PDF Summary:The Worlds I See, by Fei-Fei Li

Artificial intelligence has evolved from an academic concept to a technology that shapes daily life, from business operations to social interactions. But this rapid development has raised concerns about bias, privacy, and AI's broader impact on society. In The Worlds I See, Fei-Fei Li traces AI's history from the 1956 Dartmouth workshop through modern breakthroughs in deep learning and neural networks, explaining how technologies like transformers and GPUs have enabled machines to process vast amounts of data and perform tasks once thought impossible.

Li also addresses the ethical challenges that accompany AI's advancement. She argues that AI's ultimate impact depends on the intentions behind its creation and advocates for a human-centered approach that prioritizes fairness, representation, and collaboration across disciplines. Through her work at Stanford's Institute for Human-Centered Artificial Intelligence, Li promotes responsible AI development that enhances human capabilities rather than competing with them.

(continued)...

Li also notes that RNNs were key in early language processing advancements. These algorithms are intended to align with the linear order of words. They can swiftly deduce fundamental characteristics of text, similar to how CNNs handle visual data.

(Shortform note: In this context, “aligning with the linear order of words” means that the network updates a single memory state one word at a time, so that each new word is interpreted in light of all the words that came before it.)

The Power of Scale and Data

According to Li, scale and information are crucial for pushing AI forward. The ability to leverage extensive data is what makes large language models (LLMs) possible. The transformer represents the most significant advance in neural network design since AlexNet emerged in 2012, incorporating essential characteristics: a vast scale, the capacity to speed up training time through processing data in extensive, parallel sections, and an exceptionally sophisticated attention mechanism.

(Shortform note: This textbook on the transformer architecture supports Li’s claim that scale and information are crucial for pushing AI forward. It explains that increasing the scale and information in training unlocks new abilities in AI models. The book also supports Li’s assertion that the transformer is the most significant advance in neural network design since AlexNet, noting that most recent state-of-the-art natural language processing systems rely on the transformer architecture.)

New "multimodal" networks, which learn from diverse media including text, photos, audio, recorded speech, and video, are now acquiring the ability to produce such content too. In just a decade, algorithms have gone from having difficulty identifying what photos contain, to outperforming humans at it, to now autonomously generating completely new images that are often unsettlingly realistic and detailed. By contrast, AlexNet was initially launched with a 60-million-parameter network—sufficient to partially comprehend the ImageNet dataset—whereas transformers large enough for training on a variety of media like text, photos, and videos now have parameter counts in the hundreds of billions.

Critics of Large Language Models

Not all researchers are as enthusiastic about these developments as Li. Computational linguist Emily Bender, co-author of the influential paper “On the Dangers of Stochastic Parrots,” has been a vocal critic of the rush to build ever-larger language models. Bender and her co-authors argue that the drive to build ever larger text-generation systems trained on vast, uncurated datasets “comes with substantial environmental and financial costs, exacerbates the reproduction and amplification of harmful biases and stereotypes present in the data, and diverts attention and resources away from alternative research directions that could produce more just, accountable, and transparent language technologies.”

The Deep Learning Revolution

Li argues that deep learning has transformed AI and society. It has revolutionized businesses, led to billions of dollars in investments, and left people racing to comprehend a technology that appeared to suddenly go from a specialized academic field to a globally transformative force. However, the excitement within the tech industry has been dampened by concerns from the media, advocacy organizations, and governments. Issues like algorithmic bias, concerns over large-scale job loss, and disquieting ideas about surveillance have become prominent in the media, souring the public perception of AI. Li believes that AI's real effect will be determined by the motivations behind its development. By expanding our view of AI to purposefully encompass a beneficial effect on people and society, AI can transform the world positively.

AI’s Societal Effects Are Driven by Power Dynamics

Many thinkers disagree with Li’s assertion that AI’s real effect will be determined by the motivations behind its development. In Atlas of AI, Kate Crawford argues that contemporary AI is neither artificial nor intelligent but an extractive industry that relies on intensive use of energy and natural resources, low-paid and often invisible human labor, and massive data collection. She contends that AI systems are deeply intertwined with existing power structures and perpetuate inequalities through their material infrastructures and the social relations they reinforce. Crawford suggests that the societal effects of AI are driven by these material realities and entrenched power dynamics, rather than by the intentions of individual designers.

