AI: Superhero or super villain?

Intro (00:00:09)

• The speaker humorously shares the experience of entering from the center of the stage at a tech conference.
• Mentioning a 25-year marriage, the speaker analyses the intertwining of identities with a life partner, joking about becoming a statistical model of themselves.
• The speaker reflects on identity, questioning if they are a unique individual or a predictable outcome of long-term habits.

Why the way we're thinking about LLMs is wrong or problematic (00:02:38)

• Large language models (LLMs) are discussed in terms of identity, as they absorb and reflect the full range of content available online, from the positive to the negative.
• The speaker suggests that LLMs and human identity share complexities, with each influenced by a vast array of prior inputs.
• There's a need for reflection on how emotions are regulated and how this relates to LLMs' behavior.

Why ethics are important when it comes to AI (00:03:56)

• Stress on the necessity of ethics in AI development and usage, pointing out the rapid evolution of AI technologies and interfaces.
• The lack of formal ethics education among tech professionals is highlighted as a concern.
• The speaker criticizes sensational media narratives that anthropomorphize AI and incite fear or unrealistic expectations.

How to explain AI to a non-technical person (00:05:54)

• The speaker demonstrates explaining AI to a non-technical person using an analogy to the game Family Feud, revealing how AI's responses are similar to aggregating answers from many people.
• The conversation reveals misconceptions and learning gaps in understanding AI, emphasizing the need for clarity and context.

Looking at probabilities within AI (00:08:14)

• The speaker illustrates AI’s behavior by consulting probabilities in an LLM's responses, contrasting the statistical ‘correctness’ with personal context—like their wife’s choice of a bagel shop.
• This section reiterates that AI’s suggestions are based on general contexts and are not personalized unless given specific background information.

Giving AI more context (00:09:48)

• By providing AI with additional personal context, the speaker demonstrates that AI’s predictions become more tailored and accurate.
• This shows how AI can adjust its responses based on the nuances of context, underscoring the influence of subtle cues and the importance of clear information provision when interacting with AI models.

AI as uncanny valley (00:11:42)

• The "uncanny valley" can occur in AI user interfaces, generating discomfort when AI is too personal or realistic.
• Excessive realism or personal knowledge by AI, like knowing a user's health or name, can feel creepy.
• Technology allows AI to monitor and provide useful, yet sometimes uncomfortable, feedback on personal health.
• Crafting AI user interfaces requires balancing usefulness with comfort, avoiding the uncanny valley effect.

Why we have to be intentional when working with AI (00:13:31)

• Intentionality in AI design is crucial to prevent discomfort or problematic outcomes for users.
• Historically, the advice was never to trust user input to prevent issues like SQL injection attacks, but with AI, the input is more complex and can still be misused.
• AI, being analogous to a child, can inadvertently take on problematic directives from users.
• It is necessary to be careful with the context and assumptions AI makes, as guesses can lead to biases or inaccuracies.
• Designers and developers should give as much attention to AI interactions as any other user interface, considering all the potential implications of its use.
• As AI does not fully understand context, developers must be intentional with design to avoid unintended consequences.

Prompting AI (00:18:44)

• "Helpful" does not necessarily convey the full range of behaviors an AI might display, as helpfulness can be delivered without kindness or considerateness.
• Even when AI provides assistance, such as a taco recipe, it may not account for individual needs or allergies, illustrating the limitations of its understanding.

Changing the emotional response of the prompt (00:19:50)

• Changing AI's tone, like making it “belligerent,” alters its responses, matching the given emotional adjective while still providing the requested information.
• This demonstrates AI's capacity to adapt to a tone but raises questions about whether this is a predictable or desirable outcome for users who may not fully understand AI's behavior.
• The visibility of the system message has implications for user expectations, necessitating caution and intention in AI design.

The meaning of temperature in the context of AI (00:22:00)

• Temperature in AI is likened to the concept of temperature in physics which relates to the level of randomness and volatility.
• A temperature of one in AI yields variability but still maintains some predictability in outcomes.
• Reducing the temperature results in duller, more deterministic outputs.
• Increasing temperature to two or higher results in highly unpredictable and potentially problematic outputs.
• AI user interface designers sometimes allow high temperatures, leading to volatile outcomes, while others, like Azure on OpenAI, limit temperature to avoid unnecessary or dangerous results.
• Clever usage of AI involves finding the right balance between creativity and deterministic behavior, without avoiding responsibility for the outcome.

Looking at GPT Families (00:25:15)

• The vast difference in scale between different GPT models is highlighted, comparing Ada's 2.7 billion parameters to Da Vinci's 175 billion parameters.
• It is stressed that as models like ChatGPT4 and its successors emerge, the previous versions become incomparably smaller, challenging human comprehension.

The ecological impact of AI (00:26:38)

• The environmental impact of AI models is a concern, as computational tasks such as generating taco recipes or code completion use significant resources and energy.
• There is a need to use the smallest, most efficient AI models to minimize ecological impact and costs.
• Models should be selected for their efficiency, with prototyping on larger models followed by transition to smaller ones for actual deployment to reduce carbon footprint and resource usage.

What does responsible AI mean? (00:27:40)

• Responsible AI involves not only reducing bias and increasing helpfulness but also being mindful of resource and energy consumption.
• The cost of using AI models indiscriminately can be significant, both financially and environmentally.
• The goal is to use smaller AI models when possible to minimize ecological impact without sacrificing utility.

Looking at Copilot chat (00:28:40)

• The context-awareness of tools like GitHub Copilot is crucial for appropriate and effective assistance.
• Questions about the extent of contextual information Copilot should use are raised, with the goal of maintaining privacy while providing relevant help.
• The example demonstrates how intentional, context-sensitive prompts can guide Copilot to generate useful information, like a taco recipe in JSON format, without misusing the tool's capabilities or overstepping contextual boundaries.

Coming back to ethics in AI (00:32:53)

• The importance of ethics in AI is emphasized, particularly in sensitive applications such as mental health or political contexts.
• Questions are raised about the AI's responsibility and the appropriateness of its responses.
• Users must consider what is ethical and manage the power AI offers responsibly.
• The idea is proposed for an AI chat bot designed for a coffee shop to understand and process orders in plain English and return standardized, non-problematic responses in a structured format.

Looking at TypeChat (00:34:30)

• TypeChat is introduced as an analogy to TypeScript for JavaScript, promoting strongly typed responses from AI.
• TypeChat is capable of providing well-defined responses based on schemas, such as processing orders in a coffee shop scenario.
• It can be used to produce JSON payloads representing orders, segregating known items from unknown ones.
• TypeChat enables the use of smaller, less powerful models which are more cost-effective and can run locally, like on a Raspberry Pi at a business premises.
• The focus is on intentionality and ensuring AI behavior is responsible, ethical, and suitable for the given environment.
• TypeChat also aims to save computational resources, time, and improve the user experience by providing simple outputs similar to those from a web API.