Munich AI Safety Student Initiative

Advanced AI is arriving fast. The tools to govern it are not.

We unite students, academics and experts to learn and work on the most pressing risks posed by advanced artificial intelligence.

Join our WhatsApp group See what we work on

Two branches Reading groups every 2 weeks Open to all Munich students

Who we are

A student initiative working on the risks that come with advanced AI

We are students, academics and experts across Munich's universities who think the hardest problems in AI safety are worth working on now, not later. You do not need a background in AI safety to join. You need curiosity and a willingness to read.

01

We learn together

Structured introductions and regular reading groups, with topics curated by the club leads and members. Everyone starts somewhere, and the reading groups are built so you can join without prior knowledge.

02

We work on real projects

Each branch runs its own project, from research on compute verification mechanisms to building a database of reported AI safety incidents. The aim is to identify where further work is needed and then contribute to it ourselves.

03

We connect Munich to the wider field

Workshops with AI safety organisations and startups, talks, and introductions to people already working in the field. We are part of the broader Safe AI Munich community, so students and professionals meet.

What we work on

Two branches

Pick the one that fits your interests, or follow both. Each branch has its own lead, its own reading group and its own project.

Branch 01

Compute Verification

Industry leaders and experts are calling on governments to slow down the development of frontier AI, but the technology needed to do so doesn't exist. We are here to build it.

For any international agreement on AI to work, countries need to be able to check that others are keeping to it. Verification mechanisms that account for and evaluate the use of AI compute in a privacy-preserving manner can make such an agreement possible.

  • Structured introduction to compute verification
  • Regular reading group
  • Research projects on open problems
Read more about this branch

Through a structured introduction to compute verification and a regular reading group, we will learn about these mechanisms and identify where further work is needed, so that we can conduct research projects and contribute to it ourselves.

Compute is one of the few inputs to frontier AI that is physical, concentrated and countable, which is what makes it a promising lever for international agreements. The open questions are technical as much as political: how do you prove what a cluster was used for without exposing the model or the data?

Source: pacingthefrontier.com

Join this branch Reading list
Branch 02

AI Governance

Our current focus is on raising awareness and gathering interest around how AI development is actually regulated.

We aim to create transparency when it comes to regulatory guidelines for AI development, as well as the inconsistencies we mark as dangerous in AI reporting.

  • Reading groups every two weeks through the Winter Semester
  • Two workshops with AI safety organisations and startups
  • Internal project: an AI safety incident reporting database
Read more about this branch

We currently plan to hold reading groups during the Winter Semester every two weeks starting in October, where relevant topics are curated by club leads and members. We have also planned two workshops with AI safety organisations and startups.

Our team's internal project will be to build a reporting database, where we will compare different third-party research publications on AI safety incidents, aiming to gather complete data on the matter.

Join this branch Talk to the lead

Events

Reading groups, workshops and socials

Everything we run is listed on our Luma calendar. Subscribe there to get new events automatically, or join the WhatsApp group to hear about them first.

Calendar coming soon

The calendar is being set up

Our Luma calendar goes live together with the Winter Semester programme. In the meantime, the WhatsApp group is where every session gets announced.

Get event announcements

Reading list

Start here: compute verification

The reading group works through these in order. You do not need to have read anything in advance to turn up to a session.

  1. Start here

    Pacing the Frontier

    The clearest overview of why verification is the missing piece for any international agreement on frontier AI. Our structured introduction follows its framing.

  2. Paper

    What Does It Take to Catch a Chinchilla?

    Yonadav Shavit's proposal for verifying rules on large-scale neural network training via compute monitoring. The reference point for most later work on the topic.

  3. Paper

    Computing Power and the Governance of Artificial Intelligence

    Why compute is a uniquely governable input to AI development, and what that implies for policy. Useful shared context for both branches.

The governance branch curates its own list each semester with the club leads and members. Ask in the WhatsApp group for the current cycle.

Get involved

Ways in

Whether you want to read along quietly or run a project, there is a way in.

Join a reading group

The lowest-effort way to start. Pick a branch, show up, read along. No prior background needed.

Work on a project

Contribute to the compute verification research or the incident reporting database. Good for a thesis, a side project, or a semester of serious work.

Bring something of your own

Give a talk, propose a reading, run a workshop, or help organise. Tell us what you want to do and we will help make it happen.

Contact

Talk to us

The WhatsApp group is the fastest way to reach anyone in the initiative. For anything that needs a longer reply, email works too.

Compute Verification Lead

Diego Müller

Questions about the compute verification branch, its reading group or research projects.

Governance Lead

Begüm Balat

Questions about the governance branch, the workshops or the reporting database.

Everything else

Find us online

Day-to-day conversation happens on WhatsApp. We post updates and events on LinkedIn.