Li also discusses enabling technologies and landmark achievements.

Enabling Technologies

Li explains that GPUs enabled deeper neural networks to be developed. Because they were affordable, researchers could create larger networks than previously possible. However, merely increasing the number of layers didn't provide a cure-all. While networks with increased depth initially had more accurate results, they soon hit a plateau. Too many layers caused the signal to degrade as it traveled through the network, stopping training and making the system unusable. The status quo needed to evolve in both size and innovation.

(Shortform note: Modern deep-learning theory explains why this happens. In a standard multilayer network, the signal is repeatedly multiplied by weight matrices and nonlinearities. This causes the signal to either shrink toward zero (vanishing gradients) or grow uncontrollably (exploding gradients) as it passes through many layers. Empirically, this produces exactly the training plateaus and failures Li describes. The deeper the network, the more severe the problem becomes, which is why simply adding layers doesn't work.)

In 2015, an innovation was introduced: ResNet, or the Deep Residual Network. The enormous network was comprised of 152 layers, yet it incorporated an innovative design that allowed certain layers to be skipped during training. This let distinct images impact smaller subregions inside the network. Once fully trained, the system would ultimately utilize its entire depth, but each training example wasn't required to extend throughout. The result was a greater number of layers, which boosted performance and allowed the network to absorb more data, but it also allowed the signals to flow freely without degradation. ResNet’s error percentage of 4.5% beat the estimated human accuracy rate of 5.1%.

Highway Networks

Before the Deep Residual Network, Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber introduced “Highway Networks,” which used skip connections to train very deep neural networks. The authors found that adding gated skip connections allowed information to flow more easily through the network, making it possible to train networks with hundreds of layers. This approach addressed the problem of vanishing gradients, where the signal weakens as it passes through many layers. By providing shortcut paths for the signal, Highway Networks made it easier to train deep networks and improved their performance on complex tasks.

Landmark Achievements

Li notes that the ImageNet Challenge in 2012 marked a pivotal moment for AI and machine learning. The AlexNet team showed the potential of extensive datasets, high-speed graphics processing units, and neural networks with multiple layers. Their algorithm was inspired by the way mammalian vision works, with every network layer integrating more details until the object comes into view. The network could autonomously refine its sensitivities based on the training dataset. This resulted in a system able to recognize objects with accuracy comparable to human vision.

(Shortform note: While AlexNet’s performance in the 2012 ImageNet Challenge was impressive, it didn’t quite match human vision. According to Olga Russakovsky et al., the best-performing convolutional neural network in the 2012 challenge had a top-5 error rate of about 15%. In contrast, human annotators had a top-5 error rate of roughly 5%. This suggests that, at least in 2012, human recognition accuracy on ImageNet was still significantly higher than that of the leading automatic system.)

Responsible AI: Ethics, Governance, and Future Directions

Li argues that artificial intelligence's rapid development has led to ethical and governance challenges, including biased algorithms, privacy concerns, and the need for informed consent in data collection. The public is increasingly aware of these issues, and media outlets are starting to cover them more. AI's impact will depend on the motivations of those developing it. If the goal is to benefit people and society, AI can be a positive influence.

The Origins of Data Privacy Concerns

The current debates over biased algorithms, privacy, and informed consent in data collection have roots in earlier battles over data privacy. In the 1970s, the rise of computerized record-keeping sparked fears of a “dossier society” where personal information could be easily accessed and misused. This led to the creation of data-protection laws in Europe and the US, which established principles like informed consent and the right to access personal data. These early debates set the stage for today’s discussions about how to balance technological innovation with individual rights and societal well-being.

Li also discusses guiding principles and mechanisms for building and deploying responsible AI.

Guiding Principles for Responsible AI

Li believes that developing AI should prioritize human well-being. She argues that AI ought to be created to boost human abilities, not vie with them. The aim should be to better human lives, rather than simply improving process efficiency. This requires a holistic understanding of real life and a commitment to valuing the shared dignity of everyone worldwide.

(Shortform note: In Design Justice, Sasha Costanza-Chock argues that the people most affected by a technology should be the ones to decide whether it’s used. They suggest that when a new technology is introduced in your workplace, school, or community, you should gather a small group of people who are most affected by it, especially those with the least power. Document the harms or benefits you experience, and use that record to demand changes or reject the technology if it undermines people’s dignity.)

