Anapoly AI Labs

Category: diary

A public record of the development of our ideas

  • Working with AI

    Transparency label: AI-assisted. NotebookLM worked with Nishigaya’s post to outline the five levels.

    In a recent post on SubStack, Nori Nishigaya wrote:

    Being effective in your use of AI isn’t just figuring out the ideal prompt. There’s so much that goes into skillful use of AI. Done poorly, it can become a tool for chaos. This post lays out the framework that I’ve been exploring as a way of taming this chaos and getting really powerful results from AI.

    His framework captures the moment perfectly, and it resonates with my own learning and experience. Nishigaya’s beautifully clear explanation, which identifes the five levels outlined below, is a powerful (though deceptively simple) aid to thinking about the human | AI interaction.

    See also this infographic or this one or this one, all produced by NotebookLM.

    In my (small-scale) project to develop an agent for a recruitment agency, I am currently working at Levels 1 to 3. Once the harness has proved its worth, we’ll explore Level 4.

    The Five Levels of Working with AI

    Produced by NotebookLM from Nishigaya’s post

    Level 1: Prompt Engineering – The foundational skill of writing effective prompts.

    Level 2: Context Engineering – Shaping the environment and giving AI memory by providing custom context, files, and notes so the AI can formulate its thinking around your specific needs.

    Level 3: Harness Engineering – Creating skills, scripts, verification steps, and feedback loops to make AI output more repeatable, trustworthy, and verifiable.

    Level 4: Agentic Engineering – Deploying autonomous agents equipped with identity, goals, memory, and a strong harness, transforming them from a tool into an active, autonomous member of a team.

    Level 5: The Murmuration – The complex coordination of multiple autonomous actors (both humans and agents) moving together with emergent coherence and a shared purpose, drawing on principles from organizational design.

  • Momentum is gathering

    Transperency label: human only

    When Randolph Lalonde released the first “broadcast” of his Spinward Fringe space opera in 2008, AI was entirely in the realm of science fiction. Lalonde imagined AIs as autonomous agents capable of making complex decisions, sometimes with unintended consequences. He portrayed AIs both as companions (personal assistants worn on the wrist, shipboard intelligences) and as adversaries (corporate-controlled AI enforcers). His AI characters often formed deep, personal bonds with humans, raising questions about trust, dependency, and emotional attachment to machines. Fast forward to where we are today, and these AI themes are remarkably relevant to current debates.

    The pace of AI development is accelerating, and we are on the cusp of seeing autonomous agents as an everyday reality. Alongside that, a capable, open source AI can now run on a desktop computer. If I set up my own personal or business AI on an Apple Mac Mini, newly-emerging software tools will let me access it securely from almost anywhere through my laptop, tablet or smartphone, and thus probably from the smartwatch on my wrist as well.

    Tailscale is one such tool. It is like a private, secure internet just for you or your business, allowing you or your team to access files, apps, or AI tools from anywhere, without exposing sensitive data online. It does this by creating a secure, peer-to-peer mesh network in which devices connect directly to each other rather than through a centralised server. This network operates over but is entirely private from the internet.

    If you are running an AI locally on your own server, you can use Tailscale to connect to it securely from your laptop, tablet or phone. No data leaves your control, and you can use the AI without having to put your data at risk. The direction of travel is clear; momentum is gathering and it will be impossible to stop. The best society can do is try to steer it.

  • Ride the wave

    Transparency label:human only

    The message I am getting loud and clear is that there will be two groups of people: those who delay applying AI for whatever reasons, and those who put it to practical use and ride the wave of its evolution. The former will find themelves left behind, unable to catch up, and out of work. The latter will be valuable and sought after.

    The wave is now at the stage of agents; there seems little doubt that these will be operating widely within a year. It’s useful to know how to create and orchestrate them. It’s doubly useful if you understand business well enough to know where, when and how to deploy them. It’s triply useful if you have the understanding and skills to engineer the business information (on which agent success depends) in such a way that it is suitable for use by AIs whilst also being suitable for people. 

