AFFINE II · 4 – 29 January 2027

2027 Seminar Design

Here is how we plan to run the month. We are still preparing, so details will change.

Three aims

1. Space to overhaul your assumptions

We don’t have good frameworks for thinking about optimization and learning. Yet the temptation to just jump into the field based on our initial assumptions is strong. It is easy to start treating the field as if it were normal science with its foundations settled.

Progress is going to involve overhauling key assumptions and criteria, over and over. This can leave you even more confused than before, as things you previously took for granted are no longer things you can hold true.

You get an appearance of negative progress, which is implicitly punished in most contexts. Expectations of output demand holding on to a single framework for a long time, and legibility abhors confusion.

In January we want to create space for confusion and overhauling. We will not push for short-term legible output, and we strive for a culture where infectious questioning is held in high regard. We will ask questions without necessarily resolving them.

2. Work on the questions, not only read about them

Nobody understands these problems well yet, so there is no body of material you can study and then know what to do. We want you to sit down and try to make progress: write down an attempt, see where it breaks, try again. Reading supports the attempts; it does not replace them. Trying is where we can help you most, so the workshops, the daily writing hour and the time with mentors are built around it.

You are welcome to argue that one of our questions is the wrong question. We ask one thing first: understand it well enough that the people who take it seriously would agree with how you state it.

3. Get out of the way of conversations

Most of the good stuff happens between people. So we keep talks to one a day, leave large blocks of open time, and run most workshops in pairs or small groups. Participants, mentors and visitors come from agent foundations, philosophy of science, artificial life, physics, biology, applied category theory, cybernetics and complex systems, so the person at the next whiteboard will often see the problem differently from you.

Visitors come with questions rather than lectures: each spends a few minutes on questions they find interesting, then their days with participants.

A typical day

Days share the same shape; the content changes. This is the current plan.

Approximate. The green block, the writing hour, is the one mandatory session and the one fixed time. Open time runs alongside the workshop and office hours.

Morning

An optional workshop: an exercise in tackling one specific question with one method, often in pairs. A common one is to write down a plan for an aligned AI that learns, or to take apart someone else’s plan.

Open time

Most of the day. Work on a question with another participant, talk with a mentor or visitor, read, write, or walk in the countryside. Mentors and visitors hold office hours after lunch.

17:00–18:00

Research journalling, the one mandatory session. We write to think: you start from a question without knowing what you will write. It works well, especially for people who have never done it.

Into dinner

One talk. Past examples: “Map of Agent Foundations”, “Introduction to Logical Inductors” and “On Corrigibility”.

The venue is in the Portuguese countryside, with walks right outside and whiteboards everywhere.

An example from the first seminar

To exemplify what we mean by this we will share a prototypical discussion which happened at our first seminar, as an example of what we want to have happen again.

A central aspect of the alignment problem is embedded agency. Traditional theoretical approaches towards modeling agents fail to carry over when modeling embedded agents: for example, a standard Bayesian agent has a prior over hypotheses, each of which is a complete probabilistic model of the environment. The nice theorems (convergence to the truth, low regret) assume realizability: the true environment is one of the hypotheses in your prior.

For an agent embedded in the real world, that assumption is false. The world contains the agent, other agents as smart as it, and far more detail than it can represent, so the true environment can’t sit inside its hypothesis space. Without realizability, the Bayesian guarantees can fail badly. Infra-Bayesianism is a modification of Bayesian reasoning for agents that can’t fully model their world. It provides a mathematical theory of what a coherent embedded agent can look like.

Coming from recent advances in theoretical neuroscience is a completely different approach, called Active Inference. Active Inference tries to answer the question: “can we describe intelligent behavior as we see in the world around us from first (physical) principles?” It is related to the theory of Predictive Processing, which says that our brains are prediction machines that minimize prediction error. Active Inference basically says that all systems that maintain their self-organization over time (approximately) minimize prediction error, by minimizing an upper bound of it called free energy. This is also a description of what a coherent embedded agent can look like.

Participants were introduced to both ideas during the seminar; some of them had a background understanding of Infra-Bayesianism and some had a background in Active Inference.

In the readings and workshops that we held, people inevitably tried to make sense of these somewhat notoriously complicated topics. Some of our fellows gave a presentation on the intersection.

Active Inference and Infra-Bayesianism actually don’t share all that much in common except that they both describe (very different) visions of what a coherent (somewhat Bayesian) embedded agent can look like. Their approach towards normativity is very different. Infra-Bayesianism’s main hope is to build a mathematical language in which you can state and prove things about agents in a world bigger than themselves, which could be used to build some alignment protocol on top. Active Inference was not developed for the purpose of alignment, but rather as a general framework for thinking about intelligent (primarily biological) systems as being approximate Bayesians.

This kind of debate connects to a much more general debate of the role of theoretical approaches to alignment. Theories can be descriptive (what best explains our observations of the natural world) vs prescriptive (aiming for knowledge like “X algorithm works best for Y problem”). They can be motivated almost entirely mathematically, or make various levels of contact with other fields and empirical observations. These discussions seem to naturally bottom out in some degree of philosophy of science. You can see the following post and the following comment, which is essentially the spiritual successor to the discussions had at AFFINE I.

We are still very confused about this intersection, but it is an interesting confusion which might lead to less confusion in the future. Being productively confused is in the spirit of AFFINE.

The questions

The questions we care about, and why we are not sure they are the right ones yet, are on the Approach page.

Apply

Applications close on 22 November 2026. Dates, costs and the FAQ are on the program page.

Applications for the January 2027 seminar are closed. Ask to be notified about the next one, or read more on the program page.