Ivar Frisch

I have a background in philosophy and AI research. Specifically, in the past, I have focussed on research integrating the two. On the one hand, I have written papers about the philosophical implications of AI technologies on the relationship between humanity, the human mind, nature and technology. On the other hand, I have used philosophical theories to build different AI models. For example, based on Immanuel Kant's theory of cognition I build a neurosymbolic computer vision model to learn interpretable causal rules on videos. Finally, I have done (multi-agent) LLM evaluations where the experiments are inspired by philosophy. For example, I wrote a paper evaluating the personality consistency and linguistic alignment between interacting LLM populations by applying Robert Brandom's theory of linguistic inferentialism and a paper about myth making and cooperativeness in LLM populations (funded by COSMOS institute and the cooperative AI institute). Similarly, during my work at ACS research, I am investigating LLM psychology using a mix of mechanistic interpretability techniques and (psychoanalytic) philosophy of the self to understand if LLMs have self models, among other things. The relation between philosophy and AI (how to best combine these two disciplines) is one im still investigating, but im coming to believe that the 3rd path I sketched out is currently the most effective one; using philosophy to inform new experimental paradigms and critiques of LLM evaluations. We could give an example of such an critique by appealing to Robert Brandom's philosophy again. For Brandom, agency and linguistic meaning making are always normative and socially embedded. Your status as an agent only counts (only is real) when it is conferred upon you by other agents and vice versa, when you confer that status upon other agents. By conferring roles and status upon one another, we enter into normative relationships (i.e. we ought to treat each other in such and such way) structuring our individual behavior. Yet, at the same time, this allows for meaning meaking to occur and new linguistic structures to arise; by affirming others as normative agents, I have to accept their points of view as 'equal' to mine, as something to deal with. Even if its whole different from mine. Similarly, if other agents' linguistic use of concepts is very different than my own (such as life and death), I can't simply ignore that, but im forced to reconceptualize my own concepts such that they encompass the other agents' view too. As such, the multi-agent gap, and the attempt to bridge that gap, allows emergent phenomena to arise. This has a clear application to AI safety research. Current AI safety research is still mostly focussed on single model behavior and experiments. Models are often studied in isolation. Although, the field of multi-agent AI has existed for long, mechanistic interpretability still often does not engage with this. Leading researchers to claim we need to adopt a more interactionist paradigm (https://arxiv.org/pdf/2601.10567). So, I would really like to contribute by this. One way would be to see if we could use more of robert brandoms' inferentialism to construct something like multi-agent mechanistic interpretability; building inferential linear probes and toy models (and other methods) to distinguish views of multiple agents from one another and the different views which an agent might hold internally (Reasoning models generate societies of thought: https://arxiv.org/abs/2601.10825)

LinkedIn

LinkedIn

Projects

No grants recorded.