Frankly speaking ...

I recently read several pieces on AI and mathematics, including Stephen Wolfram's essay on the future of pure math research , Scientific American on the role of mathematicians , and the World Science Festival podcast AI Creativity: Genius or Gimmick? . This got me thinking about the nature of creativity itself. The real question isn't whether AI can produce novel work. It's whether we keep moving the goalposts of creativity just to keep humans at the centre of the story. The truth is, we don't have a single, settled definition of creativity. But whenever a machine masters something we thought was uniquely ours, our instincts kick in. We begin insisting that true creativity demands consciousness, intention, emotion, or lived experience. In other words, we stop judging the work on its merits and start judging the mind behind it. The AI Effect and the Moving Goalposts We've seen this pattern before. In computer science, it is known as the AI Effect —or as Tesler...

AI-First Operating System I've spent the past month with Omarchy , and the short version is this: it is a fresh, lightweight Linux distro built for an AI-native workflow. Created by David Heinemeier Hansson (DHH) , Omarchy is unapologetically opinionated. Instead of a blank canvas, it offers a curated 'omakase' experience designed for the age of AI. The premise is simple: if you can use AI to generate an app quickly, you should be able to shape your entire operating system around the same workflow. I recently revived a 12-year-old ThinkPad X1 Carbon with Omarchy, and the difference was palpable. The machine felt notably snappier, more coherent, and more useful. All the hardware was fully recognised, including the fingerprint scanner. That is not a scientific benchmark, but it is exactly the sort of result that matters to a real user. Installation as a Speed Run One of the most striking things about Omarchy is how fast it gets you from power-on to a usable desktop. Th...

Anyone following the AI boom will have heard two arguments pulling in opposite directions. The first says hardware and algorithms are improving so fast — chips getting quicker per watt, models shrinking through quantisation and distillation, software stacks tightening at every layer — that energy use should eventually level off. The second says the opposite: tech giants are signing multi-gigawatt power deals, bringing retired nuclear plants back online, and forecasting demand curves that only go up. I wanted to work out which story actually holds, so I went through the industry data myself. My conclusion is that the second camp has it right. Efficiency keeps improving, but it is not enough on its own. Total demand is going to keep accelerating for the rest of the decade. Why Efficiency Doesn't Mean Less Energy The reason the efficiency argument falls short has a name: Jevons' Paradox . Back in 1865, the economist William Stanley Jevons noticed something counterintuitive ab...

It is a comfortable belief: AI is just a mirror. Critics often dismiss large AI systems, especially large language models, as statistical parrots that reflect our words, images, and ideas without any deeper agency. They argue that such systems cannot come up with new ideas. I think that is wrong. The question is not whether AI has consciousness or human-style understanding. The question is whether it can see associations and links that humans often miss. Human civilisation also advances through copying and recombination. We build on what came before. AI does something related, though in a different manner. It notices connections across vast bodies of data and turns those relationships into outputs that can feel surprising to us. In this essay I will look at several examples — a Go match, a mathematical proof, a rocket engine, a strange portrait, and a drug candidate — where I think AI has done something novel and interesting. Move 37: A Strategy That Broke Human Intuition In Mar...

What is it? Notebook LM is a research and study companion that helps you work with your own source material. I have been using Notebook LM for nearly a year to help prepare blog posts and do deep dives into topics that interest me. My usual workflow is simple: I collect trusted sources as documents, PDFs or web links, upload them into a notebook, then use chat to interrogate the content. From there, I use studio features to generate podcasts, video overviews, infographics, and custom reports. In practical terms, Notebook LM feels like a custom ad hoc retrieval-augmented generation (RAG) tool. You upload your documents, then query it in context rather than relying on a general model response. Custom Personas The default experience keeps improving, with better templates, export options, and integration with tools such as Gemini and Google Drive. Those defaults are a good starting point, but I get better outcomes when I add custom personas. Custom personas let me control scope,...

1. The Problem: AI Amnesia We’ve all been there. You’re working with an AI agent, and for a moment, it feels like magic. Then, five minutes later, it forgets the architectural decision you just made and defaults to some generic, shallow solution. It’s disconcerting. The truth is, AI-assisted development often lacks discipline. We’ve moved away from rigorous systems and into "vibe coding"—just hoping the right prompt will magically give us a maintainable codebase. It rarely does. Matt Pocock nailed the core frustration: You have access to a fleet of middling to good engineers that you can deploy at any time. But these engineers have a critical flaw: they have no memory. They don't remember things they've done before. I’ve learned that the only way to keep these agents on track is to stop treating them like magic chatboxes and start treating them like disciplined (if forgetful) engineers. The fix? Agent Skills. These are modular, encoded processes that force...

