Georgia Tech AI Safety Initiative Retrospective 2025-2026

• AY 2025-26 was an outlier year for Georgia Tech’s AI Safety Initiative (AISI), with 15+ members placed in AI safety roles. In this post, we distill our most important advice for other university groups. - Key takeaways:
• Deliberately identify potential talent in the fellowship, heavily invest high-context organizer time in great people.
• 1:1s are probably your most neglected tool.
• Being professional and punctual is underrated.

In this post, we'll share general lessons we've learned from organizing that can hopefully be applied to other AI safety groups. We're still running many of the same initiatives as before — fellowships, reading groups, research projects, and general meetings — but we think we've gotten higher quality members and had more organizational success compared to the previous academic year.

Since our last update, more than 15 members have been placed in paid fellowships & full-time roles. We’ve accepted 150+ people into our fellowships and grown to 50–100 active members.

Intro fellowships #

Our intro fellowships continue to be our most impactful offering. They’re how we find talented, dedicated members. Our Spring acceptance rate was much higher than the Fall’s due to a higher volume of applications in the Fall. Because selection is noisy, we recommend running more fellowships rather than fewer (conditional on promising applicants). We missed out on several great applicants in Fall because we were capacity constrained, and we cannot be sure that the same hasn’t happened in the Spring.

Statistics:

• Applicant demographics, divided between technical and policy.

• Fall

• Technical: 150 applications of which 67 undergrad, 34 master's, 20 PhD, 29 non-students

• Technical: ~58 accepted of which 18 undergrad, 18 master's, 9 PhD, 13 non-students | 28 in-person, 30 online

• Policy: 94 applications of which 19 undergrad, 29 master's, 8 PhD, 38 non-students

• Policy: ~22 accepted of which 5 undergrad, 5 master's, 3 PhD, 9 non-students | 5 in-person, 17 online

• Spring

• Technical: 88 applications, of which 50 undergrad, 29 master's, 6 PhD, 3 non-students

• Technical: 68 accepted of which 38 undergrad, 22 master’s, 6 PhD, 2 non-students | 50 in-person, 18 online

• Policy: 27 applications, of which 15 undergrad, 7 master’s, 2 PhD, 3 non-students

• Policy: 21 accepted, of which 9 undergrad, 7 master’s, 2 PhD, 3 non-students | 11 in-person, 10 online

• Rough attendance stats, measured as ending cohort size as a percent of starting cohort size

• 85% attendance for policy

• 50% attendance for technical

• Feedback form results

• Fellow engagement was mixed – many stopped showing up or doing readings completely; others were extremely engaged.

• Cohort sizes often get too small and rarely get too large. Sweet spot seems to be 5-6 so best initial sizing is probably 10-12 for technical and 6-8 for policy

See our technical and policy syllabi.

Speaker Events #

We ran 5 speaker events in Fall, 3 in Spring. We are the biggest AI safety organization in Atlanta, providing a home for safety-minded individuals around the city.

• We had a median of ~45 in-person attendees and ~10 remote attendees in Spring.
• We were able to find speakers through multiple routes: personal networks, alumni connections, Georgia Tech professors, networking events (OASIS, Action Potential, Global Challenges Project, etc.), and cold-emails.
• Speaker events not only increase visibility but create a scenario that feels low-stakes for people to come up to organizers and ask questions.

Placements #

Our members secured external opportunities including MATS, Astra, METR, Anthropic Fellow, IAPS, ERA, Pivotal Fellowship, Horizon Fellow, CG grant(s), GA House of Reps, American Enterprise Institute, Meta Alignment Team and many SPAR streams.

Other activities #

• Graduate student focused empirical reading group

• ~8 average attendance with ~5 regulars

• 1 research project/extension birthed.

• Faculty engagement: we had 1:1s with ~8 new professors, which has led to:

• Setting up several new projects including a collaboration with Goodfire, and interpretability work on AlphaFold.

• Getting verbal commitments from 6+ faculty to contribute to an institutional research center for AI safety, and a go-ahead from the College of Computing Dean.

• Acquiring a dedicated conference room for coworking and fellowship for free!

- Two trips for high-potential members and organizers:
- Policy-focused Washington DC trip (partially organized by GT AISI)

• ControlConf in Berkeley

• Two-hour ARENA-style technical workshop dedicated to AI security (attended by 22 safety fellows)

• A “mixer” with another professional-oriented organization. Overall it did not lead to productive collaboration, and we are moving away from spending time on socials without a specific purpose

In rough order of importance.

Build around exceptional people #

Focus on a few highly motivated and agentic members. It takes frequent contact for someone to consider switching career paths. With finite organizer time for efforts beyond the fellowship it's better to invest in quality individuals rather than casting a wide net. Almost everyone in AISI who has completed competitive fellowships or is likely to work in AI safety long-term expressed interest in AI safety early on.

Operationalized:

• Invest in high-potential people deliberately: refer them to opportunities, recommend them to research projects, send them to workshops, and tell them you think they would be good organizers. Raising ambitions is often a self-fulfilling prophecy.

• Signals of high potential include:

• Impressive feats within their own fields.

• Ability to update world models and re-plan in light of new evidence.

• Clear, first principles, strategic thinking, being aware of uncertainties.

• Reliability and dedication to their work.

• We have a big-little system where new team members are paired with experienced organizers for weekly/biweekly 15-30 min 1:1s.

• If you're not super well resourced, prioritize having fewer motivated people rather than many uninterested members.

Have more 1:1s #

1:1s are probably your most underused tool.

