Goal Hijacking
How Intrinsic Motivation Can Derail Goal-Directed Behavior
The Hammer and the Nuke
Why AI Robustness is a Necessary Condition for AI Alignment
Turing versus Scherbius
The card game
A Tutorial on Ciphers
From Caesar to a Simplified Enigma
Intelligent Immunity
What if intelligence evolved within the immune system?
Approximate Reasoning: Computational Rationality in Humans and LLMs
Why 'Flaws' in Thinking Might Be Optimal Strategies for Incomplete Models
Typicality
In high dimensions, typical is unintuitive
The MAP solution is not the best solution?
MAP produces solutions that are not typical
Projection into the typical set: PITS
A new approach to solving inverse problems
Diffusion posterior sampling
A review of recent work
Typical set of arbitrary distributions
Constructing the typical set for arbitrary distributions
Typical set of arbitrary distributions
Proving that the image of the typical set is the typical set of the image
Score functions, denoising and diffusion
Diffusion is just stacked denoising score matching!
The behaviour of neural flows
Neural nets can struggle to learn very simple flows.
Pilot waves
The classical alternative to quantum mechanics?
Graph enumeration
The orderly enumeration of graphs
Language models are all you need
for SMILES-based chemistry
Automatic integration
Like automatic differentiation, but for integration
Requests for research
Some ideas from my masters.
The fable of the caterpillar
A fun intro to my masters topic; abstraction for efficient reinforcment learning.
The advantages of backward reasoning.
A simple exploration of what can be gained by reasoning backwards from your goal.
Real time bandits
Rewarding a twitter bot is more complicated than I imagined.
Inference via interference
Learning by controlling the propagation speed of signals.
Automated science
Model-based reinforcement learning, symbolic AI, and the limits of efficient learning.
Unsupervised skip connections
How can we use ladder nets in the unsupervised setting?
Covering letter for my first PhD proposal
Principles of neural design, automated science, general reinforcement learning
Visualising dataset alignment
Dataset alignment represented as tensor networks
Conserved complexity
A conservation law for algorithms?
Deep learning
An intro to deep learining for COMP421.
Representations within linear algebra
We can use linear algebra to represent; linear operators, algebras, computations, symmetry and more.
A functional type of non-linear algebra?
Functional programming plus neural network architectures.