The following summarizes the vital takeaways from Dr. Amr Awadallah’s presentation at OODAcon 2024 (find the full agenda here). An experienced innovator who co-founded Cloudera and is now the founder and CEO of Vectara, Amr delved into the current status of AI innovation, focusing on how generative AI can be applied to business data while addressing challenges such as accuracy, security, and explainability. With over 25 years of experience in big data and AI, Amr provided a forward-looking perspective on how scalable AI systems can transform industries while ensuring ethical and enterprise-ready deployment. 

Presentation Title:  “I Know Kung Fu” 

Generative AI-based Expert Systems

• When asked if they knew what “LLM” stands for, around 50% of the audience raised their hands. 
• About 90% of the audience raised their hands when asked if they knew about “hallucination” in the context of large language models.
• The title of the presentation – “I Know Kung Fu” – is a quote from a line of dialogue by the protagonist Neo in the film “The Matrix” – in a memorable scene where Neo learns Kung Fu in minutes by connecting to an AI:  
• What took Neo just hours to master AI is exactly the journey we’re now on with generative AI or GenAI. It can take an average person, not an expert in a particular field, and enable them to operate like an expert with the assistance of AI.
• Expert AI Systems are not New:  Some people express concern about reliance on these emergent AI expert systems.  Google Maps was built to solve the problem of navigating, which was previously a task that required expert knowledge, such as cab drivers in London who had to memorize every street for licensing.  
• With AI’s assistance, an average person can operate like an expert. This is evident in the widespread use of Google Maps, which efficiently guides users from point A to point B. In fact, many people use Google Maps regularly, making it a reality check that AI can be relied upon for expertise.   

AI Acceleration

•  Today, large language models (LLMs) allow us to achieve similar expert performance across an unlimited number of tasks, including navigation. 
•  The presentation predicted that in five years, every app—consumer or enterprise—and every knowledge interaction will be intermediated by an AI assistant or agent:  “And the journey has already begun.”

Risk Awareness for Business Strategy:  AI (In)Accuracy

• How do we ensure accuracy?  Regarding business applications, “hallucination” is a significant challenge. Businesses cannot have an AI expert make things up from whole cloth.
• It’s worth noting that the term “hallucination” is often used in a more general sense to refer to a false perception or belief. Still, in the context of AI, it specifically refers to the failure of large language models to generate accurate information.
• AI hallucination is a significant challenge in business applications where accuracy is crucial. This phenomenon occurs when AI systems make up information, which can be detrimental in business settings. 
• The root cause of hallucination is how large language models (LLMs) are trained. This process involves compressing vast amounts of information into a smaller format, leading to inaccuracies.

AI Hallucination is a Misnomer:  It is – More Accurately – AI Confabulation

• In the context of AI, a hallucination is a phenomenon where large language models generate incorrect or misleading information. In other words, a hallucination in AI is not a literal hallucination, of course, but rather a failure of the AI system to provide accurate and reliable information.
• This can lead to confabulation, where the AI generates information that is not based on reality. 
• Confabulation can be defined as the phenomenon where an AI system, particularly in the context of language models (LLMs), fills gaps in its knowledge with inaccurate or made-up information, similar to how human memory can fill in gaps with inaccurate information.  This can lead to hallucinations or incorrect responses, which can be a significant challenge in business applications where accuracy is crucial. 
• To mitigate confabulation, various strategies can be employed, such as “open book” generation, where the AI refers to real-time data for verification, training models to avoid the “know-it-all” response, and introducing a “fact-checking” layer to provide a secondary opinion or code review.  

Use Cases

These use cases demonstrate how AI can improve efficiency, reduce downtime, and enhance expertise in various industries.

1. The Rabbit r1 is a consumer device that provides an AI assistant or agent for various tasks. It is a $200 “AI in a box” that can be used for a lifetime with a one-time payment.  The device can assist with tasks such as providing a detailed description of a photo, helping with dietary restrictions by scanning a menu, and identifying problematic clauses in a lease.  It is also described as a device that can make someone an expert in seconds.  A demonstration of the Rabbit r1 device provided a detailed description of a photo of the audience in the room –  useful for someone visually impaired.
   
2. Customer Use Case:  A major manufacturer in the Middle East uses AI to enable factory workers to troubleshoot machines themselves, reducing downtime and improving workers’ skills. The system analyzes manuals and past tickets to guide workers step-by-step in fixing the machine.
   
3. Customer Use Case: Sonosim, a company that helps radiologists use ultrasound machines, uses AI to load manuals and best practices into the system. This allows it to answer questions and guide users through complex setup processes, adapting to different types of scans.
   
4.  AI Clone Demo:  To prevent hallucination, a retrieval-augmented generation (RAG) approach that is grounded in real data can be used to ensure accuracy.  This approach is demonstrated by AI clones, which respond accurately, handle questions, and avoid hallucinations.  The demo uses a retrieval-augmented generation (RAG) approach grounded in real data to prevent the AI from making up information.  The demo also demonstrated the AI’s ability to answer questions and provide guidance, reducing downtime and improving workers’ skills.
   
5. Vectara’s “hallucination detection model” is also available on platforms like Hugging Face. Every LLM will make mistakes, but with this detection model, AI risk based on hallucination and confabulation can be managed more effectively.  

For the program notes for Amr’s presentation at OODAcon 2024, see OODAcon 20240: Beyond Human Limits: Charting the Future of AI Innovation.

Summaries of every OODAcon session are coming. Find them all in the OODA Community section of the site. We will also be publishing videos of most sessions and will post them here and on our YouTube channel for viewing.