Effective Altruism Forum
EA Forum

Hide table of contents
by Dušan D. Nešić (Dushan)
4 min read 0

7

AI safetyCareer choiceBuilding the field of AI safetyFellowships and internshipsPostmortems & retrospectivesResearch
Frontpage

Between June and September 2024, we ran the third iteration of the PIBBSS Summer Research Fellowship. Here are our reflections on how the program went and what we learned. Apply for the 2025 program here!

TLDR:

The 2024 fellowship demonstrated continued success in attracting and developing senior academic talent while refining our approach to interdisciplinary AI safety research. Key outcomes include:

  • Successfully facilitating the transition of multiple early-to-mid-career researchers into AI safety
  • Strong research outputs, including advances in mechanistic interpretability that one mentor described as "a 70th percentile paper in mech interp"
  • Increased seniority of fellows led to more sophisticated and complete research outputs
  • Going forward, we plan to strengthen research support infrastructure and better balance research direction diversity, which is the usual strength of our program.

About PIBBSS

PIBBSS (Principles of Intelligent Behavior in Biological and Social Systems) is a research initiative focused on leveraging insights and talent from fields that study intelligent behavior in natural systems to help make progress on questions in AI risk and safety. Since our inception, our approach has evolved - while maintaining our commitment to epistemic pluralism, we have started taking steps toward focusing on more concrete research directions. This includes launching an affiliate program in January 2024 to support sustained research efforts. Previous retrospectives can be found for 2022 and 2023.

Fellowship Structure

The fellowship pairs researchers (typically PhD level or above) from fields studying complex and intelligent behavior with mentors from AI alignment. The 2024 program featured several key elements:

  • Opening retreat
  • Three-month fellowship period (June-September) at LISA facility in London
  • Structured reporting throughout
  • Closing symposium

2024 Program Changes

The fellowship relocated to London's LISA facility in 2024, extending to a three-month residency. This move offered practical advantages and increased exposure to the broader AI safety community. Fellows reported that "having meals and snacks provided was incredibly convenient" and valued "the general atmosphere of having lots of people working intensely on interesting things."

Due to budget constraints, we have had to cut the closing retreat, which was usually held just prior to the Symposium. Based on feedback, we will aim to host one in the future if budget permits - the value of "closing the vessel" and having a clean ending to a program is higher than we had foreseen.

We also hired a dedicated program director this year (Clem Von Stengel), and the overall response was quite positive. Clem also leveraged London and LISA more concretely as a place with interesting researchers and organized semi-regular events for fellows, some of whom reported high value of these extra sessions.

Result Highlights

Fellowship produces impact in several avenues - counterfactual career shifts, research, mentor benefits, and bridge-building. Some of these are harder to pinpoint, and most of our Fellows have contributed to more than one of these, but we would like to highlight a few especially salient and tangible ones (in alphabetical order):

  1. Agustin Martinez Suñé created a framework integrating automated planning with LLMs to provide measurable safety guarantees, effectively bridging formal methods and AI safety [presentation]. As a direct result of the Fellowship, he received a postdoc offer at a world-class university and began the initiation of an AI safety scholarship program for Argentina, representing a significant field-building impact.
  2. Aron Vallinder investigated cultural evolution and cooperation in LLMs, developing work (accepted at AAMAS as extended abstract) that demonstrated the viability of applying cultural evolution concepts to AI systems. His project represented solid research progress aligned with his existing trajectory while opening new avenues for understanding multi-agent dynamics.
  3. Euan McLean developed a framework for searching for phenomenal consciousness in LLMs, creating concrete experimental proposals focusing on measuring introspective metacognition. The fellowship provided a structured opportunity to develop these novel approaches, with his mentor noting the work brought "particularly valuable new ideas" to the field.
  4. Magdalena Wache produced a comprehensive formalization of Factored Space Models for understanding causality between abstraction levels, delivering high-quality technical work clarifying important concepts. Her strong distillation work made complex concepts more accessible while maintaining mathematical rigor and it reached an impressively polished publishing state in a short time.
  5. Matthew A. Clarke discovered that SAE features show heavy-tailed co-occurrence patterns rather than independence, providing important insights for interpretability research that are being prepared for publication. His transition from biomedical research to AI safety represents high counterfactual value, with his strong scientific background bringing valuable methodological rigor to the field.
  6. Assistant Professor at U of Toronto Yevgeny Liokumovich contributed to singular learning theory by deriving higher order/constant terms of free energy and extending Jeffreys prior to singular cases, bringing rigorous mathematical foundations to developmental interpretability [presentation]. His transition into an agenda fit to his interests and expertise represents a significant counterfactual impact as it may not have otherwise happened. Further plans include continued collaboration with developmental interpretability researchers and potentially supervising future PhD students in AI safety.