Li also discusses the ethical basis along with technical and procedural safeguards.

Ethical Foundations of Responsible AI

Li argues that artificial intelligence must uphold human decency. It should focus on people and be equally dedicated to scientific and human interests. This means working together and being respectful, but also willing to face reality. Ultimately, AI should uphold the integrity of the global community.

(Shortform note: In 2021, UNESCO published a list of principles for the ethical use of artificial intelligence. These principles were negotiated by 193 countries and are meant to steer the technology toward the shared goals of the global community. These principles include respecting human rights, promoting diversity, and ensuring that AI is used for the benefit of all people.)

Li also believes that fairness and representation must guide AI creation. While AI can be beneficial, it can also cause harm. The tech industry’s lack of diversity causes biased algorithms that work poorly for people who aren't white or male. The internet often portrays daily life primarily from a male perspective and a viewpoint centered on whiteness and the West, which makes it difficult for technology to understand other people. This problem is worsened by algorithms that aren't adequately tested and decisions that are dubious. Li argues that the lack of fairness and representation in AI is an issue that needs resolution.

Algorithmic Harms Are a Product of Society

Legal scholar Andrew D. Selbst and his co-authors argue that the harmful behavior of algorithms is shaped by the broader sociotechnical institutions and power structures surrounding a system, rather than by the properties of the algorithm itself. They argue that algorithmic harms are often the result of complex interactions between technology, social norms, and institutional practices. For example, an algorithm designed to predict criminal recidivism may produce biased outcomes not because of flaws in the algorithm itself, but because it relies on historical data that reflects systemic biases in policing and sentencing.

Technical and Procedural Safeguards

Li explains that ethical and privacy concerns are crucial in studying AI. Surveillance and the collection of substantial amounts of data may raise privacy issues. In healthcare, for example, the use of AI-powered sensors to monitor patients and staff may be viewed as invasive and dehumanizing. The use of artificial intelligence in employee monitoring, facial recognition, and social media has also raised concerns about privacy, bias, and the potential for misuse. The opacity of AI systems and the potential for discrimination in decision-making algorithms are other major ethical concerns. The rapid pace of AI development and deployment by large corporations has outpaced the ability of regulators and researchers to address these issues, leading to an increasing feeling of unease and mistrust among the public.

The Dangers of Opaque AI Systems

The opacity of AI systems and decision-making algorithms can lead to discrimination and mistrust because it makes it difficult to understand how decisions are made and to identify and address potential biases. For example, if an AI system is used to make hiring decisions and it is opaque, it may be difficult to determine whether the system is discriminating against certain groups of people. This lack of transparency can erode trust in the system and make it difficult to hold the system accountable for its decisions. Additionally, if an AI system is used to make decisions that have a significant impact on people's lives, such as in the criminal justice system or in healthcare, the lack of transparency can have serious consequences.

Mechanisms for Building and Deploying Responsible AI

Implementing Responsible AI Practices

Li argues that implementing responsible AI requires collaboration across diverse fields. AI isn't an independent science. The highest potential for AI is its integration with additional domains and the expertise of other professionals. The greatest work happens in communal scientific areas, with international partnerships that fearlessly traverse borders.

(Shortform note: To implement responsible AI, consider creating a dedicated “responsible AI team-science unit” within your organization. In Enhancing the Effectiveness of Team Science, Nancy J. Cooke and Margaret L. Hilton argue that complex, collaborative research is most successful when organizations move beyond informal cooperation and create formal team-science structures.)

Collaborative Ecosystems for Responsible AI

Li explains that the Stanford Institute for Human-Centered Artificial Intelligence (HAI) promotes collaboration across various disciplines to create human-focused AI. It brings together experts from different fields, including law, political science, humanities, neuroscience, and economics, to work alongside AI students and researchers. This collaboration helps them understand how AI affects employment, wealth, and power dynamics in society.

Form an AI Social-Impact Council

To mirror Stanford HAI’s cross-disciplinary approach, consider forming a standing AI social-impact council within your organization. This council should include representatives from various departments—such as HR, legal, finance, and operations—to ensure diverse perspectives. Give this council veto power over any AI system that could impact employment, wealth distribution, or power structures within your organization. This ensures that ethical considerations are integrated into decision-making processes from the outset.