early 2019, as some predicted that deep learning might soon reach its limits. (15m6s)
• Despite a drop in AI publications in 2021 and 2022, the release of ChatGPT in December 2022 marked a resurgence in AI activity, leading to a period described as an "AI summer," although predictions of another AI winter persist. (16m10s)
• There is ongoing debate about the capabilities of large language models (LLMs) in achieving artificial general intelligence (AGI), with some experts like Yann LeCun expressing skepticism about their potential to reach AGI. (17m16s)
• Rodney Brooks, who founded iRobot, has been involved in AI for a long time and maintains a blog where he made predictions about technological advances, including AI, autonomous vehicles, and space flight. In 2018, he predicted that the next significant development in AI would occur between 2023 and 2027, and it would be based on existing research. (17m40s)
• Brooks identified ChatGPT as the next big thing in AI, noting that the Transformer paper, which is foundational to ChatGPT, was published in 2017. He considers this a successful prediction. (18m32s)
• In January of the current year, Brooks predicted another AI winter, suggesting a period of slowed progress in AI development. He acknowledged the difficulty in predicting the exact timing of such events. (19m5s)
• During the late 1990s, a period considered an AI winter, research in intelligent systems and behavior-based robotics was conducted, contrasting with traditional AI approaches like 3D modeling and planning. (19m56s)
• Neural networks, once considered outdated technology, gained renewed interest and relevance with developments like AlexNet in 2014, illustrating the cyclical nature of AI advancements. (20m48s)

A* Search Algorithm and its Applications

• The discussion focuses on AI algorithms that have stood the test of time, specifically mentioning the A* search algorithm, which remains relevant and well-known even outside the AI community. (22m28s)
• The A* search algorithm was invented in 1968 by Peter Hart, Nils Nilsson, and Bertram Raphael at the Stanford Research Institute, originally for Shakey the robot's path planning. (24m21s)
• There is a discrepancy in historical accounts, as the report for Shakey the robot mentions using a breadth-first search rather than A, suggesting that A was considered too slow for grid-based search at the time. (24m59s)
• Older research papers often contain inventive solutions to problems, such as defining obstacles and using tangents for navigation instead of modern grid-based searches. (25m29s)
• The discussion highlights the optimization of robot path planning by reducing the search space, allowing the robot to navigate from object edge to object edge efficiently. (25m57s)
• There is a mention of representing images on punch cards, which is considered an interesting topic for future exploration. (26m18s)
• AAR (A* algorithm) is compared to Dijkstra's algorithm, with the key difference being that AAR is point-to-point, while Dijkstra's calculates the distance to the entire world. (26m28s)
• A heuristic is used in AAR to prioritize paths that bring the solution closer, similar to how a person would navigate from Amsterdam to Paris by following a direct path and backtracking if necessary. (26m50s)
• The versatility of the A* algorithm is noted, with applications in chess and video games like Super Mario, where heuristics help in decision-making and optimizing game states. (27m35s)
• The algorithm's adaptability is emphasized, with its use in various applications beyond mapping, and the importance of incorporating heuristics to improve search efficiency. (28m27s)
• A potential future discussion is suggested on comparing A* with reinforcement learning, focusing on selecting the next best action based on the current state and desired outcome. (28m40s)
• A* can be slow if the heuristic is not effective or if there is a large space to explore, as visualized in examples like the Wikipedia article on search horizon expansion. (29m8s)
• Contraction hierarchies are used in mapping to simplify navigation by contracting parts of the graph, allowing for efficient travel over long distances by ignoring smaller streets and focusing on higher-level navigation. (29m52s)
• To prevent cycles in navigation, it is important to keep track of which nodes have already been expanded or visited, ensuring that the fastest path is found without unnecessary expansions. (30m14s)
• Since 1968, mapping algorithms have improved, but specific routing needs, such as those for trucks with weight restrictions or sports cars seeking scenic routes, often lead companies to revert to the A* algorithm due to its flexibility in handling various constraints. (30m43s)
• The heuristic used in mapping can be customized to include factors like distance, road width, or windiness, allowing the A* algorithm to guide results according to specific constraints or desires. (31m55s)
• The A* algorithm is versatile, capable of handling different constraints by determining whether a path is feasible for a given vehicle, such as a truck with weight limitations, and avoiding expansion of nodes that do not meet the criteria. (32m32s)
• Factorio is a highly addictive factory-building game that appeals to those who enjoy automating tasks and solving problems, particularly programmers, due to its engaging and challenging nature. (33m0s)
• In a game scenario, players can shoot aliens from a distance, prompting the aliens to plan a path to attack the player's base. This pathfinding can be slow on large maps, especially when navigating around obstacles like lakes. (33m47s)
• To address this, a modified heuristic was implemented, allowing for faster pathfinding by dividing the world into larger tiles and calculating whether each tile is crossable. This approach helps direct the search more efficiently. (34m15s)
• The modified algorithm uses a heuristic that estimates the distance back to the target location, enabling quicker navigation in the desired direction while still employing a standard A* algorithm. (35m20s)
• The A* algorithm is described as a flexible and efficient tool, likened to a Swiss army knife, due to its adaptability and ease of implementation. (35m52s)
• There is a discussion about the history and academic debates surrounding the A* algorithm, including claims about its optimality and subsequent challenges and validations by researchers. (36m14s)
• The A* algorithm is praised for its reliability and efficiency, having persisted through various AI development phases, similar to how mammals survived after the extinction of dinosaurs. (37m25s)

