Blog

<h3>The Meaning of Words</h3><p>We all intuitively know that the meaning of a word depends on its context, typically the words around it. For example, sending an email that only contains the word ‘cat’ communicates nothing more than you are referring to a kind of animal.</p><p>Conversely, if I asked you to complete the sentence, "The cat sat on the...", there's a good chance you'd say 'mat', and so did my preferred Large Language Model (LLM) when I tried this out.</p><p>But why?</p><h3>How LLMs Learn Word Relationships</h3><p>As you probably know, LLMs are 'trained' by processing massive amounts of text from the internet, books, and other sources. It doesn't 'understand' words in a human sense (it doesn’t even see them); instead, each word is first converted into a unique numerical identifier, called a 'token'. During this training, the LLM's main goal is to learn how these tokens relate to each other through probability calculations.</p><p>For example, when given 'the cat sat on the...', the LLM processes these words as a sequence of tokens. It then calculates a probability for every possible next token (representing a word) in its vocabulary. Based on whether its prediction is right or wrong, the model adjusts its internal settings. This continuous adjustment helps it get better at calculating probabilities, essentially learning the patterns and relationships between words in the vast training data.</p><p>The core of this process lies in 'embeddings'. These are not the simple token numbers themselves, but rather complex numerical values assigned to each token. Initially, these embedding numbers are random. But as the model trains, it constantly adjusts these embedding values based on which words appear together. This allows it to ‘model’ the statistical relationships between words. Patterns form in these embeddings because language itself has patterns in how words relate to ideas and concepts. </p><p>For example, if you list animals, you're listing words that relate to the concept of an animal. While we might think of these relationships as 'concepts' or ‘semantics,’ the model itself doesn't 'understand' them like a human does. It's simply a complex system of statistical patterns that, when combined, influence the probability of one word following another.</p><h3>Changing the Meaning of Words</h3><p>The truly amazing piece is how these large language models dynamically adjust the meaning of words based on their immediate context. Think of it like this: the meaning of a word isn't fixed; it shifts depending on the other words around it in a sentence. This is why ‘context’ (the words you put in) is so important: it's fundamental to how every word gains its meaning through the LLM’s mathematical mechanisms, and this meaning, in turn, determines the next word produced.</p><p>For example, consider the word "cat."</p><ul><li>If I say, "The cat purred contentedly on the rug," the word "purred" influences the meaning of "cat" towards a domestic animal. The LLM adjusts the numerical representation (embedding) of "cat" to reflect this "pet animal" meaning.</li><li>Similarly, in "He's a cool cat with a smooth jazz tune," the words "cool" and "jazz" influence the meaning of "cat" towards a stylish or hip person. The LLM adjusts the embedding of "cat" to reflect this "stylish person" meaning.<br></li></ul><p>This dynamic adjustment of a word's embedding based on its context is crucial for predicting the next word. If the LLM has adjusted "cat" to mean "pet animal" because of words like "purred," it will then give a higher probability to words like "food" or "toy" as the next word. And if "cat" is adjusted to mean "stylish person" due to words like "cool," then "music" would become a more probable next word.</p><p>We don't actually ‘know’ what concepts (like pet animal or stylish person) words are being linked to, as the model builds these concepts through a statistical process based on how the words are associated with other words in the training data (which includes website content, book texts, and so on).</p><h3>How the LLM Predicts the Next Word: Back to our Cat</h3><p>This dynamic process means the meaning of each word is constantly shaped by its surroundings. When the LLM predicts 'mat' for "the cat sat on the...", it's because the preceding words, through statistical patterns learned during training, point towards concepts related to where a cat might sit.</p><p>Here's the truly mind-bending part: when the LLM predicts the next word, it doesn't re-process the entire sentence from the beginning. Instead, as each word is processed, the full context of the sentence up to that point is cleverly encoded into the numerical representation (embedding) of the last word.</p><p>So, in "the cat sat on the...", the final word 'the' holds all the contextual information that a cat is looking to sit on something. It is only this last, context-rich word (token) that the model uses to calculate the probability of the very next word.</p><p>This might seem counter-intuitive, but it's possible due to the enormous complexity of the mathematical 'embeddings'. Imagine representing each word's meaning not just in 3D space, but in thousands of dimensions (e.g., 3000). This allows for an incredibly rich and unique numerical representation for every possible thought, idea, or sentence context. </p><p>Each unique combination of words creates a unique value in the embedding of the last token that captures the full context, which then guides the prediction of the next. In other words, the full idea is captured in one set of numbers; this is possible because those numbers describe a relationship between that word and all the other words in the vocabulary.</p><p>This brings us back to the idea of context. A word means what it means by its relationships to other words. In the case of large language models, that full context, that meaning of what a word means in the context of everything that's come before it, is captured in this number for each and every token, and that is what drives the next word in the sequence.</p><p>It’s an amazing, almost mind-bending idea, but the key takeaways are far more pragmatic:</p><ul><li>The models never retrieve data; they only provide probabilistic outputs.</li><li>Where facts matter, you need to verify the outputs.</li><li>Driving that probabilistic output to be most relevant and valuable is largely dependent on how well the model was trained and the words you put in to drive the context for the next word in the sequence.</li><li>The workflows that can most benefit from this are those where creativity and the surfacing of previously unseen patterns and concepts are valuable.</li></ul>

