Talking to Machines, Part 1: The Rapid Evolution of Human–Machine Collaboration
A peek into how changes in AI “thinking” are reshaping how we work with it
Talking to Machines is a three-part series exploring the rich changes in interacting with AI and what that means for improving how we work with it.
In this series, I’m looking into practical ways of bringing generative AI into our workflows from my perspective as a human-centered designer, and my experience leading product design for teams, products, portfolios, and global information systems.
Part 1 offers a conceptual anchor for where we are in AI’s evolution, a bird’s eye view of how LLMs work, and a foundation for us to explore prompt and context engineering in the next two parts.
Subscribe to get future essays and support this work.
Or give this post a like :)
A growing toolbox for working with AI
We’re at an inflection point in how we interact with AI systems. Large language model (LLM) behavior is rapidly evolving. The technology is becoming increasingly powerful, and at the same time, more collaborative.
The ways we as AI operators can influence their outputs are advancing in sophistication and specificity. More control is emerging for engineers, and some of it is now reaching us through new features.
We can move beyond just typing into a prompt. Influencing how an LLM interprets and responds is looking more and more like real-world team collaboration — from role-playing to get new perspectives, to providing “how-to” guides for specific projects, to curating knowledge we can reference over time.
It’s a classic personal computing shift. The MS-DOS command line to graphical user interfaces (GUIs), only this time within AI. As that shift plays out, I’m catching myself using it more as a thought partner.
For many use-cases, such as organizing complexity or exploring potential paths, the process is simply faster. Yes, there are limitations, many requiring subject matter expertise and verification to avoid being wrong. We know about AI hallucinations.
That’s where knowing how they work is helpful. Or more precisely for my interests, how they “think". That under-the-hood insight is both fascinating to explore and instantly applicable. It’s analogous to learning about psychology to help collaborate more meaningfully with other people. Just here it’s a machine.
This three-part series will break down LLM-based generative AI in an accessible way, focusing on how we might better collaborate with it as an emergent toolset. I will dive into some basics of LLMs, what prompt and context engineering are, how they are distinct and related, and offer simple mental models and analogies to help make it all more approachable. I love digging into this, and hope it proves useful and interesting to you.
Context engineering is the new context
Lately, “context engineering” is being talked about as the “new prompt engineering”. And in my opinion, this framing as a replacement is more a catchy misnomer. The “context” really sits a layer up. It includes the prompt, among many other things.
My first uninformed thought? Westworld had it right. If you are making AI more believable and effective, you need deliberate context for it beyond a general “trained on all the things” LLM. Memories, specific information, and preferences.
What largely makes people unique are information and perspectives built from our own personal contexts. Our responses depend on what we’ve been exposed to, what we recall, and how we make sense of and prefer to present information. In a nutshell, our experiences, knowledge, and personality — all from a life lived up to now.
Context engineering seems like something along those lines. Specificity in inputs creates specificity in outputs. It’s sensible that machines designed by people for people would operate similarly. Let’s explore that, starting from the basics of how LLMs work.
And a friendly FYI before you dive in, there are a few practical examples sprinkled in this series, but the focus here is on unpacking the concepts, not offering a list of quick tips. Those are always easy to find as you need them.
A bird’s-eye view of how LLMs respond
A bit of conceptual understanding is incredibly helpful in working better with AI tools. Beyond being interesting, it’s pragmatic knowledge. This is the direction our daily tools are headed. The “psychology” and “neurology” of machines sheds light on how LLMs respond, naturally improving how you work with them.
They do not all “think” the same way, especially between providers. They have been trained on different data, with variations in approaches and core guidance. Even the same prompt sent to the same LLM will likely give varying responses, by design. It helps the AI seem more “creative” and “human”, creating the illusion of collaboration with a real entity.
The knowledge of how to get the responses you want will strengthen with practice and become second nature over time. You may also begin to favor different models within and between providers for different types of tasks. It’s much like collaborating with people, where you develop a rolodex of go-to collaborators for different specialized topics. Or you may just be fine with a generalist.
Now let’s get into some conceptual basics.
Responding on the fly
The first bit to know — LLMs are not retrieving stored answers. Responses are generated in real time using probabilities of what word should come next, based on their training data. Each word is generated based on what came before, guided by patterns it has learned from training. You see them working in real time before your eyes.