Science Fiction and Early Technological Advancements

• A book titled "The Moon is a Harsh Mistress" by Robert Heinlein, written in 1966, features a self-aware computer with associative neural networks, which is notable given the era it was written. (38m15s)
• In the book, the computer creates a photorealistic CGI version of a person, a concept that was ahead of its time and still challenging to achieve today. (38m50s)
• Robert Heinlein's earlier works from the 1940s and 1950s included advanced space technology but lacked the presence of computers, which he later incorporated in the 1960s with self-aware machines. (39m19s)
• Heinlein's stories from the 1950s and 1960s sometimes featured technology like cell phones, which were not yet invented, showcasing his imaginative foresight. (39m52s)
• The first successful manned flight was conducted by the Montgolfier Brothers in 1783, a significant historical milestone. (40m30s)
• Benjamin Franklin reportedly witnessed the Montgolfier Brothers' flight and responded to a remark about its uselessness by comparing it to a newborn baby, highlighting the potential of new inventions. (40m53s)
• The modern parachute was also invented in 1783, and a frameless parachute was used in a balloon trip in 1797, marking early developments in parachute technology. (41m14s)
• The concept of parachuting originated with people jumping from buildings, but later balloons were used to jump from higher altitudes, with the first parachute demonstration conducted by someone in 1785 as a means of safely disembarking from a hot air balloon (42m6s).
• The first parachute demonstrations were conducted with a dog as a passenger, similar to the Russians sending a dog into space to test the parachute (42m18s).
• George Kayy invented a glider, essentially an airplane without a motor, and referred to it as a "governable parachute" in the Mechanics Magazine on September 25, 1852 (42m38s).
• The term "governable parachute" was used to describe gliders or airplanes before they were commonly known as such, as people were familiar with parachutes but not airplanes (43m0s).
• George Kayy's idea for the governable parachute involved hanging it behind an air balloon and releasing it to steer the parachute, but he lacked a fast enough motorboat to test his idea (43m28s).
• The illustration of Kayy's governable parachute resembles a massive boat or chariot, with Kayy sitting inside (43m55s).

Smart Watches and Personal Experiences

• The host of the podcast, Roland, switched from an Apple watch to a Garmin watch, which has provided more accurate timekeeping and longer battery life, lasting up to 17 days (44m53s).
• The Garmin watch also works with Roland's iPhone, showing messages and providing a map during a 25-kilometer hike, using only a quarter of its charge (45m32s).