PDF Summary:Co-Intelligence, by Ethan Mollick

Artificial intelligence is rapidly changing how we work, learn, and live our lives. In Co-Intelligence, Ethan Mollick explores how AI systems function and what their capabilities mean for society. He explains the mechanisms behind large language models and image generators, including how they're trained and why they sometimes produce unreliable results.

Mollick examines how AI can enhance productivity by handling repetitive tasks and supporting creative work, while also addressing concerns about bias, copyright issues, and potential misuse. He discusses AI's applications in education, where it can personalize learning experiences and assist teachers, and considers broader implications for the workplace and economy. This guide offers a framework for understanding when to collaborate with AI, when to work independently, and how to navigate the challenges and opportunities that come with this technology.

(continued)...

Self-Training AI

The idea of AI learning from data it helped to create has been explored before. In 2020, Qizhe Xie et al. published a paper on “Self-training with Noisy Student,” which showed that a model could be made stronger by training it on examples labeled by another model. This approach improved accuracy and robustness on ImageNet classification benchmarks. The authors found that when the student model is larger than the teacher and is trained with noise on both labeled and unlabeled data, self-training becomes significantly more effective. This research anticipated today’s interest in recycling AI-generated content for further training.

Mollick adds that after pretraining, AI companies use human feedback to adjust AI models. This method is called RLHF, which stands for learning by reinforcement through feedback from humans. Companies hire people to evaluate AI answers based on various qualities. This input is subsequently utilized to further train and adjust the AI's behavior according to human preferences. This method enhances effective replies and diminishes poor ones.

(Shortform note: RLHF is part of a broader research tradition called reinforcement learning, which studies how agents can learn to make decisions that maximize long-term rewards. This field has developed mathematical frameworks and algorithms that allow agents to learn from trial and error, adjusting their actions based on feedback from the environment. Sutton and Barto (2018) provide a comprehensive overview of this field, covering key concepts such as value functions, policy optimization, and exploration-exploitation trade-offs.)

Uses and Implications of Group Intelligence

Mollick believes AI could dramatically transform our approach to life, education, and work. It can assist in transcending the boundaries of human science, resulting in advancements in how we comprehend the universe and ourselves. He adds that artificial intelligence can make scientific discoveries faster, increase productivity, and broaden educational access. Additionally, it can unlock human potential by serving as an inspirational resource and a means of delegating time-consuming tasks. For instance, AI-based counselors and aides can support individuals in enhancing themselves through innovative methods. Mollick also notes that robots and autonomous agents with AI capabilities could dramatically decrease reliance on human labor and simultaneously grow the economy.

(Shortform note: In Power and Progress, economists Daron Acemoglu and Simon Johnson challenge the idea that AI will simultaneously reduce the need for human labor and increase prosperity. They argue that many current uses of AI fall under what they call “so-so automation,” where companies adopt technologies that slightly lower labor costs but don’t significantly boost productivity. This kind of automation can displace workers and weaken their bargaining power without creating broad-based increases in incomes or shared prosperity.)

This would necessitate a significant reevaluation of our societal and work paradigms. Reduced work hours, a guaranteed minimum income, along with additional policy shifts could become feasible as demand for human labor declines. We'll have to discover different meaningful activities for our leisure time, since our lives now are largely centered on work. Mollick suggests that AI companions might be more engaging conversational partners than many people, leading some to prefer interacting with AIs over humans. Entertainment powered through artificial intelligence will deliver tailored, distinct experiences that blend gaming, narratives, and films.

(Shortform note: If we reduce work hours and spend our newfound leisure time with AI companions and entertainment powered through artificial intelligence, we may become less inclined to engage with other people. In Alone Together, Sherry Turkle argues that as we become more reliant on technology for social interaction, we risk losing the ability to form meaningful connections with others. She suggests that we may become so accustomed to the instant gratification and low-stakes interactions provided by AI that we lose the motivation to put in the effort required to build and maintain real-world relationships. This could lead to a society where people are increasingly isolated and disconnected from each other.)

However, artificial intelligence may also be used for destructive purposes, such as hacking computer systems, creating hazardous chemicals and pathogens, and amplifying the efforts of militaries and criminals. Governments might employ AI to combat terrorism and crime that use AI, posing the risk of AI-based authoritarianism, since omnipresent surveillance allows democracies and dictators to exert greater control over citizens.

(Shortform note: AI-based authoritarianism is a form of government in which artificial intelligence is integrated into the data-gathering and decision-making systems of a country. This allows governments to track the activities of their citizens in great detail and automatically adjust benefits or punishments to keep political opposition weak. This approach allows governments to maintain control without relying solely on visible, brute-force repression.)

Next, we will explain the core characteristics and limitations of AI, and then discuss its practical applications and strategic responses to it.

Core Characteristics and Limitations of Co-Intelligence

Observable Behaviors and Output Quality

Mollick explains that AI might produce generic outputs without context or constraints because it often adheres to the common patterns in the text it was taught from. To produce more engaging and practical results, you should add background and boundaries.