    Alongside that, it is essential to keep some friction in what people do. In helping me prepare this project brief for a small recruitment agency, my AI assistant went straight to the hosted solution. I directed it down the path of demonstration, piloting, and only then a fully engineered solution because I learnt that approach in my previous experience. The people who build enough friction into their use of AI so that they are always learning, these are the people who will continue to have value; others will simply become replaceable by agents. 

  • The pace is accelerating

    Gas Town would have been science fiction 18 months ago.

    Like so much of that genre, it has become reality. Steve Yegge created a real system architecture that has demonstrated it is capable of directing, harnessing, and controlling the software development work of 30 agents. The human overseer has 30 terminal windows open to monitor what is happening and enters commands at great speed throughout the process; hence his advice not to try it unless you are a top echelon software engineer not prone to insanity.

    But although it is chaotic, it works, and we can be sure that this demonstrator (the fourth version Yegge has produced over the past year) will soon evolve into a better version of a multi-agent software factory.

    As for the crazy Moltbook experiment, that uses OpenClaw. This is an open-source platform allowing users to run autonomous AI agents locally that integrate with messaging apps to execute tasks and interact with other agents on a network.

    Google just released Universal Commerce Protocol UCP), an open-source standard to power agentic commerce, providing a unified language for AI agents to discover products and execute secure transactions with businesses.

    Both OpenClaw and UCP utilize the Model Context Protocol (MCP) to enable agents to do things in the real world. OpenClaw agents can use MCP-based skills to potentially transact with merchants adopting the UCP standard.

    This is science fiction becoming reality before our eyes.

  • Beyond Words: The Rise of Large World Models

    Transparency label: AI-assisted

    We’ve spent the last few years watching Large Language Models get remarkably good at text. But something more fundamental is emerging: models that don’t just understand language, but reality itself. These are Large World Models, and they represent a different approach entirely. Where LLMs predict the next word in a sentence, LWMs predict the next state of a dynamic environment.

    There is an excellent documentary about Deep Mind and its extraordinary achievements both before and after acquisition by Google. Towards the end, it explains that Demis Hassabis’s team at Google DeepMind have developed Genie, a model that can generate playable games from a single image after having learnt game physics and mechanics entirely from videos. Fei-Fei Li’s World Labs and companies like Runway are pursuing similar approaches, training models on vast amounts of video to learn not how to arrange words, but how the physical world actually works: 3D geometry, object permanence, the laws of physics.

    The distinction between a Large Language Model which knows how scenes are described and a Large World Model which knows how they behave is important. The shift from static text processing to dynamic world simulation opens a path towards embodied AI: systems that can navigate physical environments rather than simply generate text. These will range from advanced robotics to fully interactive, virtual simulations.

  • An AI power-user’s perspective

    Transparency label: AI-assisted

    In his YouTube channel, Dylan Davis shares his thoughts on how we can make good use of AI. This infographic (produced by NotebookLM using only Dylan’s video as its source) shows five ways power users do so.

    Importantly, the fifth habit he highlights is not letting AI do all the thinking for us. If we delegate all our reading and summarising to AI, we risk losing the ability to focus for extended periods. We need to read a sufficient number of long-form documents ourselves, in order to retain that abiliy. Think of attention span as a muscle: we must exercise it by doing some hard work, or it will wither away.

    In another video, Dylan explains how we can create a GPT (ChatGPT) or a Gem (Gemini)to generate professional Nano Banana Pro image prompts. I used NotebookLM to write a standard operating procedure for this, based entirely on source material in the video. Users act as Creative Directors providing context, while the Gem serves as a Technical Prompter, transforming inputs into optimized JSON code to ensure consistent, high-quality visual outputs from Nano Banana Pro.

    JSON (JavaScript Object Notation): a lightweight data-interchange format based on a subset of the JavaScript Programming Language. It is easy for humans to read and write. It is easy for machines to parse and generate.

