AI Safety mitigations: Lightning talks

Join us at AI Safety Awareness Group Boston for an insightful session to catch up on the state of AI risk mitigations in a 5-6min lightning talk (e.g. Pecha Kucha) format! If you want to discover like-minded experts, threat models for AI-based systems, or ways to help public awareness, come connect with the Boston AI safety community!

We invite all participants to spend between 15-45min to summarize a few key points that you've learned, as a 5-6min presentation. You can pick any particular facet of the broader AI Safety field. Given the short time to speak (and prepare), this presentation format intentionally makes it best to pick only one narrow topic and convey your points mostly orally, using very light slides (showing only a few words or 1-2 visuals).

Some ideas may be:
- What lessons did tech companies learn in ML safety in the 2010's?
- What are the most significant safety wins of the last 5 years?
- Is there a hierarchy of alignment challenges?
- Are there big safety differences between foundation models?
- Does it matter where a model was built?
- Should we trust models built for "constitutional" safety?
- In 2025, is there a viable adoption of governance methods?
- Is mechanistic interpretability research limited to "measuring" models?
- Has reinforcement learning brought any safety benefit / to what limits
- What vulnerabilities are common in large language models? Covering the impact of prompt injection, jailbreaks, data leakage etc.
- What are the minimum practices to avoid low-cost abuse?
- How should tech companies compare the criticality and efficacy of risk mitigations (e.g. open-source, home-grown, and commercial solutions)
- What are critical security practices when developing with AI?
- Are access controls even relevant in AI applications?
- Where does privacy break down in AI agents?
- How are threat actors capitalizing on AI adoption?
- How has our economy and society already changed / what changes are widely expected to happen in 6, 12, 18 months?
- Will any paradigms or architectures support positive safety outcomes, or what risks are still unsolved?

If you need any materials to support your topic, check out our recommended readings/courses/resources on AI risks and mitigations.

Code of conduct:
Food, drinks, pets, and weapons are strictly prohibited by the Boston Public Library.
This is a forum for open discussion between participants, intentionally open to new ideas. Disparaging or harassing other participants is unacceptable. We want AI Safety Awareness to be a safe and productive environment for everyone. It’s not about what you look like: what counts is what’s in your mind and how you present yourself.
We do not condone harassment against any participant, for any reason. Harassment includes deliberate intimidation and targeting individuals in a manner that makes them feel uncomfortable, unwelcome, or afraid.
Participants asked to stop any harassing behavior are expected to comply immediately. We reserve the right to respond to harassment in the manner we deem appropriate, including but not limited to expulsion without refund and referral to the relevant authorities.