(Shortform note: Think of this as a mini design brief: State your role, your audience, your constraints, and one clear success criterion. This gives the AI a clear framework to work within, leading to more focused and useful results.)

Mollick further explains that AI might exhibit bias because it’s trained on biased data. It can also be biased because the people training it are biased. AI firms are working to tackle bias in various ways, each with its own sense of urgency. Some of them just cheat, like the image generator DALL-E, which covertly inserted the word "female" into various prompts to create an image of "a person" so as to compel some gender diversity not found in the training data.

(Shortform note: Some people might argue that this isn’t cheating. In Weapons of Math Destruction, data scientist Cathy O’Neil argues that it’s an ethical obligation to counteract the structural discrimination in the data that generative systems are trained on. She explains that algorithms are designed to codify the past, not invent the future. If we want algorithms to be fair, we can’t just feed them biased data and hope for the best: We have to decide what kind of world we want to encode and then deliberately design, test, and correct our models so they stop amplifying existing inequities and start actively counteracting them.)

Another method could be altering the training datasets to cover a broader range of human experiences, though, as we've observed, collecting training data presents challenges. A frequent method to lessen bias involves human correction of AI systems, exemplified by a technique where humans guide reinforcement learning, used to refine LLMs. This approach lets human evaluators discipline the system if it generates harmful material, such as racism or incoherence, and to praise it for creating quality material. Through RLHF, what's generated gets better over time in various aspects, becoming more objective, precise, and useful. However, biases aren't guaranteed to disappear. During this phase, the predispositions of human evaluators and the organizations managing their work may begin to affect the AI, potentially leading to novel biases.

(Shortform note: Christiano et al. were among the first to propose the idea of humans guiding reinforcement learning. They had humans compare short clips of an AI agent’s behavior and choose which one they preferred. These preferences were used to train a reward model that the AI then used to improve its behavior. This approach allowed the AI to learn complex tasks without needing a hand-crafted reward function, showing that human feedback could effectively guide reinforcement learning.)

Practical Applications and Strategic Responses to Cooperative Intelligence

Collaborative Strategies Using Co-Intelligence

Mollick suggests using AI to perform tasks that are repetitive or time-consuming. By doing so, you can free up your time for more meaningful work, making your job more enjoyable and allowing you to use your skills and capabilities more effectively.

(Shortform note: While using AI for repetitive or time-consuming tasks can make your job more enjoyable, it can also have a downside: It can erode your skills over time. For example, if you rely on AI to write emails, you may lose your ability to write them yourself. This can leave you vulnerable if the AI system fails or is unavailable.)

He also recommends collaborating with AI by assigning tasks according to strengths and weaknesses. The activities fall into three categories: Only Me Assignments, Assignable Duties, and Automated Responsibilities. Just Me Tasks are tasks that you do yourself, either because AI can’t handle them or because you don’t want to assign them to AI. Delegated tasks are assignments you give AI and then check over. Automated Tasks are things you assign to AI without reviewing them.

Mollick notes that AI is skilled at some things and bad at others. Similarly, you have strengths in some areas and weaknesses in others. You can leverage artificial intelligence to complete tasks you're not good at, while handling the things that artificial intelligence struggles with. You can also use AI to do things you’d prefer not to handle.

How to Assign Tasks to AI

When deciding which tasks to assign to AI, consider the consequences of a mistake. If a mistake would be catastrophic, keep the task in the Only Me category. If a mistake would be inconvenient but not disastrous, assign it to the Assignable Duties category. If a mistake would be easily reversible, assign it to the Automated Responsibilities category. For example, if you’re a doctor, you might keep diagnosing patients in the Only Me category, assign writing patient notes to the Assignable Duties category, and assign scheduling appointments to the Automated Responsibilities category.

Societal and Workplace Transformations

Mollick believes AI is transforming education by enhancing learning experiences and altering how instruction is delivered. For instance, it enables educators to create engaging, personalized learning experiences, such as games, activities, evaluations, and models. AI instructors can customize teaching for individual students' requirements and modify material according to their performance. This allows teachers to spend more time interacting with students and providing personalized instruction. AI can also assist teachers in crafting lectures that are more engaging and transforming passive lectures into active ones. Learners are currently utilizing artificial intelligence for educational purposes, and instructors are leveraging it to prepare for class.

The History of Educational Technology

The history of educational technology is a story of recurring cycles of innovation, hype, and disillusionment. In the mid-20th century, psychologist B.F. Skinner developed mechanical “teaching machines” that presented students with programmed instruction and immediate feedback. Skinner envisioned these machines as a way to provide individualized instruction and free teachers from repetitive tasks. However, as Audrey Watters explains in Teaching Machines, these early efforts faced significant challenges in practice. Teachers often found the machines cumbersome and time-consuming to use, and students quickly lost interest in the repetitive exercises. Later computer-based systems like PLATO and TICCIT promised more engaging, interactive learning experiences, but they too struggled to live up to their initial hype. Watters argues that these technologies often failed to address the underlying power dynamics in education, as they were typically designed by technologists rather than educators.