The Cat Sat on the Mat: How LLMs Work

<p><br></p><h3>The Convergence of Knowledge Work and Software Coding</h3><p>A fundamental convergence is happening at an accelerated pace: the worlds of software development and everyday office productivity are merging. AI-native coding tools are not just helping developers write code; they are optimizing the creation of digital ‘artefacts.’ This is more relevant to your daily work than you might think.</p><p>Consider the daily function of a business professional. We coordinate people, effort, and resources often to produce specific artefacts: proposals, reports, strategic plans, and critical emails. The purpose of these artefacts is to drive action or change the state of the business. In essence, we are orchestrating workflows to produce a predictable outcome. This has a parallel to software development. While the technical skill sets have been distinct, the underlying purpose, the execution and scaling of knowledge work, is virtually identical.</p><p>For those without a software development background, it can be illuminating to view coding in a different light. At its core, software development is the creation of a set of documents (source code) that adhere to a perfectly logical and unambiguous set of rules. These rules seek to eliminate the potential for misinterpretation inherent in human language, allowing a machine to execute a predictable outcome. It is precisely this need to translate human intent into a machine-readable format that made software development the first and most fertile ground for Large Language Models (LLMs).</p><p>The core function of LLMs in coding is translating ambiguous human language into the precise syntax a machine understands.</p><h3>Leaders, Take a Lesson Here</h3><p>So, what does this mean for a business leader?</p><p>It means this same translation capability is rapidly being embedded into your own toolset, from spreadsheets to presentation software. The AI will be there to translate your teams intention into new artefacts and automated workflows. The primary bottleneck will no longer be technical execution, but the scope of your ambition. The limiting factor will be an inability to reimagine what you and your team can achieve with these new capabilities.</p><p>This is why I study the adoption of AI in software development: not to become a coder, but to understand the mental models of developers. I am learning how they think, manage complexity, and ensure products are shipped reliably. Leaders should be doing the same. </p><p>Soon, you won't need to write code to orchestrate complex digital work. But this capability will be wasted if your thinking remains anchored in today's workflows. The challenge is to start thinking like a developer: systemically, logically, and with an eye toward automation and scale.</p><h3>It Seemed Kooky...For a Few Weeks</h3><p>As a practical example of this convergence, I am writing this article not in a word processor, but in an AI-powered code editor called Cursor. I am experimenting with the tools of tomorrow to change how I work today, and the initial results are revealing.</p><p><em><strong>[Edit:</strong> Since I first wrote this article the convergence has accelerated. All of the major vendors are landing 'coding tools' on desktops; coding tools that can do far more than 'code', coding tools that are enabled with agentic capabilities - meaning the work of artifact creation is getting automated at the output layer, not just through inputs from Large Language Models. Using Cursor, or Claude Code, or OpenAI's Codex, or Google's CLI for Gemini or similar, for Knowledge Work, and not just software coding is very quickly becoming not so 'Kooky'.</em></p><h3>Getting onto the Front Foot</h3><p>To navigate the path forward, leaders need a framework for thinking, not a feed of tech announcements. </p><p>If you are focused on separating the signal from the noise and understanding the strategic implications of AI, I invite you to <a data-mce-href="https://calendly.com/keytatkinson/ai-literacy-program-for-leaders-introduction" href="https://calendly.com/keytatkinson/ai-literacy-program-for-leaders-introduction" target="_blank" rel="noopener">connect</a> and explore how growing your AI Literacy can shift you to a proactive posture on AI risks and opportunities.</p><p><br></p>