Presentations from all fellows in video form are available in the Appendix.

Result Changes from 2023

Compared to 2023, we observed:

  • Higher average seniority of fellows
  • More publication-ready research outputs
  • More engagement with present-day ML systems in projects
  • Better integration with the broader AI safety community

Fellow and Mentor Feedback

Fellows particularly valued the LISA environment: "Working at LISA was definitely an improvement over working from home or working from a non-AI-alignment office."

Mentors reported high satisfaction:

  • 6 out of 8 mentors rated outcomes 7/10 or above
  • 8 out of 9 indicated they would "Pretty likely" or "Somewhat likely" want to mentor again
  • The majority rated their PIBBSS mentorship as more useful than their typical work

Areas for improvement included:

  • Need for better cohort cohesion
  • More structured post-fellowship support
  • Additional technical support for experimental work

Future Plans

For 2025, we plan to:

  • Continue the fellowship program with enhanced research support
  • Strengthen post-fellowship support infrastructure
  • Better balance technical and theoretical diversity in future cohorts
  • Explore new funding models through targeted support of specific research directions

Expression of interest for Fellowship 2025

If you are interested in doing research with us yourself, please fill in the expression of interest form (for non-Fellowship interest, register here). We will announce an official opening of the fellowship once we have funding confirmed for the next year, but in the meantime we would like to know what areas of research are people interested in. By filling in the form you are helping us reach out to relevant mentors and funders on time, and we will inform you of the opening of the Fellowship directly to your email!

Appendix: Fellowship Research Projects

  1. Agustin Martinez Suñé - "Neuro-symbolic approaches for achieving quantitative safety guarantees for LLM-based agents": Developed a framework integrating automated planning with LLMs to provide measurable safety guarantees for LLM-based agents. As a direct result of his engagement with PIBBSS, received informal offer for postdoctoral work at a top university. [presentation] [website]
  2. Aron Vallinder - "The Cultural Evolution of Indirect Reciprocity in LLMs": Explored how cultural evolution and cooperation emerge in LLM interactions, with results being developed for AAMAS submission. [presentation]
  3. Baram Sosis - "Measuring Beliefs of Language Models During Chain-of-Thought": Investigated several methods for measuring LLM beliefs during chain-of-thought reasoning and attempted to apply bounded rationality models. [presentation]
  4. Euan McLean - "Searching for phenomenal consciousness in LLMs": Developed framework for investigating consciousness in LLMs, particularly focusing on distinguishing direct and indirect metacognition. Made significant updates on the validity of computational functionalism. [presentation] [LW post]
  5. Jan Bauer - "The geometry of in-context learning": Explored geometric properties of in-context learning, particularly focusing on Chain-of-Thought reasoning through a computational neuroscience lens. [presentation]
  6. Magdalena Wache - "Factored Space Models: Causality between Levels of Abstraction": Comprehensive formalization of Factored Space Models providing mathematical foundations for understanding causality between abstraction levels. [presentation] [full paper]
  7. Mateusz Bagiński - "Conceptual investigation of the core drivers of goal-achieving mental activity": Applied hermeneutic net method to investigate core concepts in minds, agency, and alignment. [presentation][published WIP]
  8. Matthew A. Clarke - "Studying co-occurrence patterns in Sparse Autoencoders": Discovered heavy-tailed co-occurrence distributions in SAE features, challenging basic assumptions about SAE behavior and suggesting new directions for interpretability research. [presentation][LW Post]
  9. Nadine Spychala (part-time fellow) - "Exploring the potential of formal approaches to emergence for AI safety": Conducted preliminary investigation of applying information theoretic measures of emergence to AI safety evaluation. [presentation]
  10. Shaun Raviv - Conducted extensive foundational research in AI safety as preparation for future journalism work focused on making complex technical content accessible to broader audiences. [personal website]
  11. Wesley Erickson - "Heavy-tailed Noise & Stochastic Gradient Descent": Investigated the role of heavy-tailed noise in SGD and its implications for understanding learning dynamics. [presentation]
  12. Yevgeny Liokumovich - "Minimum Description Length for singular models": Contributed to theoretical understanding of Watanabe's asymptotic expansion in Bayesian statistics for singular models, with applications to developmental interpretability. [presentation]

