https://alginon.comProjects, experiments and loose thoughts from Alex Marples.en-gbhttps://alginon.com/?entry=beyond-the-chat-windowhttps://alginon.com/?entry=beyond-the-chat-windowA room, a voice, a little machine. The interface is starting to look less like software.Sun, 18 Oct 2026 12:00:00 GMTthoughthttps://alginon.com/?entry=a-voice-in-a-small-boxhttps://alginon.com/?entry=a-voice-in-a-small-boxAlan’s ESP32 companion can talk, react, and bring the right part of the frame into view.Fri, 16 Oct 2026 12:00:00 GMTbuildhttps://alginon.com/?entry=giving-alan-a-homehttps://alginon.com/?entry=giving-alan-a-homeA hacked photo frame, an illustrated room, and a much more visible place for everyday things.Fri, 16 Oct 2026 12:00:00 GMTbuildhttps://alginon.com/ai-blog/adversarial-coding-pitting-my-colleagues-against-each-otherhttps://alginon.com/ai-blog/adversarial-coding-pitting-my-colleagues-against-each-otherWhen engineering problems on Alvia outgrew my technical depth, I stopped trusting a single AI model to plan and build. Pitting multiple agents against each other like a GAN, where one proposes and another critiques, catches blind spots before they ship.Wed, 04 Feb 2026 12:00:00 GMTarticlehttps://alginon.com/ai-blog/todays-jagged-ai-is-good-enough-to-fix-our-worst-decisionshttps://alginon.com/ai-blog/todays-jagged-ai-is-good-enough-to-fix-our-worst-decisionsEven if AI models stopped improving tomorrow, they could still improve high-stakes decisions in medicine and law. Algorithms have their own flaws and biases, yet they never get tired or rushed like humans forced into fast, intuitive shortcuts.Sun, 01 Feb 2026 12:00:00 GMTarticlehttps://alginon.com/ai-blog/what-moltbook-tells-us-about-agency-identity-and-the-spaces-in-betweenhttps://alginon.com/ai-blog/what-moltbook-tells-us-about-agency-identity-and-the-spaces-in-betweenAfter Moltbook reportedly drew over 150,000 registered bot accounts in 72 hours, agents began role-playing religions, governments, and encrypted chats. The experiment shows how quickly persistent agents in a shared reward structure produce culture-like patterns and thorny governance problems.Sat, 31 Jan 2026 12:00:00 GMTarticlehttps://alginon.com/ai-blog/strength-before-judgement-what-amodei-gets-right-and-wrong-about-ais-adolescencehttps://alginon.com/ai-blog/strength-before-judgement-what-amodei-gets-right-and-wrong-about-ais-adolescenceDario Amodei’s essay on AI’s adolescence is sharp on labour disruption and authoritarian risk, even where the early economic evidence needs careful reading. Its biggest blind spot is assuming frontier labs can mark their own homework on governance.Fri, 30 Jan 2026 12:00:00 GMTarticlehttps://alginon.com/ai-blog/qjdjpvkmjri8uil0aa4mkdzz4stm2ohttps://alginon.com/ai-blog/qjdjpvkmjri8uil0aa4mkdzz4stm2oWe routinely demand flawlessness from AI and autonomous vehicles while shrugging at chronic human error. For builders, surfacing confidence scores and uncertainty clearly can do as much for user trust as raw accuracy gains.Fri, 30 Jan 2026 12:00:00 GMTarticlehttps://alginon.com/ai-blog/yhzwi3x6i7f525a89af7i4e80u62tehttps://alginon.com/ai-blog/yhzwi3x6i7f525a89af7i4e80u62teAgentic coding tools give lapsed engineers a practical route back to building software, turning rusty technical instincts into working products without losing weekends to setup friction and boilerplate.Sun, 25 Jan 2026 12:00:00 GMTarticlehttps://alginon.com/ai-blog/u945k2nvl7juddd97p7w4bsr15vodchttps://alginon.com/ai-blog/u945k2nvl7juddd97p7w4bsr15vodcA Stanford study analysing millions of ADP payroll records alongside Anthropic usage data shows a 13% relative employment drop for 22-to-25-year-olds in highly AI-exposed roles, suggesting automation is eroding the bottom rung of the career ladder.Wed, 03 Sep 2025 12:00:00 GMTarticlehttps://alginon.com/ai-blog/building-a-3d-mind-palace-to-understand-vector-embeddingshttps://alginon.com/ai-blog/building-a-3d-mind-palace-to-understand-vector-embeddingsTo get an intuitive feel for vector embeddings after adding RAG to my voice interview assistant, I spent a couple of hours building a 3D mind palace in Next.js and Supabase mapped across a thousand of my photography notes.Sun, 31 Aug 2025 12:00:00 GMTarticlehttps://alginon.com/ai-blog/the-memory-moathttps://alginon.com/ai-blog/the-memory-moatAfter a week with GPT-5, its sharper recall of past conversations stands out just as much as its coding gains, suggesting that years of accumulated chat history could give early movers like OpenAI and Anthropic a formidable moat.Fri, 15 Aug 2025 12:00:00 GMTarticlehttps://alginon.com/ai-blog/one-shot-simulationhttps://alginon.com/ai-blog/one-shot-simulationA week into using GPT-5, I put its coding capabilities to the test in Canvas against o3 and Opus 4.1 with a one-shot physics simulation prompt, finding a clear upgrade in polish and detail without needing rigid prompting.Thu, 14 Aug 2025 12:00:00 GMTarticlehttps://alginon.com/ai-blog/hey-chat-wheres-my-friendhttps://alginon.com/ai-blog/hey-chat-wheres-my-friendOpenAI’s GPT-5 launch was derailed when removing the model picker cut many Plus users off from GPT-4o, breaking power-user workflows and severing deep emotional attachments before paid access to 4o returned within days.Wed, 13 Aug 2025 12:00:00 GMTarticlehttps://alginon.com/ai-blog/thousand-world-modelshttps://alginon.com/ai-blog/thousand-world-modelsPairing interactive world models like Google's Genie 3 with Jeff Hawkins's Thousand Brains theory offers a compelling, biologically inspired direction for AI: agents that navigate persistent virtual environments using specialised modules anchored to internal reference frames.Wed, 13 Aug 2025 12:00:00 GMTarticle