AI expert Michael Littman explains what AI is and isn't, how it works, and discusses concerns about job displacement and malicious use. They delve into the rise of AI tools in image generation, explore biases in language models, and discuss safeguards against misuse of AI. Listeners can also learn how to get started with AI using publicly available tools and chat bots integrated with search engines.
AI models have limitations and are not yet capable of human-level intelligence.
Quality data is crucial for training AI models and ensuring accurate results.
AI tools can enhance human capabilities, but they require human supervision and won't completely replace jobs.
Deep dives
The Rise of AI: From Image Recognition to Language Generation
Over the past few years, there has been an explosion of AI tools, driven by advancements in deep learning and neural networks. It all started with the breakthrough in image recognition, where deep networks showed unprecedented accuracy. This led to the development of large language models, such as chat GPT, which are trained to predict the next word in a sequence. These models have proven to be incredibly versatile, not only in language tasks but also in generating images, sounds, and even code. However, it's essential to note that these models have limitations and are not yet capable of complete human-level intelligence.
The Challenges of AI: Quality of Data and Training
One of the main challenges in AI is the quality of data used for training the models. The data needs to be carefully curated to ensure accurate and reliable results. While these models can perform impressive tasks, they are not infallible and can generate incorrect or nonsensical outputs. Debugging the models can be challenging, as it's often hard to pinpoint whether the issue lies in the data, the training process, or the model itself. Another challenge is keeping these models up to date with new information, as they are fixed in time and cannot easily incorporate new facts or knowledge.
Language Models and Translation
Language models, like chat GPT, are not language-specific and can be applied to various languages. While they may be trained on English-centric data, they have shown the ability to translate and converse in other languages. However, their performance may vary based on the amount and quality of available training data. Efforts are being made to improve language model coverage in underrepresented languages and cultures. The emergence of multi-language models, like the Bloom project, aims to provide broader language representation.
The Impact of AI on Jobs and Overblown Concerns
The impact of AI on the job market is a topic of debate. While AI tools have the potential to automate certain tasks and replace some jobs, the idea that they will completely replace human workers is overblown. These tools are still far from being sufficiently reliable and independent. They require human supervision and are best seen as tools to enhance human capabilities rather than replace them. It's more likely that AI will shift the job landscape, creating new roles and requiring specialized skills. It's crucial to focus on preparing the workforce for these changes and ensuring that AI is used responsibly and ethically.
The potential of machine learning tools
Machine learning tools can bridge the gap between what computers can do and what people want them to do. They can empower people by enhancing their capabilities, although it may lead to more concentrated jobs. It is important to navigate this space carefully, considering policy, education, and staying informed about upcoming shifts. Skill sets like prompt engineering are already emerging as a new requirement to effectively utilize these language models.
Safeguards and challenges of language models
Language models, like chat GPT, can provide amazing power when asked the right questions. However, they require caution as they can be manipulated for malicious purposes. Safeguards are being implemented to restrict output on topics like bomb-making. There are ongoing efforts to watermark and detect AI-generated content, but it is a cat-and-mouse game. Validating valid content through digital signatures may provide more security. It is crucial to explore protections and transparency to counter disinformation and misuse of language models.
Unless you've been living under a rock, you've seen several news stories about AI, machine learning and so-called Large Language Models. While tools like ChatGPT hold a lot of promise, many are deeply concerned about AI replacing jobs, generating potent malware, and being used in phishing and disinformation campaigns. Today I will ask AI expert Michael Littman to explain clearly what AI is and what it isn't, how the technology actually works, and what we should and maybe shouldn't be worried about.
Michael Littman is a computer science professor at Brown University who has won several prestigious teaching awards while studying machine learning and the implications of artificial intelligence. He serves as division director for Information and Intelligent Systems at the National Science Foundation and is also a Fellow of the Association for the Advancement of Artificial Intelligence and the Association for Computing Machinery.
Interview Notes
Gathering Strength, Gathering Storms: The One Hundred Year Study on Artificial Intelligence https://ai100.stanford.edu/gathering-strength-gathering-storms-one-hundred-year-study-artificial-intelligence-ai100-2021-study
Code to Joy book preorder: https://www.amazon.com/Code-Joy-Everyone-Should-Programming/dp/0262546396/
Michael Littman’s website: https://www.littmania.com/
Gandalf AI challenge: https://gandalf.lakera.ai/
ChatGPT: https://openai.com/blog/chatgpt
Stable Diffusion: https://stability.ai/stablediffusion
Canva Image Generator online: https://www.canva.com/ai-image-generator/
Paperclip Maximizer: https://en.wikipedia.org/wiki/Instrumental_convergence#Paperclip_maximizer
Further Info
Nominate someone for a challenge coin: https://fdsd.me/quest
Get your Firewalls Don’t Stop Dragons Merch! https://fdsd.me/merch
Give the gift of privacy and security: https://fdsd.me/coupons
Send me your questions! https://fdsd.me/qna
Support our mission! https://fdsd.me/support
Subscribe to the newsletter: https://fdsd.me/newsletter
Check out my book, Firewalls Don’t Stop Dragons: https://fdsd.me/book
Generate secure passphrases! https://d20key.com/#/
Table of Contents
Use these timestamps to jump to a particular section of the show.
0:00:56: Dragon coin promo update
0:01:51: Interview preview
0:03:15: What is Artificial Intelligence, really?
0:05:36: Is it a mistake to anthropomorphize AI?
0:08:50: What is AI versus machine learning?
0:11:59: How does AI differ from normal computer code?
0:14:49: What is a large language model or LLM?
0:18:45: What does it take to create an LLM?
0:22:04: Why are these AI models limited to certain points in time?
0:26:46: How are these chat bots leading people to believe they're sentient?
0:28:54: What was behind the AI explosion in late 2022?
0:32:29: How to AI systems generate images from text prompts?
0:35:36: How are AI systems affected by their training data?
0:40:24: Which concerns about AI are justified and which are overblown?
0:44:55: What sorts of jobs may be impacted by AI?
0:47:15: Is there an art to creating AI prompts?
0:48:43: Can you trick AI systems?
0:51:42: How do we detect AI output? How should we restrict this technology?
0:56:19: How can we try out these AI systems to learn more?
0:59:26: What's the next big thing in AI?
1:02:12: Why should people learn to do a little coding?
1:05:27: Wrap-up
1:07:01: Gandalf AI game
1:08:19: Upcoming interviews
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
