Research

This analysis conceptually connects current obfuscation results with some of MIRI's more pessimistic threat models (e.g., deep deceptiveness, RAAPs) and suggests how we might unify them all under a common framework.

A report on the interface design workshop where we explored how to create culture and technology to replace the current less imaginative ways of interfacing with AI and connected it to meaningful progress on alleviating risks from generally intelligent systems.

CEEALAR hosted these talks emphasise the difference between grown connections and modular connections, AI and traditional software, slow integrated caring vs plug and play sensors and actuators. Sahil goes on to ask what is the Self that wants to preserve itself, grow, is situationally aware and holds motivations

The phase 1 prototypes of live interfaces involving ~20 people, culminated in 8 interface prototypes with infrastructural cohesion (final demo video).

After the AISC 2025 cohort finished, we had three teams working on their interfaces. The Soloware Platform, Live Discernment and Live Conversation Threads. This video is a presentation of the prototypes Aayush Kucheria, Kuil Schoneveld, Aditya Adiga, Jayson Amati have built.

You can read the transcript here , Sahil and Abram discuss problems of AI addiction and AI addictiveness: when does a FOOM become a POOF due to wireheading-like dynamics?

You can read the transcript here , Sahil often makes remarks about preciousness and irreplaceability, claims about things that can't be done by mere simulation or information transfer. Abram is as unconvinced by these.

You can read the transcript here , Abram and Sahil will often/endlessly talk about 'substrates'. Substrate-relevance is an interesting idea, where we can't understand the algorithms that are running on a medium without taking into account the details of the physical make-up. In this conversation we connect decentralization with decentralized cognition, both for threat models and responses to them (substrate-sensitive risk management).

You can read the transcript here , Abram and Sahil continue their conversation on the significance of decentralization in AI and live machinery, still including differentiated AI interface design, but especially in connection with centralized training of AI models.