As the user you have influence over this process with your prompt, your previous messages, and any other context given to the LLM while it “thinks”. Expanding on this leads directly into context engineering, which we’ll explore later. For now, know that this context-setting impacts the probabilities of what comes next by influencing what is “front of mind” for the LLM.
Especially in niche areas, they may generate text or other outputs that seem to regurgitate existing content, occasionally to the point of plagiarism. Not excusing this, but in domains with limited references, the result is explainable. Still, it creates a huge gray area. The lively philosophical, ethical, and legal debate, which you have likely heard about, centers here.
Where should we draw the line on what data can be used to enable LLMs? Depends on who you ask. And for me, the striking parallel to ponder is that we respond in much the same way, based on what we’ve read over time. After all, none of us is “one mind”. As innately social beings, we are passively shaped by everything we intake from others. Are we all plagiarists by nature? That’s a rabbit hole for another time.
Limited by a “Context Window”
Today’s LLMs are limited by their “context window”, or quantity of text that a model can consider while generating responses. Think of it like our “working memory” capacity — the amount of information we can store in our heads as we problem solve and perform tasks. Like people, there is an upper limit. And like different people, the window sizes are different between models. But the theme remains, it’s not infinite.
LLMs, as they pull from what they know from training, only work with what is still in that context window (e.g., your prompt, prior messages still in range, etc.). A window that moves along with the conversation. As you may have deduced, the longer the chat, the more likely the thread of conversation gets lost by the AI. The earlier messages and criteria you set will be forgotten without extra mechanisms in place. They are naturally ephemeral, existing outside the current context window.
For us as AI operators, this memory limitation in our tools can be incredibly frustrating. Longer conversations naturally become cyclic. Responses begin contradicting previous constraints while vital details seem to disappear. Somewhat like working with a curious colleague (of course, never me!) that can drift off-course out of excitement, losing the anchor of what was established early on.
Ways of approaching this issue seem to mimic our own psychology, getting into forms of structured memory. Some tools are beginning to store user and project-level preferences from conversations — you may have seen “Memory updated” appear in ChatGPT. We also can increasingly pre-load structured information and documents, analogous to strategy-setting and standard operating procedures (SOPs) for teams and organizations, helping to stay more on course throughout a body of work.
Evolution like this will continue. Clever engineers will keep tinkering in the growing field of “context engineering”, and with that, more AI operator skills will keep emerging.
No true knowledge, awareness, or beliefs
Even though it may seem otherwise, as with well-documented cases of AI going off the rails into extremism and misinformation, LLMs do not have beliefs about the world, awareness of what they are saying, or knowledge built from experiences.
They don’t “know” things, have “awareness”, or check their accuracy. They are trained, then generate words one at a time based on the quality of their training. The old “garbage in, garbage out” trope.
Probabilistically generating each next word will result in outputs that appear as believable knowledge objects. But that’s much different than knowing information through exposure, understanding what's what, then accurately applying that knowledge contextually. Their apparent confidence is not a measure of accuracy.
LLMs can drift off course while responding, with a minor word variance compounding into an egregiously inaccurate output. Or one that simply reflects its training data, which may or may not have included some random corner of the internet with interlinked blogs or Reddit posts with dubious peer review.
Fascinating techniques are emerging to mitigate these issues, but the field is fairly nascent. Smart context engineering (which we’ll jump into in a later post) is extending in-the-moment responses beyond LLM training data, such as with retrieval augmented generation (RAG). But the base principle still stands. They are simply predicting the next most likely word, based on a fed corpus of information and guidance.
What looks like knowledge, awareness, or beliefs is, in reality, nothing more than the output of mathematics. An output shaped by biases, intentional or inadvertent, from those who trained the model.
This Talking to Machines post is the first of a three-part series. If you found this insightful or interesting, hitting the like button helps me know it resonated. And if you want to follow along, or help this thinking reach more people, subscribing is the best way to do that.
Here are following two:
Talking to Machines, Part 2: Crafting clear conversation with Prompt Engineering
Talking to Machines, Part 3: Designing the AI’s world with context engineering
Keep thinking and making. And be well.
Have a wonderful day.
- Peter
Read Part 2 next:
Talking to Machines, Part 2: Crafting Clear Conversation with Prompt Engineering
Talking to Machines is a three-part series exploring the rich changes in interacting with AI and what that means for improving how we work with it.
Westworld DID have it right 💯
Thanks for the good 😊