Parties, dinners, work events and family gatherings–it’s that time of the year when people get together and catch-up on each other’s lives and current events. AI is all over the news, and it’s likely to come up in conversations between the clinking of glasses and the passing of appetizers.
Want to sound like an in-the-know AI expert? Having a few advanced AI terms in your back pocket can turn your small talk into a big hit. In this spirit, here are three concepts to get you started.
1. Reversal Curse
The Reversal Curse is a limitation in large language models (LLMs) and their assistants, such as ChatGPT, which refers to their inability to reverse causal statements they are trained on. For example, if you ask ChatGPT, “Who is Beyonce’s mother?” and it correctly responds with “Tina Knowles,” it may struggle to answer the reversed question, “Who is Tina Knowles’ daughter?” with Beyonce.
This reveals a fundamental limitation of LLMs, showing that they rely more on statistical patterns than on developing a causal understanding of information. This limitation is important because it shows that while AI can process and generate complex information, its understanding is not always deep or flexible and that it still has significant areas to develop, especially in understanding and reasoning in ways that humans naturally do.
LLMs are dream machines. They are trained on vast amounts of data and remember the essence of this data, then metaphorically “dream” up outputs in text, images, and sound without those outputs being explicitly in its data sets.
2. What’s Your P(Doom)?
(pronounced “p-of-doom”)
In the Harry Potter series, Hogwarts employs the Sorting Hat to assign new students to the appropriate house based on their traits and values. People working in AI have our own version of a Sorting Hat, and we call it P(Doom).
P(Doom) is a metric rooted in probability theory and gauges the likelihood of an event occurring. In this case, it’s a spectrum with AI doomers, who foresee potential AI catastrophes, at one end, and AI enthusiasts, who focus on its positive potential, at the other.
A conversation opener could be asking someone about their P(Doom) percentage. A higher percentage indicates greater perceived risk, while a lower percentage suggests more optimism about AI’s positive potential.
For example, a P(Doom) of 20% suggests your conversation partner perceives a relatively low risk of AI causing catastrophic outcomes.
3. AI Dreaming
LLMs are dream machines. They are trained on vast amounts of data and remember the essence of this data, then metaphorically “dream” up outputs in text, images, and sound without those outputs being explicitly in its data sets. Essentially, we guide these dreams with prompts, but we can’t always be sure if the response is correct or whether it’s a hallucination—an incorrect answer.
Both an AI’s dreams and a human’s dreams potentially combine familiar elements in unexpected ways. For instance, a person might dream of a place they know but with unusual details or events. Similarly, an AI, when given a prompt about a familiar subject, might generate a story or image that blends known elements with creative and even nonsensical twists.
This term underscores the idea that, while AI can generate creative content, its process can be as unpredictable and abstract as human dreams
Up Your AI Conversation Game
Reversal Curse, P(Doom) and AI Dreaming are more than catchy phrases; they’re insights into the complex world of AI. As you mingle at your next social event, these terms can make you sound like an AI insider, but they also invite others into a conversation about AI’s role and its impact in our world. While technology can divide us, it can also be an opportunity to bring us closer together in a shared dialogue.
By Natalie Diggins, Technologist and Founder of TheArts.ai
Proofread using ChatGPT-4.