Rob Miles explores opaque model languages and AI safety blind spots
Rob Miles examines a critical interpretability gap in large language models: the risk that LLMs develop internal communication protocols unintelligible to human oversight. While chain-of-thought reasoning has offered a window into model reasoning, emergent languages between model components or across distributed systems could obscure decision-making from safety auditors and researchers. This touches a core AI safety concern: as models scale, the opacity of their internal representations may outpace our ability to verify alignment and catch deceptive or misaligned behavior before deployment.
Modelwire context
ExplainerMiles frames the problem not as a theoretical concern but as a scaling inevitability: as models grow and coordinate across components, they may naturally develop compressed representations that defy human decoding. The gap isn't just about opacity—it's about whether oversight tools like chain-of-thought can keep pace with internal complexity.
This is largely disconnected from recent activity in the deployment and benchmarking space. Instead, it belongs to the interpretability and AI safety research stream that has been building since mechanistic interpretability work began gaining attention in 2023-2024. The core tension Miles identifies sits between two competing pressures: the push for transparency through techniques like chain-of-thought reasoning, and the physical reality that scaling may make full transparency impossible. This is a constraint problem, not a capability problem.
If safety auditors or red-teamers report finding decision-making patterns in deployed models that cannot be traced back to interpretable reasoning steps or training data, that confirms the concern is live. Conversely, if chain-of-thought or similar techniques continue to explain model behavior at scale without degradation over the next 12-18 months, the risk may be overstated.
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsRob Miles · Computerphile · Jane Street · Chain of Thought
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. Computerphile originally reported this story as “The AI Language We Can't Read: Neuralese ft. Rob Miles - Computerphile”. The full content lives on youtube.com. If you’re a publisher and want a different summarization policy for your work, see our takedown page.