Connor Leahy: We lack understanding of intelligence and neural networks, the unpredictability of AI could lead to loss of control, and GPT models have revolutionized AI capabilities | The Peter McCormack Show

AI's unpredictable growth could lead to a future where humans lose control over advanced systems.

Listen on The Peter McCormack Show

Share

Add us on Google by Editorial Team Mar. 25, 2026

Key Takeaways

• We lack a comprehensive understanding of how intelligence and neural networks function.
• The unpredictability of AI could lead to a future where humans lose control over AI systems.
• Neural networks differ significantly from traditional programming, operating more like a growth process than a coded instruction set.
• GPT models have revolutionized AI by learning complex tasks independently as they scale.
• The release of GPT-2 was a landmark moment, marking a significant leap in AI capabilities.
• The transformer architecture is foundational to modern AI, underpinning many current applications.
• Despite advancements, the inner workings of neural networks remain largely mysterious.
• Neural networks process vast amounts of data through complex mathematical operations, which are not fully understood.
• AI models like ChatGPT use extensive data and user information to generate contextually relevant responses.
• Understanding intelligence, whether in humans or AI, remains a significant challenge.
• The rapid evolution of AI technologies raises important questions about future human-AI dynamics.
• The complexity and opacity of neural networks highlight the challenges in AI research and development.
• The psychological impacts of AI on users are an emerging area of concern and study.
• AI’s ability to provide tailored responses showcases the sophistication of current systems.
• The potential shift in power dynamics between humans and AI underscores the need for careful oversight.

Guest intro

Connor Leahy is CEO and co-founder of Conjecture, an AI safety research company based in London. He previously co-founded EleutherAI, where he helped develop GPT-J and GPT-NeoX, among the largest open-source large language models at the time. Leahy advocates for pausing development of advanced AI systems to address unsolved alignment risks.

The mystery of intelligence and neural networks

• We do not understand intelligence or how neural networks work, despite having built them.

  — Connor Leahy

• The complexity of intelligence in both humans and AI remains largely unexplored.
• Understanding neural networks is crucial for advancing AI research and development.

• It’s very important to understand is that we do not understand intelligence we don’t know how the brain works.

  — Connor Leahy

• The gap in understanding highlights the uncertainty surrounding AI technologies.

• We sure as hell don’t know how these neural networks work either.

  — Connor Leahy

• The unpredictability of neural networks poses challenges for AI researchers.
• Despite advancements, the fundamental mechanisms of neural networks are still unknown.

The unpredictable future of AI control

• We may wake up one day and find that we are no longer in control of AI.

  — Connor Leahy

• The potential loss of control over AI systems raises ethical and practical concerns.
• Future dynamics between humans and AI could shift dramatically without proper oversight.

• I think this happens before extinction happens like the thing I expect to happen is that one day we wake up and we’re just not in control anymore.

  — Connor Leahy

• The rapid evolution of AI technologies necessitates careful monitoring and regulation.
• Understanding AI’s trajectory is critical for maintaining human oversight.
• The unpredictability of AI systems underscores the need for robust safety measures.
• The potential for AI to operate beyond human control is a significant concern.

The revolutionary impact of neural networks

• Neural networks operate fundamentally differently from traditional programming.

• The way AI worked was the new technique was called deep learning or neural networks.

  — Connor Leahy

• Unlike traditional programming, neural networks grow and learn from data.

• Normal programming you write code… neural networks are very different it’s more like you grow them.

  — Connor Leahy

• The distinction between traditional programming and neural networks is crucial for understanding AI.
• Neural networks represent a paradigm shift in how AI systems are developed.
• The growth process of neural networks allows them to solve complex problems.
• Understanding the function of neural networks is essential for advancing AI technologies.

The transformative power of GPT models

• GPT represents a significant shift in AI capabilities due to its general-purpose nature.

• The crazy thing about it was is that as you fed it more data and as you gave it more computing powers you made the neural network bigger.

  — Connor Leahy

• GPT models learn complex tasks independently as they scale.

• It learned first you know how to spell words then it learned how to do sentences then paragraphs.

  — Connor Leahy

• The release of GPT-2 marked a pivotal moment in AI development.

• For me my like oh shit moment was in 2019 with the release of GPT two.

  — Connor Leahy

• GPT models have revolutionized AI by learning without explicit human instruction.
• The scalability of GPT models highlights their transformative potential.

The foundational role of transformer architecture

• The transformer architecture revolutionized neural networks and is foundational to modern AI applications.

• In 2017 a group at Google discovered a new way to kinda build a neural network called a transformer and it changed everything.

  — Connor Leahy

• The transformer architecture underpins many current AI applications, including image and voice generation.

• All the neural stuff you see today whether it’s AI you know image generation voice generation you know chat GPD all of this is based on what’s called the transformer.

  — Connor Leahy

• The development of the transformer architecture marked a significant advancement in AI technologies.
• Understanding the significance of the transformer architecture is crucial for grasping modern AI evolution.
• The transformative impact of the transformer architecture underscores its importance in AI research.
• Despite its foundational role, the inner workings of transformers remain largely mysterious.

The complexity of neural network operations

• Neural networks operate by processing billions or trillions of numbers through various mathematical operations.

• When you think of a neural network the way you should think of is billions even trillions of numbers.

  — Connor Leahy

• The underlying mechanisms of neural networks remain largely unknown.

• We don’t know what any of this means… we have some guesses at some of it.

  — Connor Leahy

• Understanding the complexity and opacity of neural networks is crucial for AI research.
• The mathematical operations within neural networks highlight the challenges in understanding their function.
• The complexity of neural networks underscores the need for further research and exploration.
• Despite advancements, the inner workings of neural networks remain a mystery.

AI’s ability to generate contextually relevant responses

• AI models like ChatGPT can generate contextually relevant responses by leveraging vast amounts of data.

• I’m blown away by what it can do because it adds in a little bit more than that it adds in what it already knows about me.

  — Connor Leahy

• AI models utilize user-specific information to enhance response accuracy.
• Understanding how AI models process and utilize user data is essential for grasping their capabilities.
• The sophistication of modern AI systems is evident in their ability to provide tailored responses.
• AI’s ability to generate contextually relevant responses showcases its transformative potential.
• The use of extensive data allows AI models to deliver more accurate and relevant responses.
• Understanding the mechanisms behind AI’s response generation is crucial for advancing AI technologies.

The ongoing challenges in understanding intelligence

• We do not fully understand how intelligence works, whether in humans or in neural networks.

• It’s very important to understand is that we do not understand intelligence we don’t know how the brain works.

  — Connor Leahy

• The complexity of intelligence remains a significant challenge for researchers.
• Understanding the current limitations in neuroscience and AI development is crucial for advancing the field.
• The gap in understanding highlights the need for further research and exploration.
• The complexity of intelligence underscores the challenges in AI research and development.
• Despite advancements, the fundamental nature of intelligence remains largely unexplored.
• The ongoing challenges in understanding intelligence highlight the need for continued research and innovation.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.