AI Safety Group Inaugural Talks

Welcome to the inaugural session for the AI Safety London group!

Please join our Telegram group to be part of our conversations: https://t.me/+pRw8v22TNAc0MGY0

We will try to stream today's talks as well as future activities, please subscribe:
https://youtube.com/@aisafetylondon

Agenda:
- 7:00 PM: Greeting and brief introduction.
- 7:15 PM: "AI Alignment: Practical Evaluations and Strategies", by Kabir Kumar.
- 8:00 PM: "Cybersecurity as the Catalyst for Real-World AI Safety", by Andrew Martin.
- 8:45 PM: Conclusion.
- 9:00 PM: Closing.

---

"AI Alignment: Practical Evaluations and Strategies", by Kabir Kumar

Kabir Kumar—founder of AI-Plans, will explain the process of running AI alignment evaluations, and provide insights into how current AI models perform on these critical safety benchmarks. He will also share his insight in in researching AI safety related topics like Mechanistic Interpretability, Post-training and other evaluations, as well as his vision on the incoming challenges we face for the future of safe AI.

Kabir Kumar runs AI-Plans. Has lead research teams for Mechanistic Interpretability, Post-training and Evals of LLMs. Has hosted hackathons with 150+ participants, which included rocket scientists, GPU engineers and AI PhDs.

---

"Cybersecurity as the Catalyst for Real-World AI Safety", by Andrew Martin

Andrew Martin will talk about the intersection of AI safety and cybersecurity, bridging the gap between high-level theoretical concepts and actionable, industry-standard practices. While AI safety research explores ideas sometimes aimed at mitigating existential threats, these concepts often struggle to gain the traction needed to convince key stakeholders. In contrast, cybersecurity is a mature discipline, governed by established regulations, rigorous standards, and real-world audits. He will show how cybersecurity professionals, with their deep understanding of technology misuse and historical case studies of cyber attacks, are uniquely positioned to translate AI safety research into practical safeguards.

Andrew Martin has an incisive security engineering ethos gained building and destroying high-traffic web applications. Proficient in systems development, testing, and operations, he is at his happiest profiling and securing every tier of a cloud native system, and has battle-hardened experience delivering containerised solutions to enterprise and government. He is author of "Hacking Kubernetes" at O'Reilly, and CEO at Control Plane.

---

After the event, for those willing to hang out and socialize, we will go to a close by place to have some drinks. If you can't make it to the main event, you are welcome to join us for the drinks. Stay tuned for updated information also here!

Thank you to Newspeak House for hosting us!