A tutorial on how to use Clifford Algebras and SageMath to represent geometric objects as vectors in an algebraic manner amenable to direct computation.
A proposal to create joint text-EEG embeddings by having participants read language model outputs while wearing EEG headsets, aiming to increase the bandwidth of human feedback for AI alignment.
Proposes using GPT-4 to generate group-specific AI constitutions for diverse demographic and philosophical groups, then having those constitutions debate in a round-robin competition to produce a democratic AI governance framework.
Argues that nurture (training data and procedure) matters more than nature (architecture) for sufficiently capable AIs, and that the Bitter Lesson implies convergent abstractions and even convergent values.
Traces an intellectual path from ontology maps to cyborgism via Neuralink, arguing that augmenting humans into superintelligence is more promising than trying to align a separate external superintelligence.
Proposes enhancing human intelligence via brain-computer interfaces by connecting electrodes to an artificial neural network, potentially enabling gradual mind uploading and helping humans stay competitive during the AGI risk period.
Extends the REPL agent framework to propose a solution for ARC’s Eliciting Latent Knowledge problem, learning an ontology map between machine and human state spaces.
Introduces a formal type-theoretic framework for agents based on read-eval-print loops, defining agents as 3-tuples of Read, Eval, and Print functions.