What Business Leaders Can learn From Coders

<h3>Talking 'Bout My Generation</h3><p>For leaders of my generation, the path to 'AI nativeness' isn't theoretical—it's experiential. True understanding requires getting hands-on.</p><p>For those of us who went to school before the birth of the internet, research was a physical process: a library, a card catalogue, a Dewey Decimal number, and occasionally, a microfiche reader. A decade later, this shifted to keyword searches on AltaVista or early Google.</p><p>Today you ask your computer a question using AI, and if you're like me you don't use the keyboard, you talk to it.</p><p>Often, the end result of an investigation is a new kind of output, such as a prototype website. I find this is often the most effective way to iterate on an idea and deepen my understanding. This isn't about becoming a developer. It's about collapsing the time between idea and execution. I can now test a hypothesis or visualize a new customer experience in minutes, not weeks. </p><p>This speed of iteration is a fundamental shift in how we can approach innovation and de-risk new ventures. That's a 10 to 15 minute process end-to-end, mostly running in background and by using my voice it requires no more than a minute or two on the keyboard.</p><p>I don't have to imagine where my idea can go, I can see it and immediately iterate....immediately. This represents more than an increase in speed; it's a fundamental change in my method of working and thinking.</p><h3>Mindset before Mastery</h3><p>Setting the technology aside, here's the key point. I can accomplish significantly more today than I ever could before. If I'd stuck with the way I used to work in my twenties, I'd be irrelevant in the workforce.</p><p>If your mindset is that your future tools of productivity are Word, Excel, Outlook, Teams or some kind of slightly more contemporary tooling - think Slack, Notion, all of them with AI bolted in; it's time to extend your thinking. This bolting in is important and will continue to expand, but it is just the early response from existing vendors as more extensive changes start to wash through.</p><p>Relying solely on the AI features embedded in our current software stack (Microsoft 365, Slack, etc.) is a defensive posture. It helps with incremental efficiency, but it binds us to old mental models and blinds us to the offensive opportunities of building entirely new, AI-native processes. </p><p>The real disruption will come from rethinking workflows from the ground up, not just enhancing old ones.</p><p>Having been deeply hands on with various AI tooling and transformed my own workflow and processes to leverage them, and having seen the remarkable uplift in output I am now capable of, my urgent message to you is this: it is a strategic imperative to understand what is coming, not by reading about it, but by experiencing it.</p><h3>Hands On Experience Leads to Strategic Insight</h3><p>I really believe the best way to understand the significant impact and opportunity of AI is to get hands on and build your own insights and understanding through experience. It's one thing to hear about a magic trick. It's a completely different experience to see one.</p><p>However...and maybe you can relate to this...more than one of my colleagues has talked about being too busy to dig in and learn more...don't fall into that trap. There's no room for complacency.</p><p>Don't get me wrong, I'm not a wide-eyed optimist for how great AI is. There's proper challenges to deal with. Leadership in this era requires moving from the passenger seat to the driver's seat. You cannot effectively lead this transformation from the sidelines.</p>

AI Leadership: Why Hands on Experience Matters

<h3>Functional Fixedness<br></h3><p>Adopting Generative AI is in many ways a human transformation. This is because we tend to build and stick to mental models of the world and the tools within it to avoid cognitive overload. This is illustrated by the concept of "functional fixedness," famously shown in the 'Candle Problem'.</p><p>In the Candle Problem, subjects are given a candle, thumbtacks, and matches and asked to fix the candle to a wall so it won't drip. The solution involves emptying the thumbtack box, tacking it to the wall, and then placing the candle on the box. The difficulty lies in seeing the box not just as a container but as a platform (which requires overcoming the fixed idea of its usual function).</p><p>This tendency towards 'functional fixedness' is a barrier to truly leveraging AI. We see an AI tool through the lens of its most obvious, initial functions, like writing text or summarizing, and mentally categorize it as such, much like the subjects who saw the box only as a 'tack container'.</p><h3>Status Quo &amp; Loss Aversion</h3><p>This initial, narrow view then feeds a second, more powerful bias: Status Quo Bias. If an AI is perceived as just a slightly better drafting tool, why should one change an entire workflow for what appears to be a marginal gain? The resistance to change feels logical because the perceived benefit is small and contained.</p><p>This inertia is compounded by Loss Aversion, where the pain of a perceived loss looms larger than the pleasure of a potential gain. The effort to learn a new process, and the fear of ceding control or devaluing one's existing skillset, feels far more threatening than the incremental benefit of the 'better summarization tool' we believe the AI to be. Consequently, the true, game-changing opportunities of using AI are not considered, because our initial functional fixedness has already devalued the prize.</p><h3>Reframing</h3><p>Overcoming these cognitive biases requires a fundamental reframing of the technology. We must move beyond viewing AI as a deterministic tool for retrieving known information, (using it like a knowledge base). Much of its true power lies in its non-deterministic nature as a creative engine, capable of generating what may appear as novel solutions.</p><p>The opportunity, therefore, is not just in optimizing existing processes, but in identifying entirely new problems that it can solve when leveraging your unique competencies, data, and networks to create new forms of value.</p>

Are You Biased Against AI?

