Will We Be Haunted by a Non-Hallucinatory AI?

Lloyd Watts discusses the significant challenge posed by hallucinations in large language models (LLMs), such as ChatGPT. While these models often generate fluent and useful responses, they occasionally produce incorrect or misleading information, referred to as “hallucinations.” Watts highlights that this problem, acknowledged by major tech companies like Google, remains unsolved despite their advanced efforts. He argues that hallucinations are inherent to the architecture of LLMs, which rely on deep neural networks and transformer technology. This unresolved issue represents a multi-billion or even trillion-dollar problem, impacting the market valuation of tech giants dependent on AI advancements.

Watts takes the audience through the hallucination problem, arguing that hallucinations are inherent to the architecture of LLMs because LLMs do not have an axis of truth in their process. “We need the ability to have it have some source of truth,” says Watts. “Right now all it has is a training set from which it extracts statistics and it learns about likelihood. The most statistically likely next words, that’s what it produces — likely next words, not necessarily true sentences.” Watts proposes that introducing a new attribution network, which would act like a hippocampus for the LLM, would fix this problem.

“To build a truthful and trustworthy system,” says Watts, “you also need memory, confidence, and judgment, and if you add elements into the system that do that, you have a chance of building something that’s truly intelligent.”

Andrew Mayne highlights the rapid progress in AI and the improvements in handling issues like hallucinations in AI models. He emphasizes the practical applications of AI today, such as writing code and enhancing creativity. AI amplifies human potential and has transformative capabilities for the economy and job creation, fostering economic growth and increased demand for both humans and robots — leading to even a future of zero unemployment.

Mayne presents a very optimistic view of an AI future, based on the history of technological advancement. Cameras didn’t replace artists. Phonographs didn’t destroy music. Rather, those inventions expanded art. “AI is a tool. Tools amplify creativity,” argues Mayne. “We can’t predict where it will go, but we can know that opportunities will expand.”

Lloyd Watts is Founder and CEO of Neocortix.

Andrew Mayne, former researcher and prompt engineer at Open AI, is the Founder of Interdimensional AI.

Don’t miss out on COSM 2025, November 19-21 in Scottsdale, Arizona. Visit cosm.tech for more information.