1:1s with fellow organizers — this is a great way to build a core group and create an environment for AI safety discussion. Many AI safety groups seem fragmented in that core organizers don't know a lot about each other’s interests and plans, and this may hurt coordination, information sharing, and tight feedback loops. We specifically recommend a big-little system to help onboard new organizers.

1:1s with new people and intro fellows — for someone to conclude that they should change their default career trajectory, they need to do more than read blogs and listen to speakers. Everyone has a different model of the world and perception of AI safety. 1:1s help you figure out what is going on with them and open up space for mentorship and personalized career advice. AISI leadership probably has ~four 1:1s with new folks every week in total.

Three major success stories that came from 1:1s.

• Yixiong went to a dinner with Parv on a random Tuesday night and quickly onboarded him as an organizer over the coming week. Parv did the Astra fellowship and is now Chief of Staff at METR.
• Parv and Yixiong had a ‘1:1’ with Glenn to hear out his research. This turned into sustained dialogue and Glenn is now on MATS extension, starting a new org around data attribution for LLM personas.
• Andrew got introduced to AISI through a personal friendship with Yixiong and a pizza 1:1 with Ayush.

Professionalism is still underrated #

Building a brand is important, and the ceiling for good marketing is quite high. There is a lot of competition for the attention of university students. For students to consider switching away from default, prestigious paths, your group needs to pattern-match to the same level of professionalism and impressiveness they expect from top-tier opportunities. If your marketing has typos, if your elevator pitch is unconvincing, if your QR code doesn't work, if your food doesn't show up, if your meeting is chaotic and poorly designed, people might lose interest without even getting involved in the club. For people to take your claims seriously, you need to show that you are serious.

Concretely:

• Be detail-obsessed with anything public facing. Avoid typos and awkward phrases. Dress well and prepare talking points in advance for club fairs and talks. Emails to important people should be reviewed by another person.
• Your marketing should be everywhere,

• *especially at the start of the year when people join new clubs the most. Figure out what mailing lists you can occupy, where you can put posters, etc. Do analytics to understand what works best. - Be conscientious about operations. Events should start on time. Food should be there when you say it will. Fellowship decisions should come back promptly.

• Always have a call to action. After every meeting, every event — what should members do next?

Leverage low-context organizers while helping them develop context #

There are a lot of operational tasks that ideally should never be touched by high-context organizers. Your research lead shouldn't be filling out finance spreadsheets. Your policy lead shouldn't be handling the logistics of fellowship applications.

Doing ops work well is a strong signal for generalist fieldbuilder qualities. We currently run a two-week work-trial with an initial cohort, then select those who clear the bar. So far, this has been working well because it gives us signal that you can scarcely get from applications and interviews:

• How fast they are at responding and completing tasks.
• Conscientiousness and agency.
• Whether they work well with the team.
• Whether they care enough about the mission to put in real hours.

Speaker Events #

There are several improvements we could have made to our speaker events. Most boil down to doing the obvious things in a timely manner.

• Get logistics and speakers sorted around 6 weeks in advance, to avoid venue and catering-related challenges.
• Cold emailing speakers 2-3 months in advance, so they have time to find an availability in their schedule.
• Get talks that focus more on strategy and motivate the case for ‘why AI safety’
• Some of our talks were explainers of technical papers our invited speakers worked on — however, these were hard for audiences to understand and very narrow in scope.

Changes to the Fellowship #

• Fellowships often suffered from people not doing the required readings. We are considering moving readings back in-session and increasing the duration of each session — this has been recommended by other organizers. - We are modifying our curriculum to focus more on strategy and motivating AI safety instead of diving deep into technical problems and papers.
• We’re prioritizing in-person fellowships since engagement seems easier to maintain.
• We’re setting up intra-fellowship socials between the second and third session where the AISI organizer takes the cohort out to a coffee shop and has unstructured discussion.

Post-Fellowship Involvement #

• We’ve historically struggled with getting fellowship graduates who didn’t end up joining our team to do meaningful research. Folks generally do not yet meet the bar for MATS-type fellowships and there is a lack of faculty that work on problems in AI safety. There’s no ‘obvious next thing’ to do. To solve this, we’ve tried:

• Giving project ideas, sorting people into teams, and being mentors ourselves. Only two out of 6 such projects are expected to have a publishable outcome, but both took way longer than the ideal timeline. The main failure mode was a self-reported lack of time from participants and AISI managers. This setup also did not select the best people because it is low-friction to say yes first and then slack off later.

• We got people to read the SPAR projects page, pick a few mentors, propose ideas & followups, do 1-2 week research sprints — and we offered to help them reach out for external mentorship. However, no fellows independently pursued projects after this.

• This is a problem we’re still attempting to resolve. At some point, people will need to be agentic to contribute to AI safety. We’re not sure this is a problem worth solving — it may be a feature, not a bug!

Deprecated Rapid Upskilling Cohorts #

• We decided to axe our rapid upskilling cohorts. These were ARENA-style accountability groups, where we provided mentorship upon request. Out of the ~40 participants, only 2 still work with us. A demanding curriculum just cannot be hosted in a low contact way. Solution: We have pivoted to hosting 2-3 hour, self-contained, sub-field specific workshops (shoutout Meru!) and helping ARENA develop bite-sized notebooks for sub areas such as LLM attacks & defenses, emergent misalignment, evals, etc. This is much less costly and achieves similar outcomes for motivated individuals.

If you're building a university AI safety group and want to talk, [reach out](mailto:board@aisi.dev). We're happy to share more specifics on curriculum, fellowship pipelines, or institutional strategy.

[Discuss](https://www.lesswrong.com/posts/DgLGD3A5CZXH7hbzP/georgia-tech-ai-safety-initiative-retrospective-2025-2026#comments)