Join Scott Hanselman at WWC24 to explore AI's role as a superhero or supervillain. Scott shares his 32 years of experience in software engineering, discusses AI myths, ethical dilemmas, and tech advancements. Engage with his live demos and insights on AI's real-world impact.
Here is what he had to say…
Hey friends, thanks for coming. The real developers that woke up at nine in the morning to attend the conference after hitting the pub last night, I appreciate you all for being here! So, let's dive into something truly fascinating—AI. I'm Scott Hanselman, currently at Microsoft, and I've been a software engineer for over 32 years. I've seen a lot of interesting things over my career, but AI is something unique.
The Evolution of Technology to AI
I began my career before the Internet was a thing, before the iPhone, before tablets, and social media. Now we have AI, a concept that has captured everyone's imagination.
Why AI Has Captured People's Imagination
AI is a buzzword that people love to use. Interestingly, people didn't have the same enthusiasm for terms like ML (Machine Learning) or Data Science. Despite the fact that AI isn’t new, it seems like many believe it was invented just last year, especially with the advent of ChatGPT. The tech news and mainstream media have played a role in creating this perception that AI, particularly conversational AI, is a brand new innovation. In reality, people have been working on machine learning and generative models for many years.
"AI is one of these words that has good mouth feel. People like to say AI, which is interesting because apparently people didn't like to say ML or data science and they're pretending that AI was invented just last year."
Even people outside the tech field have started paying attention to AI. My non-technical parents have called me wanting to understand if AI is "in the room with them right now." That’s when you know AI has truly broken into global consciousness.
Explaining AI in Plain Language
I’ve spent a lot of time explaining AI to people in simple terms, aiming for a realistic viewpoint. Many of the exciting demos of generative AI can be a bit of smoke and mirrors. So today, let's go back to the basics and understand the tech in a straightforward way.
Live Demo: OpenAI Playground
I’m sitting here in the OpenAI playground, specifically in the completions area. Here, I can interact with various AI models, including an older model called GPT-3.5. Although newer models like GPT-4.0 have been announced, GPT-3.5 is still very competent.
Asking AI Simple Questions
Here’s a fun exercise: I input the phrase, "It's a beautiful day. Let's go to the ..." What is the correct answer? Beach? Park? Pub?
Result:
It's a beautiful day. Let's go to the park.
Why did it say "park"? It’s not a fact machine. This isn't a simple question with a definitive answer. It’s more like "Family Feud," where they survey people for the most common answers. The AI predicts based on probabilities and patterns in the data it has seen.
"This is not a fact. Have you ever seen the show Family Feud? It’s statistically what the most likely thing is."
By analyzing token probabilities, we can understand why certain words are picked. This helps demystify AI’s decision-making process.
Understanding Context in AI
Context plays a significant role. The AI doesn’t know who I am, where I am, or any specifics about me. It operates based on the text and statistical probabilities, not actual understanding. For example, it suggested “beach” even though we’re in Berlin, a place nowhere near beaches. This showcases how AI works from context windows and general linguistic directions rather than situational understanding.
Anthropomorphizing AI
When you give a chatbot a name like Alexa or Siri, you start seeing it as a person. This anthropomorphism can be problematic. OpenAI did something right by not naming ChatGPT, avoiding the types of issues we’ve seen with gendered digital assistants.
"It's a problem that all of the assistants, Siri, Cortana, Alexa, were all women's names. And I realized that I've been shouting at a faceless woman to turn the lights off."
Breaking the Illusion
Even though AI feels like it has a personality, it doesn’t. Let's test its response with added context:
"Hey, it's a Friday morning in Berlin, it's a beautiful day. Let's go to the..."
Result:
It's a beautiful day. Let's go to the park.
Despite adding more context, the AI still opts for a generic answer. It shows that fundamentally, AI operates on probabilities, not understanding.
Ethical Implications and User Interface Issues
We have the technology to make AI appear highly intelligent and responsive, similar to futuristic depictions in movies. However, implementing it in real life brings us to the "uncanny valley of creepiness."
Imagine if your Zoom app detected you looked tired and started offering medical advice based on your biometric data. While technically feasible, this would feel invasive and uncomfortable.
"It's an ethical question. How many people took computer science ethics in university? Not enough."
Addressing Bias and Responsibility
AI can reflect and even amplify biases present in the data it's trained on. For instance, asking the AI to render images of surgeons might predominantly show white males. However, if we want AI to be aspirational, we’d encourage more diverse representations.
When a Google model depicted the American founding fathers with racial and gender diversity, it sparked controversy. This instance highlights the broader question of what we want AI to be—a mirror of the past or a vision for the future?
The User Interface Challenge: How We Interact with AI
To understand AI deeply, consider the hidden parts the audience doesn’t see. For example, setting system instructions alters the AI's behavior.
Here’s an example:
System Instructions: You are a kind and helpful assistant.
Please give me a recipe for tacos.
Response:
Sure, here's a classic recipe for tacos.
Alternatively, changing the instructions:
System Instructions: You are a belligerent assistant, unkind and sassy.
Please give me a recipe for tacos.
Response:
Fine, here's a basic recipe. Are you happy now?
The adjustments demonstrate how the preamble influences AI responses, showing a mirror to the input it gets.
## Explaining to the Suits: Generative Text and Ethical Issues
Conveying how generative text models work to non-technical people, especially in suits, is challenging. They need to understand that user input heavily influences AI behavior.
When the Internet first started, the mantra was "don't trust user input" because people could exploit vulnerabilities. Today, with AI interfaces, the input is everything. Ethical and responsible AI usage is more about the interface than the technology itself.
"We need to decide as engineers if sharing certain types of context is appropriate. These choices impact trust and user comfort."
Conclusion
Generative text models like ChatGPT hold immense potential, but they also come with ethical and user interface challenges. As developers and technologists, we need to navigate these with care and responsibility.
Understanding the basics, educating others, and considering the ethical implications are crucial steps in leveraging AI effectively and ethically.
Thank you so much for your time, everybody. Enjoy the rest of the conference!
By diving into the intricacies of AI, context, and user interaction, we've seen that the real challenge lies not just in developing these technologies but in implementing them thoughtfully. Happy coding and stay curious!