<h3>What is Your AI Mental Model Costing You?</h3><p>In a sense, every technology wave begins with confusion and a fight between the present and the past.</p><p>When new technology emerges, it's natural for us to understand it based on what we're already familiar with. These 'mental models' can sometimes make it challenging to see the new opportunities it presents. Over time, as our perspective shifts, what was once unclear often becomes obvious.</p><p>For example, when digital photography was invented by Steven Sasson at Kodak, it was viewed primarily through the lens of how to support the existing photo printing business. Others, not encumbered with that consideration realized that the attributes of digital photos could become a key enabler of new value.</p><blockquote><p>Kodak created a digital camera, invested in the technology, and even understood that photos would be shared online. Where they failed was in realizing that online photo sharing was the new business, not just a way to expand the printing business.<br>HBR Article: https://hbr.org/2016/07/kodaks-downfall-wasnt-about-technology</p></blockquote><p>New technologies offer more than just new solutions to old problems; they create entirely new possibilities. If we can recognise and move past ingrained mental models that hold us back and learn to see things differently, we can unlock a new technology's full potential early on.</p><p>Those who examine and refresh their mental models first, move faster to leverage the new technology's advantages and opportunities.</p><h3>Are AI Hallucinations Really a "Bug"?</h3><p>We see this happening today with Generative AI. For example, it's common to view factual inaccuracies, or hallucinations, as a flaw in the system. This perspective makes sense if we approach the AI like a search engine or a database, where the main goal is to retrieve a specific, correct answer.</p><p>Yet, these are Large Language Models, not Large Language Knowledge Bases. That's not to say they can't support that goal; but there are limitations and a different mental model required. This "Knowledge Base" mental model is like Kodak seeing digital cameras primarily as a way to sell more printing (a sane response in their mental model). This way of thinking, traps users within the old rules, blinding them to risks and new possibilities.</p><p>In fact, it's a recognised danger of Generative AI how convincing it can be when presenting incorrect information. One that is compounded by a single interface for use cases of both factual enquiry and creative intent. I even ran into this problem when doing research on historical examples of technology transition for this article. A misconception that I already had, which my AI tool was parroting back to me, would have been easy for me to accept as fact, because it aligned with what I already believed.</p><h3>Turn a "Bug" into a "Feature" with a New Mental Model</h3><p><br>A fresh way to think about Generative AI is to see its varied outputs not as mistakes, but as a source of boundless creative inspiration. Since the AI's outputs are based on a model of the world, not the world itself, its ability to provide 'alternative realities' can, when used correctly, be a major advantage.</p><p>Consider for example, a planning process: you can easily use AI to generate 5 plans in parallel, analyse their differences, identify your preferences among the versions, and use the insights you discover to optimise the next iteration.</p><p>When viewed this way, the AI's variability of responses becomes a feature that helps you produce higher quality results and innovate faster. It's a process that improves the final outcome and the insights you achieve improve your own skills and effectiveness.</p><h3>Navigating a Blurred Reality</h3><p><br>New technologies add to our options rather than replacing them. Most of us still have printed photos in our homes - my kids have polaroid cameras! The key is knowing which approach is best for the situation and switching our mindset for the task.</p><p>For example, we can treat AI as a research assistant that needs fact-checking, or as a creative partner for brainstorming new ideas. Developing this flexibility allows us to use new tools to complement and improve our existing ways of working.</p><p>Ultimately, the real challenge is not just using the tools differently, but fostering AI ready mental models to complement, enhance and "yes" sometimes replace our existing ways of thinking and working.</p><h3>Try it Out</h3><p><br>Of course, it’s one thing to hear about a new way of working, it’s quite another to experience it. So here’s a way to get started.</p><ol><li><strong>Pick a Process</strong>: Think about a process you follow today that requires you to create an artifact (e.g., a plan, report, proposal, email). Choose something that is in part a creative endeavour that defines a future possibility rather than just reporting on the past.</li><li><strong>Generate in Parallel:</strong> Craft a prompt for your artifact and run it in three parallel instances of your preferred AI assistant to get multiple unique drafts.<br>Analyse the Variations: In a new chat window, paste all three of your initial drafts into this chat and instruct the AI to act as an analyst. Ask it to compare and contrast the drafts you provide, highlighting the key differences to help you identify your preferences.</li><li><strong>Iterate with Insight:</strong> Use the AI's analysis to identify preferences and your own judgment to either craft a new, more refined prompt that asks for more of the best elements you've discovered, or to blend the best elements together.</li><li>Rinse and repeat this cycle as required.<br data-mce-bogus="1"></li></ol><p>You will likely produce a better result. More importantly, you will learn something about what you consider defines a quality outcome for that process; an insight you can carry forward with you.</p><p><em><strong>Tip: </strong>Don’t burn time to craft the perfect prompt to do this: LinkedIn is flooded with prompt advice, but for this quick experiment, just make the AI your prompt writing assistant. Fire up a new chat, explain your objective, and have it craft a prompt for you. That way you can get on with the creative part. If you want this to be part of your regular workflow, you can automate it.</em></p><p><strong>Go try it. And if you uncover a new insight, share it!</strong></p><p>&nbsp;<br>&nbsp;<br>&nbsp;<br>&nbsp;<br>&nbsp;<br>&nbsp;<br>&nbsp;<br>&nbsp;</p>

