bio

Adam Jones is working to make the transition to advanced AI systems go well. He's concerned that powerful AI systems could pose serious challenges in the next couple of years: If many people control powerful AI this invites potentially catastrophic misuse (e.g. bioterrorism) - but if few people control them we might end up in a stable totalitarian state. And that's assuming humans do manage to control them...

At Anthropic, he works on reinforcement learning (RL) infrastructure. He's particularly interested in using RL to make models better at safety-relevant capabilities - for example, making models good at alignment research, or making them good for defensive technology use cases like pandemic prevention, detection and response. He also works on making AI agents safe and useful via his work on the Model Context Protocol.

Previously, he led AI safety talent programs at BlueDot Impact, including the large-scale AI Safety Fundamentals courses. He also advised the UK Government's Department for Science, Innovation and Technology (DSIT) on AI safety policy.

Outside of work, he enjoys making things! This includes writing blogs, and building popular open-source tools: his AWS Email Simulator has over 1 million downloads, his YouTube thumbnail-hiding browser extension serves 50,000+ weekly users, and his Airtable MCP server has spurred 125+ forks. Finally, he contributes media to Wikimedia Commons, with his work appearing in textbooks, academic papers, and YouTube videos with millions of views.

When not making new things, he enjoys co-operative board games, video games, sunny weather, and playing capture the flag on Hampstead Heath.

(there's also a bunch more on my GitHub)

working with me

The following make up a How to work with me manual / User Guide, which coworkers or AI virtual collaborators might find helpful:

• Preferences for communicating with me
• How I give reviews