Research

article 45 min read, 12,700 words
Management of Substrate-Specific AI Capabilities (MoSSAIC)
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. It was accepted by ILIAD and is currently part of BlueDot's Technical AI safety course
by Matt Farr et al.
theorytechnical

poster 8 min read
MoSSAIC poster presented at TAIS 2025
This is a poster that was presented at the Tokyo AI Safety Conference - the initial version of the paper submitted to ILIAD (2025): ODYSSEY
by Matt Farr et al.
theoryposter

article 18 mins
Live Machinery Workshop
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.
by Harshit
designlive-theory

video 5 hours, 32 minutes
Portability of Meaning : Talk Series during AI Winter Season at CEEALAR
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
by Sahil
philosophyalignment

video 1 hr 17 min
AI Safety Camp 2025 Demos
The phase 1 prototypes of live interfaces involving ~20 people, culminated in 8 interface prototypes with infrastructural cohesion (final demo video).
by AISC 2025 cohort
collectiveinfrastructure

video 1 hr 49 min
Presentation of Fellowship Prototypes
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.
by AISC 2025 Fellows
demoengineering

video 1 hr 56 min
Live Machinery Infrastructure with Abram and Sahil
Watch the video and read the transcript. The opening session. Sahil proposes "scale vs sensitivity" as the real polarity (not centralization vs decentralization). They discuss pace layering, the prompt as product, soloware and multiplayer, and what it means to build infrastructure for a high-actuation future.
by Sahil & Abram
infrastructurelive machinery

video 1 hr 49 min
A Less Mindless Moloch & Steam Design Philosophy
Watch the video and read the transcript. Sahil introduces "steam"—probability, salience, reification. Deep coordination looks like "being lucky from the inside." They discuss potentials as first-class citizens, counter-steam reasoning, and the stealing economy.
by Sahil & Abram
molochagency

video 1 hr 53 min
Interface FOOM: How significant are interface, really?
Watch the video and read the transcript. How significant are interfaces, really? Sahil and Abram debate whether interface development could lead to rapid capability gain (FOOM) or self-destructive wireheading (POOF). They discuss relevance realization, the membrane between human and AI, and Manhattan projects at scale.
by Sahil and Abram
FOOMPOOF

video 2 hr
Weak-To-Strong Generalization & The Metaphysics of Friendship
Watch the video and read the transcript. A weak teacher trains a strong student who surpasses them—how? Abram connects this to simulation and the "emulation barrier." Sahil argues that referential cuts can't be bridged by telling. The co-second law returns: you can't force friendship.
by Sahil and Abram
generalizationfriendship

video 2 hr
How robust is human potentiation?
Watch the video and read the transcript. You can't download kung fu. Sahil argues that forcing integration is like forcing friendship—both require real openness. They explore chip-in-brain syndrome, the co-second law, and what remains irreplaceable about embodied humans.
by Sahil and Abram
FOOMPOOF

video 2 hr 12 min
Decentralized Bodies of AI
Watch the video and read the transcript. Does the substrate matter? Abram introduces "finers vs coarsers"—whether agency extends all the way down into biological wet-ware. They discuss live theory, diffuse deceptiveness, and why integration can't be instantaneous.
by Sahil and Abram
decentralizedbodies

video 2 hr 4 min
Decentralized training, and nature vs nurture of trained models
Watch the video and read the transcript. What's baked in, what's shaped? Sahil and Abram explore how training becomes indistinguishable from nature, why language itself "shepherds" thought, jailbreaking as poor man's alignment, and whether we can escape the ad economy before enshittification.
by Sahil and Abram
decentralizedbodies

video 2 hr
Sim-o-phone
Watch the video and read the transcript. You call someone in a simulation to tell them they're in a simulation. The phone connects. Does the message land? Sahil argues integration is hard—not from stupidity but from distraction, addiction, the gravity of the world you're already in. Abram remains unconvinced.
by Sahil and Abram
simulationinterfaces







