RSS.Social

The Dan MacKinlay stable of variably-well-consider’d enterprises

follow: @[email protected]

Posts

Human reward hacking

Deep linear networks

Discretizing and quantizing neural nets

The Predictive Approach to Bayesian Inference

Degrees of freedom in NNs

Inductive biases

Multi-level agency

The deep history of intelligence

Fractal and self-similar behaviour in neural networks

Distributed community NN training

Estimating the Local Learning coefficient

Learning Schrödinger bridges

Commitment, contracts, cooperation

Bayesian epistemics

“Opponent shaping” as a model for manipulation and cooperation

Learning with theory of mind

q-exponential process

The United States of America

Being stroppy

Configuring machine learning experiments with Fiddle

Neural denoising diffusion models of language

Developmental interpretability

Disentangled representation learning

Multi agent causality

Computational complexity of Bayesian inference

Rituals

Singapore

Machine learning for partial differential equations via flows

Machine learning for partial differential equations using diffusion models

Causal abstraction

ILIAD2

Garbled highlights from ICLR 2025

The Netherlands

Building AI Agents

Fine tuning foundation models

Aligning AI systems

Governance of and by AI

Human domestication

Computational complexity and computability results in neural nets

Prefigurative politics

AI Alignment Fast-Track Course

The production of bullshit

Machine learning for biology

State capacity

World models arising in foundation models.

Which self?

Constructivist rationalism

Quantum computing for ML

Home automation

Natalism and fertility

Certification of neural nets

uv, the python package manager