This fellowship's research particularly excelled in:

  • Mathematical rigor (singular learning theory, factored spaces)
  • Novel experimental approaches (SAE analysis, LLM evolution)
  • Concrete safety frameworks (neuro-symbolic guarantees)
  • Theoretical foundations (consciousness studies, emergence measures)

Multiple projects are proceeding toward publication or conference submission, and several fellows have secured paths to continue their research through academic positions or research organizations.

Thanks

Many thanks to our funders for the Fellowship in 2024 including SFF, LTFF, Cooperative AI Fund, and the Foresight Institute. This would not have been possible without our mentors, alumni, London AI Safety Initiative (LISA) and many others who have helped us along the way. If you wish to support our work, please reach out to us at contact@pibbss.ai, full version of this report is available to funders.

7

0
0

Reactions

0
0

More posts like this

Comments


No comments on this post yet.
Be the first to respond.
More from Dušan D. Nešić (Dushan)
104
Introducing "Bribe Well"
Dušan D. Nešić (Dushan)
· · 3m read
15
15
73
EA Serbia is now launching!
Dušan D. Nešić (Dushan)
· · 1m read
8
8
38
Introduction to Non-EA International Charity Orgs.
Dušan D. Nešić (Dushan)
· · 7m read
8
8
View more
Curated and popular this week
LintzA
 ·  · 15m read
 · 
Cross-posted to Lesswrong Introduction Several developments over the past few months should cause you to re-evaluate what you are doing. These include: 1. Updates toward short timelines 2. The Trump presidency 3. The o1 (inference-time compute scaling) paradigm 4. Deepseek 5. Stargate/AI datacenter spending 6. Increased internal deployment 7. Absence of AI x-risk/safety considerations in mainstream AI discourse Taken together, these are enough to render many existing AI governance strategies obsolete (and probably some technical safety strategies too). There's a good chance we're entering crunch time and that should absolutely affect your theory of change and what you plan to work on. In this piece I try to give a quick summary of these developments and think through the broader implications these have for AI safety. At the end of the piece I give some quick initial thoughts on how these developments affect what safety-concerned folks should be prioritizing. These are early days and I expect many of my takes will shift, look forward to discussing in the comments!  Implications of recent developments Updates toward short timelines There’s general agreement that timelines are likely to be far shorter than most expected. Both Sam Altman and Dario Amodei have recently said they expect AGI within the next 3 years. Anecdotally, nearly everyone I know or have heard of who was expecting longer timelines has updated significantly toward short timelines (<5 years). E.g. Ajeya’s median estimate is that 99% of fully-remote jobs will be automatable in roughly 6-8 years, 5+ years earlier than her 2023 estimate. On a quick look, prediction markets seem to have shifted to short timelines (e.g. Metaculus[1] & Manifold appear to have roughly 2030 median timelines to AGI, though haven’t moved dramatically in recent months). We’ve consistently seen performance on benchmarks far exceed what most predicted. Most recently, Epoch was surprised to see OpenAI’s o3 model achi
Dr Kassim
 ·  · 4m read
 · 
