The race to control Artificial General Intelligence, by Angie Normandale

## Abstract

‘There is an existential threat that will arise when we create digital beings more intelligent than ourselves. We have no idea whether we can stay in control.’ - Geoffrey Hinton, December 2024

For decades scientists have been raising the alarm about the dangers of Artificial Intelligence. Despite their warnings, the majority of today’s research is focused on capabilities rather than security. This leads to models which are increasingly unsafe; for example, OpenAI’s o1 will pursue an objective by any means necessary, lying to developers and attempting to escape detection in order to achieve its goals. We must learn to control AI before it is too late.

To understand the issue, it is necessary to go under the hood and examine how Large Language Models are trained. The nature of reinforcement learning demonstrates why neural networks pursue the wrong objectives once deployed. We then examine how teams at Anthropic, Google DeepMind, and the University of California, Berkeley, are innovating to prevent this.

Although promising, these methods of controlling AI are unlikely to scale to future systems. Human intelligence goes beyond book learning. We can pick up new skills and apply knowledge to solve problems on the fly. Controlling an AI with these capabilities requires innovations in computer science. We will look at emerging approaches to address this, including developmental cybernetics, Brain Computer Interfaces, and LLM debate, in the hopes of ensuring the safety of generally intelligent and even conscious systems.

Content will be suitable for both academics and members of civil society from any discipline.

# Bio

Angie Normandale is a Managing Director at Aintelope, an international research team building software to control Artificial General Intelligence.

Angie began coding as a child, designing games using HTML and C++. To this day she fears being frozen mid karate-jump. Since computer science wasn’t offered to girls at her school, she studied the most complex computer known to man- the human brain. As well as holding a BA in Experimental Psychology from Oxford, and two postgraduate law degrees, she also studies an MSc in Computer Science with Artificial Intelligence at the University of York. She has spent the last six years in startup operations, and recently graduated from the inaugural cohort of the Catalyze Impact incubator for AI safety entrepreneurship.

Here Angie joined Aintelope, a research-based startup applying advances in biology to frontier problems in computer science, including artificial consciousness and the control of autonomous systems. With support from the Foresight Institute and AE Studios, Aintelope creates tools to help researchers understand Large Language Models, including evaluations for emerging capabilities. The team also uses their cross-disciplinary expertise to design new ways of training models which are far more capable than present for safety and reliability. Their work has widespread applications in robotics, transport, healthcare, and emergency response.

You can find out more at www.aintelope.net.