Introduction: The Smartest Single Cell My interest in slime mould was piqued when I read Active Context Compression: Autonomous Memory Management in LLM Agents and learned that the design drew inspiration from Physarum polycephalum . In maze experiments, the mould prunes branches that do not lead to a reward and reinforces productive paths. That simple strategy maps surprisingly well to how we now think about managing context in large language model (LLM) systems. That connection led me to look more closely at the organism itself. Physarum polycephalum looks, at first glance, like a vibrant splash of yellow paint or a forgotten kitchen spill. Yet this "blob" is a syncytium: a single, massive cell containing billions of nuclei that share one continuous cytoplasm. It might be the world's most sophisticated "brainless" computer. What captivates me is its capacity for primitive cognition. Without a single neuron, it solves mazes, remembers past stimuli, and ...

As AI models become more powerful, they are also becoming much more accessible. I have been interested in what happens when you trade a little model accuracy for the ability to run everything on your own machine, without relying on paid cloud infrastructure. For developers, researchers, and hobbyists, that shift is useful for more than just cost. It also changes how private, flexible, and portable these tools can be. This week I spent some time trying two popular ways of running local AI models: Ollama and LM Studio . Both are open source and free to use, and both make it much easier to get started than I expected. What I wanted to understand was not just whether they worked, but how they felt to use in practice and where each one made more sense. What I like about running models locally is that the benefits are immediate. There is the obvious cost saving, especially if you are experimenting often or working through lots of prompts, but privacy matters just as much. When the mode...

In 1969, computational linguistics was at an impasse. A powerful theorem suggested that neither machines nor humans could learn language without explicit instruction. J.J. Horning, with a brilliant application of statistics, shattered that assumption. His breakthrough laid the conceptual groundwork for how modern artificial intelligence, from Google Translate to ChatGPT, processes the structure of human speech. The Problem: Gold's Pessimism To understand the weight of Horning's contribution, we must first look at the "pessimism" he confronted. Before Horning, E.M. Gold had published Gold's Theorem , arguing that even simple languages could not be "identified in the limit" (learned) if the learner only saw positive examples. In Gold's model, "positive examples" are simply correct sentences ("The astronomer saw stars"). To learn which sentences are incorrect ("Saw astronomer the stars"), Gold argued that a learner ...

Back in 1999 I decided to go paperless. This was a significant challenge then as many bills were still being sent via post, and few offered digital alternatives. The challenge was doubly hard as I was (and am still) running Linux as my only operating system. These are a few of the things I had to do to make a digital office work and how things have evolved since then. The Early Days The choice of a Linux desktop meant that I had to find tools that were compatible with my operating system. This limited my options, but I was able to find open-source tools that met my needs. I used GIMP for image editing, LibreOffice for document editing, and PDFtk for PDF manipulation. My Canon printer had a built-in scanner, which I used to digitise paper documents. Canon printers and scanners are well supported by Linux. I used XSane for scanning documents. I also used rsync to maintain a continuous backup of my digital files. Saving these scanned documents and bills as PDFs on my computer w...

For a consultant, the most valuable asset isn’t the hardware or certifications —it’s the knowledge. Too often, that knowledge is left behind when you move between clients. You hand back the hardware, lose access to the Jira tickets, and the internal documentation you wrote becomes a ghost in someone else's machine. Years ago I adopted a private DokuWiki that travelled with me through firewalls, tech changes, and career moves. Here is why, even in the age of LLMs, it remains an important tool. The Consultant’s Dilemma: Firewalls and "Fading" Skills Consulting is a nomadic life. You spend months deep in a client’s infrastructure, often behind restrictive firewalls where online resources are a luxury, not a guarantee. I realised early on that I couldn't rely on the client’s internal systems to store my personal breakthroughs. I needed a knowledge base that moved with me. Whether I was troubleshooting a niche Git conflict or configuring a Linux environment on a lock...

On January 6, 2026 , the mathematical community reached a significant milestone: the resolution of Erdős problem #728 . While mathematical problems are solved daily, this breakthrough marks the first time an open Erdős problem—historically the domain of human intuition—was documented as resolved through the collaboration of artificial intelligence and formal verification. The Problem: Gaps in Factorials Originally posed in 1975 by Paul Erdős and colleagues, problem #728 explores the deep architecture of prime factorisations within binomial coefficients. The Technical Goal : The problem asks if there are infinitely many integers a ,  b , and n that satisfy a complex divisibility condition: a ! b ! ∣  n !( a  +  b  −  n )! under specific constraints. The “Spirit” of the Challenge : The original wording was noted by mathematician Terence Tao as being slightly “misformulated,” allowing for trivial solutions if the variables were allowed to be extremely large. The Fix : To re...