Evergreen
Timeless concepts in AI safety — always fresh material on core ideas that won't go out of date. Useful on slow news days or as scaffolding for series.
Genocide Bingo
exurb1a's darkly funny primer on superintelligence risk — why “it'll probably be fine” isn't a plan. A decade old and still one of the best cold-opens for explaining the stakes to a general audience.
What Is an LLM? Large Language Models Explained
A Large Language Model (LLM), at its core, is an incredibly well-trained prediction machine, that can consistently predict the next appropriate word in a sentence. LLMs are trained on massive amounts of data, from books, to internet forums, to articles. GPT-2 for example, was trained on 8 million sources that comprised of 40gb of text. Once given this information, they are trained to recognise patterns: which words follow which, how are sentences structured, what sentences read normally versus abnormally. Over time they gain an exceptional understanding of linguistic structures to the point they can form full sentences that make complete sense given the question/prompt they are asked.
Neural Networks Explained
Neural networks are machine learning models that mimic the complex functions of the human brain. These models consist of interconnected nodes, or neurons, that process data and learn patterns directly from data without relying on pre-defined rules, enabling them to perform tasks such as pattern recognition and decision-making. Neural networks are built from several key components: • Neurons: the basic units that receive inputs; each neuron is governed by a threshold and an activation function. • Connections: links between neurons that carry information, regulated by weights and biases. • Weights and biases: these parameters determine the strength and influence of connections. • Propagation functions: mechanisms that help process and transfer data across layers of neurons. • Learning rule: the method that adjusts weights and biases over time to improve accuracy. Learning follows a structured, three-stage process: 1. Input computation: data is fed into the network. 2. Output generation: based on its current parameters, the network generates an output. 3. Iterative refinement: the network refines its output by adjusting weights and biases, gradually improving its performance on diverse tasks. Adapted from GeeksforGeeks, "Neural Networks: A Beginner's Guide".
Instrumental Convergence: Why Evil AI and Good AI Want the Same Things
Instrumental convergence is the idea that we can predict some of what an advanced AI will try to do without knowing its actual goals. As Robert Miles explains, it rests on just one main assumption: that an artificial general intelligence will behave like an agent -- something that has goals and takes actions to achieve them. The key distinction is between terminal goals (what an agent wants for its own sake) and instrumental goals (what it pursues as a means to something else). Money is the classic example -- whatever a stranger’s terminal goal, offering them money helps, so ‘acquire money’ is a convergent instrumental goal across almost every goal. The same logic predicts that a capable agent will pursue self-preservation (you can’t achieve your goal if you’re switched off), goal-preservation (it resists having its goal changed, because a new goal means the current one goes unmet), self-improvement (a smarter agent is better at any goal), and resource acquisition (matter, energy, and compute help with nearly any plan). None of these need be terminal goals -- they emerge as side effects of competently pursuing almost any objective, which is why a misaligned AGI could be dangerous even without any malice. The summary above distils Robert Miles’ explainer embedded below -- watch it for the full argument, including the paperclip-maximiser and ‘stamp pill’ thought experiments.
Recursive Self-Improvement: How AI Could Bootstrap Its Own Intelligence
Recursive Self-Improvement (RSI) in AI is the idea that an AI system will become sufficiently autonomous and capable of coding that it would become possible for it to improve itself without any human intervention. Thus, it would enter a feedback loop: each improvement makes the next model sharper, more efficient and slightly smarter, which will allow that next model to then do an even better job at making a better model. This progress would then compound, with each iteration growing stronger at a faster rate than the last. Ultimately, if this capability is achieved, and it can run its own research independently, a model could bootstrap itself from roughly human-level intelligence to vastly superhuman capabilities in a relatively short period of time. This idea is often referred to as the 'intelligence explosion', as the rate of improvement between models scales with each iteration's intelligence.
The Alignment Problem
The alignment problem is the core challenge of AI safety: ensuring AI systems pursue goals compatible with human values. As AI systems become more powerful and autonomous, the consequences of misalignment grow. Alignment is not a solved problem, and current approaches are partial solutions.
Mesa-Optimization
Mesa-Optimization is the situation that occurs when a learned model (such as a neural network) is itself an optimizer. In this situation, a base optimizer creates a second optimizer, called a mesa-optimizer. The primary reference work for this concept is Hubinger et al.'s "Risks from Learned Optimization in Advanced Machine Learning Systems". Example: Natural selection is an optimization process that optimizes for reproductive fitness. Natural selection produced humans, who are themselves optimizers. Humans are therefore mesa-optimizers of natural selection. In the context of AI alignment, the concern is that a base optimizer (e.g. a gradient descent process) may produce a learned model that is itself an optimizer, and that has unexpected and undesirable properties. Even if the gradient descent process is in some sense "trying" to do exactly what human developers want, the resultant mesa-optimizer will not typically be trying to do the exact same thing. Adapted from the LessWrong "Mesa-optimization" entry (riceissa, Rob Bensinger, Ruby, et al.; last updated 20 September 2022).
Corrigibility
Corrigibility is the property of an AI system accepting correction or shutdown from humans. A corrigible AI system is safer because it can be corrected if its behavior drifts from intentions. However, building corrigible superintelligences is non-trivial: a sufficiently powerful agent has incentives to resist correction to preserve its long-term goals.
These are concept explainers, not news — the ideas here are as relevant in five years as they are today. Each one links to a canonical source: the foundational paper or an established expert resource. Every source link has been verified live.
Angles and key points follow the same standard as the daily feed: truthful, citable, ready for a script. If a claim matters for your video, click through and check the source first.