Overview

AFFINE's goal is the creation of a space which funnels newcomers into the kind of holistic, philosophical alignment theory that a pre-paradigmatic field necessitates. We believe that short-term publishing-requirements and legibility incentives cut against open-ended exploration of this type, leading to narrow research within insufficiently questioned, borrowed frameworks.

We want to expand our permanent staff and run a month-long theoretical alignment seminar in January 2027, the basic version of which is already funded. This would be a more ambitious and significantly refined followup to the first AFFINE Seminar, which we ran in May 2026. Additional funding would allow us to make the seminar even better by providing for instance travel support for more visitors, given the increased fuel prices.

The program takes place in Portugal, with future similar seminars also happening mostly in the EU (Ideally four times a year). We are currently running a year-long non-residential 8-person fellowship focused on loading/understanding the problem deeply, leading into work on ambitious, holistic alignment theory. Tracking this, the next seminar and the preparation for future instances of both, along with other projects, requires us to grow our permanent operations team significantly.

The seminar's goals and how to achieve them

We want to produce more researchers with subtle, robust generators for the core problems in AI safety and powerful babble-prune-loops capable of devising promising, truly novel research directions and agendas. We want to codify our discipline from a grab-bag of disparate methods and theories into a proper science capable of discerning those questions that are good trailheads in the investigation of key AI risks from those that are red herrings.

To this end, we are particularly interested in people combining a high level of STEM maturity and philosophical aptitude, with a track record of novel thinking and polymathy. These people will be found and admitted through professional scouting and marketing followed by a three-stage selection process.

We need to provide them freedom to think about what actually matters and to let them change course as they acquire new information and refine their thinking. Our seminars are therefore process- rather than project-driven. To avoid incentives towards financial stability, which tends to lie in the direction of capabilities, or at least an unfavourable trade-off of relevance in favor of legibility, we want to be able to offer stipends along with the possibility of the fellowship.

We need mentors who have those subtle models and are able to spot and convey them, many of whom we have already secured.

We need insightful speakers from neighbouring disciplines (such as agent foundations, philosophy of cognition, some areas of mathematics) who can open up lines of thought into unfamiliar problems and who attract promising participants. Many of these are also already secured.

We need workshops capable of inducing and making-accessible successful modes of conceptual research (builder-breaker, revising one's abstraction-toolkit, reflecting on and learning from the trace of one's thinking process, etc.) which are being collected from leading scholars and field-tested by our Seminar Designer.

We need a well-being staff to ensure our participants' ability to sustainably engage with difficult, terrifying subject matter in an unfamiliar social environment.

How the funding is used

We want to hire four additional full time ops employees for the next seven months since we are suffering from the luxury problem of scaling faster than expected. This amounts to $ 210k.

Another $ 150k would be used for website development, digital tools, external supervision, and work related travel.

As for the next seminar in January, we would want to spend $ 20k to commission curriculum materials from selected researchers, $10k to rent or buy additional office equipment at the venue and $ 60k to facilitate travel for valuable visitors and adjust our overall seminar travel-budget to the greatly increased oil prices.

An in-depth budget breakdown can be found here.

The team and our reason for believing that we are good at this

The current team consists of Noa "Ouro" Hölzer and Julia Persson, both previously of Orthogonal, Mateusz Bagiński, previously of Convergence Analysis, Jonas Hallgren of Equilibria Network and Pauliina Laine, previously holding a variety of project management roles.

We have successfully run a similar and very well received seminar in May, more than half of the current team consists of people who are or were alignment researchers, thus possessing deep and subtle models of the types of minds we seek to produce and we have been in close contact with similar org to gain expertise on how to scale sustainably.

We are advised by Abram Demski Tsvi Benson-Tilsen and Tushita Jha among others.

Failure cases

Failed extrapolation: Our accounting of which factors made the first AFFINE seminar work is highly speculative at this point, largely based on difficult-to-cheaply-validate gears-based models. Our attempts to foreground those aspects we believe to have been most impactful and to tweak variables elsewhere might therefore result in a significantly worse overall product. The result is a waste of talent and resources if we fail to quickly course-correct (that being said, the organising team showed capable of decisively course-correcting when, 2 weeks into the first program, we realised that the message about what the point of the program was was not quite understood, and therefore called for a meeting to clarify this, with good results; see this comment by a participant).

Failed recruitment: A significant portion of the AFFINE 1 attendees were already known to the organisers and hand-picked for this reason. It is possible that a cohort with fewer "already known to be a good fit”-individuals would be less cohesive and/or competent even after extensive screening. It is also possible that we cannot reach as many suitable people through our advertising campaign as we intend to, which would have the same effect. The result is lost money and a chilling effect for our reference class of alignment approaches.

Failed steering: There is a soft trade-off between the prestige of speakers/visitors and their average vision-alignment. We want leading scholars to confront our participants with interesting problems in the periphery of our discipline, but it is possible that an excessive number of such people steers the core conception of alignment within the cohort away from what we want it to be. This failure mode might cause someone to do ultimately deleterious work for a scaling lab under the guise of safety.

On a more general level we may fail to hire the right people or to delegate tasks effectively, leading us to not get as much use as we ought out of our expanded team, though contractors we have hired in the past have provided some evidence to the contrary.

How much money have you raised in the last 12 months, and from where?

We have received significant funding for the next seminar and the current fellowship during the last 12 months but are currently barred from publicly disclosing specifics.