Can machines learn how to behave?

🌈 Abstract

The article discusses the topic of AI value alignment - whether and how AIs can be imbued with human values. It explores the debate around whether the latest language models can truly understand concepts, or are merely "babblers" that randomly regurgitate their training data. The article argues that language models are capable of understanding concepts and can be imbued with values through language, which provides a clear route to value alignment. It also discusses the challenges around curating training data to avoid biases and toxicity, as well as the broader implications of highly capable AIs for the labor market and our sense of purpose. The article advocates for a transparent, legible, and controllable approach to defining AI values, rather than having them dictated by a narrow constituency.

🙋 Q&A

[01] The Debate Around AI Understanding

1. What are the two opposing views on whether language models can understand concepts? The article presents two opposing views:

• The skeptical view that language models are mere "babblers" that randomly regurgitate their training data, implying that real AI value alignment is out of reach.
• The view argued for in the article that language models are able to understand concepts, and thus have far greater utility, though with potential harms and risks that must be considered.

2. What are the implications of language models being able to understand concepts? If language models can understand concepts, they will have far greater utility, but this also raises urgent social and policy questions. For example, the labor market, economic model, and sense of purpose may be impacted when many information work jobs can be automated.

3. How does the article characterize the two disconnected camps in the AI ethics debate? The article states that those who are deeply skeptical about what AI can do haven't acknowledged the risk or potential of the emerging generation of general-purpose AI. Meanwhile, the existential risk community has been expansive in articulating potential harms and benefits, but considers AGI to be so distant and mysterious that it will emerge spontaneously in an "intelligence explosion" decades from now.

[02] The Potential for Value Alignment

1. How does the article argue that language models can be imbued with human values? The article argues that since language has mediated the fields of ethics, moral philosophy, law, and religion for thousands of years, by sharing language with AIs, we can share norms and values with them too. There is early evidence that this approach works, and as language-enabled models improve, so too will their ability to behave according to ethical principles.

2. What are some of the challenges and limitations around imbuing AIs with values that the article discusses? The article notes that imbuing AIs with values doesn't guarantee perfect judgment, nor does it address governance questions around who gets to define an AI's values and how much scope there will be for personal or cultural variation. It also doesn't tackle the economic problem of equitably distributing the gains of increasing automation.

3. What is the article's view on how AI values should be defined? The article argues that AI values shouldn't be dictated by a narrow constituency, but should become the legible and auditable "operating handbooks" of tomorrow's AIs, developed through a transparent and collaborative process.

[03] The Limitations of GOFAI and the Rise of Language Models

1. How does the article characterize the history and limitations of GOFAI (Good Old-Fashioned AI)? The article explains that GOFAI, which was based on logic, rules, and explicit programming, made far less headway when it came to using language, forming abstractions and concepts, or making sense of visual and auditory inputs. It notes that GOFAI's belief that reasoning could be universally applied was an ambitious but flawed agenda.

2. How does the article describe the breakthrough of Deep Learning and neural networks? The article states that Deep Learning and neural networks, which learn by example rather than relying on hand-engineered rules, were able to perform tasks like reliably recognizing bicycles for the first time. This offered powerful lessons in knowledge representation, reasoning, and the nature of "truth" that GOFAI had failed to capture.

3. What insights does the article provide about the relationship between language and meaning? The article argues that language offers a succinct way to express policies requiring human-like judgment to interpret and apply, rather than being amenable to formal manipulation like mathematical formulas or logical propositions. It notes that natural language isn't math, and words can't be manipulated like algebraic variables or run like computer code.

[04] The Challenges of Ethical AI

1. How does the article characterize the limitations of Asimov's Laws of Robotics? The article explains that Asimov's Laws, while seeming to provide a formal, rule-based approach to robot ethics, are actually nothing like theorems or laws of physics. They require human-like judgment to interpret and apply, and are subject to debate, cultural variation, and evolution over time.

2. What are the challenges the article identifies with attempts to curate "safe" training data for language models? The article argues that attempts to filter out "toxic" content from training data are misguided, as this hinders the model's ability to recognize toxicity, and disproportionately filters out underrepresented minority voices. It suggests that a more inclusive approach, including difficult content, is necessary.

3. How does the article propose that AI ethics should be defined and implemented? The article advocates for AI values to become the legible and auditable "operating handbooks" of AIs, developed through a transparent and collaborative process, rather than being dictated by a narrow constituency. It suggests using natural language to provide guidance and constraints, rather than attempting to fully program AI behavior.