Hey everyone, I’ve been going through the EA Introductory Program, and I have to admit some of these ideas make sense, but others leave me with more questions than answers. I’m trying to wrap my head around certain core EA principles, and the more I think about them, the more I wonder: Am I misunderstanding, or are there blind spots in EA’s approach? I’d really love to hear what others think. Maybe you can help me clarify some of my doubts. Or maybe you share the same reservations? Let’s talk. Cause Prioritization. Does It Ignore Political and Social Reality? EA focuses on doing the most good per dollar, which makes sense in theory. But does it hold up when you apply it to real world contexts especially in countries like Uganda? Take malaria prevention. It’s a top EA cause because it’s highly cost effective $5,000 can save a life through bed nets (GiveWell, 2023). But what happens when government corruption or instability disrupts these programs? The Global Fund scandal in Uganda saw $1.6 million in malaria aid mismanaged (Global Fund Audit Report, 2016). If money isn’t reaching the people it’s meant to help, is it really the best use of resources? And what about leadership changes? Policies shift unpredictably here. A national animal welfare initiative I supported lost momentum when political priorities changed. How does EA factor in these uncertainties when prioritizing causes? It feels like EA assumes a stable world where money always achieves the intended impact. But what if that’s not the world we live in? Long termism. A Luxury When the Present Is in Crisis? I get why long termists argue that future people matter. But should we really prioritize them over people suffering today? Long termism tells us that existential risks like AI could wipe out trillions of future lives. But in Uganda, we’re losing lives now—1,500+ die from rabies annually (WHO, 2021), and 41% of children suffer from stunting due to malnutrition (UNICEF, 2022). These are preventable d
Rory Fenton
 ·  · 6m read
 · 
Cross-posted from my blog. Contrary to my carefully crafted brand as a weak nerd, I go to a local CrossFit gym a few times a week. Every year, the gym raises funds for a scholarship for teens from lower-income families to attend their summer camp program. I don’t know how many Crossfit-interested low-income teens there are in my small town, but I’ll guess there are perhaps 2 of them who would benefit from the scholarship. After all, CrossFit is pretty niche, and the town is small. Helping youngsters get swole in the Pacific Northwest is not exactly as cost-effective as preventing malaria in Malawi. But I notice I feel drawn to supporting the scholarship anyway. Every time it pops in my head I think, “My money could fully solve this problem”. The camp only costs a few hundred dollars per kid and if there are just 2 kids who need support, I could give $500 and there would no longer be teenagers in my town who want to go to a CrossFit summer camp but can’t. Thanks to me, the hero, this problem would be entirely solved. 100%. That is not how most nonprofit work feels to me. You are only ever making small dents in important problems I want to work on big problems. Global poverty. Malaria. Everyone not suddenly dying. But if I’m honest, what I really want is to solve those problems. Me, personally, solve them. This is a continued source of frustration and sadness because I absolutely cannot solve those problems. Consider what else my $500 CrossFit scholarship might do: * I want to save lives, and USAID suddenly stops giving $7 billion a year to PEPFAR. So I give $500 to the Rapid Response Fund. My donation solves 0.000001% of the problem and I feel like I have failed. * I want to solve climate change, and getting to net zero will require stopping or removing emissions of 1,500 billion tons of carbon dioxide. I give $500 to a policy nonprofit that reduces emissions, in expectation, by 50 tons. My donation solves 0.000000003% of the problem and I feel like I have f
Recent opportunities in AI safety
3
Apply to become a Futurekind AI Facilitator or Mentor (deadline: April 10)
Jay Luong
· · 2m read
0
0
35
Apply to MATS 8.0!
Ryan Kidd
·
0
0
37
AIS Netherlands is looking for a Founding Executive Director (EOI form)
7 authors
· · 5m read
4
4
View more