  • Methodology & tool-kits

    Transparancy label:human-only

    Yesterday’s discussion crystallised some ideas that I have been teasing at recently. Namely that we could develop our labs as a place where we:

    (a) develop a methodology and its associated tool-kits for the implementation of AI in an enterprise; and

    (b) use these in a synthetic environment to help business people gain confidence in moving forward with AI.

    Our value to clients would derive from the relevance and usefulness of the methodology and tool-kits.

    Methodology: a comprehensive system of principles, processes, and rules that governs a specific discipline to ensure consistency and repeatability. It serves as a structural framework defining what must be done and who is responsible. It is applied through specialized tool-kits consisting of templates, software, and techniques that enable practitioners to apply the theoretical standards to real-world tasks.

    To pick up on the point Dennis emphasised, an early part of the methodology might be analysis of how an enterprise actually works (rather than what the non-existent, incomplete or out of date documentation says). The process for this would involve eliciting information from people in the enterprise. A tool-kit would enable AI to be configured for that purpose within the specific context of that enterprise. 

  • Do AI models have a prompt appetite?

    About once a week, I use ChatGPT to pull out the main themes from the stack of newsletters I receive. This week I was surprised to be told that different AI models now have different prompt appetites. So I got the AI to explain what this means and the implications for Anapoly.

    ChatGPT’s response to the question “What is meant by different models now have different prompt appetites?”

    It means the era of “one prompting style works everywhere” is over.

    Models used to behave broadly the same: structure your instructions, specify roles, delimit sections, and you’d get predictable results. Now the frontier models diverge. They respond differently not just in quality but in preference for instruction style.

    A “prompt appetite” is the set of conditions in which a model performs best: how much structure it wants, how explicit you need to be, how conversational you can be, and how sensitive it is to examples, delimiters, or verbosity. In other words, each model now has its own diet.

    Gemini 3, for example, performs better with example-rich, lightly structured, conversational prompts. Too much rigid scaffolding and it feels choked. GPT-5.1, by contrast, still thrives on formal instruction and well-defined control layers. Claude it is noticeably sensitive to tone and framing. Even small changes in wording can push these models into noticeably different modes of reasoning.

    For Anapoly, that has two implications.

    First, your governance and context layers can’t be one-size-fits-all. If you’re orchestrating across multiple models—internal RAG, foundation models, domain-specific SLMs—you need a prompting strategy that adapts to the quirks of each one. Your orchestration layer becomes, in part, a “dietician”: matching tasks to the right model and the right prompt genre.

    Second, this strengthens your case for standardised, model-agnostic knowledge units. If models vary in appetite, the only way to keep outputs consistent is to control the inputs: modular content blocks, transparent context, and structured reasoning artefacts that can be rephrased per model without degrading meaning.

    In short: prompt appetite is emerging as a genuine design constraint. Ignore it and you get erratic behaviour; work with it and you get reliability.

  • A three-layer instruction set for use in ChatGPT Projects

    Transparency label: human only

    Building on recent work around the three-layer model of AI context, I spent some time today working with ChatGPT to write a full three-Layer Instruction set for use in Anapoly’s ChatGPT Projects.

    If it works as hoped, strong governance is now the default. The AI will challenge weak logic, ask for clarification when the ground is shaky, and keeps its tone plain and grounded. And for the occasions when I want it to bend the rules baked into it, I added an escape hatch: a temporary override, clearly marked, that lasts exactly one turn.

    It will be interesting to see how well it works.

  • Using local-only AI in a micro-enterprise

    Transparency label: AI-assisted

    I’ve added a briefing note in the Resources area that sets out an approach for a micro-enterprise to run AI entirely on local machines. This creates a system that is private, predictable, comparatively inexpensive, and easy to expand in small steps.

    The note explains how a three-layer model for controlling the behaviour of an AI ( business context at the top, governance rules in the middle, and task instructions at the bottom) can still apply even when everything sits on a laptop, and how a small software component (the orchestrator) can keep the whole arrangement predictable and safe.

    If you’re curious about “local-only AI” for a micro-enterprise, or wondering what it might look like in practice, you should find this a useful starting point.

    Read the note: Using Local-Only AI in a Micro-Enterprise