Your AI Mental Model: Bug or Feature?

<h3>An Amplifier, Not a Knowledge Base</h3><p>The quality of what you get from a Generative AI model is directly related to what you put into it. With Large Language Models (LLMs), it's crucial to recognize any limiting mental models you might have. For example, an LLM is not a Knowledge Base for retrieving an answer. Instead, it is a creative engine, an amplifier for your intent.</p><h3>A Look Inside the Amplifier</h3><p>Large Language Models connect words with concepts. The better your words capture the concepts you care about, the more you empower the model to respond with words related to those concepts.</p><p>Mechanically, an LLM examines every word in your prompt to understand how it affects the meaning of every other word. In 'a spicy meal', for instance, the context gives 'spicy' a completely different meaning than in 'a spicy conversation'. If it is spicy food you care about, you need to make that obvious.</p><h3>The Importance of Signal Over Noise</h3><p>In practical terms, your LLM Amplifier is unconstrained. It has a gigantic map of concepts connected by words. Find the right words and you’ll surface relevant concepts you didn’t know existed and which could transform your horizons. This is why the signal matters so much.</p><p>The 'signal' you provide gets magnified. The mechanism of how the LLM works is to your advantage when you start with the right words, pointing it in the right direction. If it heads off a poor path, or if there is too much ‘noise’, the results will be lower quality and opportunities will be missed.</p><p>This is especially so because each new word builds on the last, so a bad start can cascade, and your signal can get lost in disappointing white noise.</p><h3>Turning Up the Volume</h3><p>This is the critical part. If you use your LLM with a closed-question mindset—as if it were a knowledge base—your prompts will be limited, and the responses will be pedestrian.</p><p>However, if you approach it as a creative process, you will provide the context needed to prime the pump. You can show it what’s important and what a good outcome looks like, encouraging it to explore potential rather than just confirming preconceptions. This way, you can amplify your intent by an order of magnitude and grow your own capability in the process.</p><p>The takeaway is simple: treat your LLM as a creative partner. Move beyond simple questions and start providing rich, contextual prompts. Challenge it, guide it, and give it a clear signal to amplify. The quality of your outcomes will not just improve; it will transform.</p><h3>Put It into Practice</h3><p>Ready to try it? Here’s a simple exercise to experience the difference firsthand.</p><ol><li><strong>Pick an Existing Workflow:</strong> Choose a task you perform regularly that is tedious or challenging.</li><li><strong>Ask a Simple Question:</strong> Start with a basic, database-style query, like: 'How do I do X?' or 'What are the best ways to do X?'.</li><li><strong>Tighten Your Signal:</strong> Now, enrich that request with specific context. For example: 'My goal is to accomplish X. My current process is... Based on established best practices for X, compare my approach and identify opportunities for me to get a better result.'</li><li><strong>Build a Capability:</strong> Use the analysis to create a new workflow. Ask the LLM: 'Based on that analysis, generate a checklist for a new best-practice approach I can follow. </li></ol><p>Then, find relevant case studies or stories that demonstrate this approach in action.'<br>The goal is not just to reduce the work for a single instance, but to build a reusable capability, improving your knowledge and empowering you to focus on higher-value, more interesting work.</p>