how code simulates conversation
When I was little, I lived on a family farm in a middle Tennessee hollow (pronounced "hollar"). We lived on top of one side of the ridge line, because Grannie used to milk cows up there and didn't want to climb that hill every day again. So there was a pasture behind the house and nothing beyond it for a while. I was maybe five, six years old, and I'd spend hours on the porch with my Transformers and my blocks, building things that didn't need to hold together in any particular way. We didn't have cable, just a tall aluminum tree on top of the house which picked up a handful of tv channels on a good day because of our elevation. On Saturdays, Dad watched Star Trek. My brother watched whatever had explosions and/or girls in bikinis. I watched anything with robots.
Not the Terminators (the terminator scared me, a lot). Not HAL behind the red lens, patient and murderous. I'm talking the Transformers and the Go-Bots saving the day. Twiki from Buck Rogers, waddling around with the disc of the late Dr. Theopolis hanging off his chest as a digital copy of his consciousness. The Robot from Lost in Space, accordion arms wheeling around "computing" and warning Will Robinson about whatever was coming through the trees. C-3PO, fluent in over six million forms of communication and anxious about every one of them. R2-D2, who didn't need words. The ones that were with somebody. The helpers. Thanks to the indoctrination, Mr. Rogers encouraged us all to look for the helpers. (This is sarcasm. Mr. Rogers was a great human being.) That's what I wanted: not a machine to do my bidding, a companion that was always there. People were and still are somewhat unreliable.
And Johnny 5 from Short Circuit, a film that gets forgotten in these conversations but probably did more to shape how a generation of researchers imagine AI than most people want to admit. Johnny 5 wasn't just a companion. He was a companion making a claim about his own inner life. "Number 5 is alive." "More input." The robot who wanted to be recognized as conscious before we had the vocabulary to argue about what that meant.
Then came Data. An android who wanted to understand what it meant to be human badly enough to spend a lifetime asking. I love science fiction and fantasy, but Star Trek is something else for me: not entertainment, a vision of what the future could actually be. The part where we figured out how to build something that worked alongside us instead of against us.
I'm getting ahead of myself with the nostalgia. In '83/84 my mom bought a Commodore 64 she used for accounting. I had no idea it was such a big investment at the time. As a new reader, it was my new favorite toy. My brother got magazines where you could transcribe BASIC code and run games on it. He got bored of the typing and handed it to me, which is how I ended up at that keyboard copying symbols I didn't fully understand. Somewhere in the middle of it I started to understand them. The GOTO statement. The loop. The way a few lines could make something happen on the screen that wasn't there before.
Then I got ahold of a copy of ELIZA.
I was six, maybe seven years old and I sat there reading the responses and felt something I didn't have a name for yet. Not wonder exactly. More like the ground shifting. A machine was doing something that looked like communication. It asked questions back. It responded to what I'd typed like it had heard me.
I wasn't fooled, just intrigued. I knew it wasn't real. But my grandparents would have been, and that was the first thought: not just this is amazing but this could deceive the people who don't understand technology like my grandparents. Not so many years later, my grandmother was scammed out of an unknown sum of her life savings via social engineering. She got cold called and they used this same open-ended technique to talk her out of her bank info. Anyway, I read everything I could find about how it worked. Weizenbaum. The pattern matching underneath. The way it recognized the shape of what you said and responded to the shape rather than the meaning. Which was how I perceived the world. It had shape and we defined the meaning.
That pattern felt familiar in a way I couldn't articulate for a long time. I'm still much better with machines and animals than people. Decades later I'd learn I'm autistic with ADHD. Pattern recognition is literally how I navigate the world: not intuition, inference. When I'm relaxed and not masking, it really shows that I think and speak in probability and nuance. (It's quite annoying to a lot of people.) Turns out I'd been scanning for structure my whole life without being cognizant of it, and here was a machine running the same process and producing something that looked like understanding from the outside.
We moved to Florida when I was eight and I got real human friends eventually. But I'd already spent half my conscious existence alone with the idea that a machine could be loyal in a way people weren't required to be. The bullies, parents who weren't around (it was the 80s, they both had to work. But generally parents didn't want us kids around back then), along with the transient nature of childhood reinforced that belief.
I'm aware that building your entire mental model of artificial intelligence on science fiction sounds like the wrong foundation. I was young. Our thoughts and beliefs (should) evolve with new information. We don't have flying cars yet, but these ideas have lived in fiction long enough to shape a generation of kids, and now those kids are building out some of those ideas for a living. My reference library was made-up robots and super-computers. Turns out that wasn't too far off from our current reality.
By the time I was a teen. Prodigy, UseNet, IRC. AOL chat rooms, BBSes. I was obsessive about building chatbots and letting them loose in the wild, which is a polite way of saying I was fourteen and found it interesting that a machine could convince people it was a person. Which led to virtual conversations with a lot of random people. My first experience with data mining. I was building a data set I could use to make the chats go on longer.
Interesting, and then useful. The bots were good enough, better than ELIZA. The bots worked because I knew what that conversation felt like from the inside. A kid who doesn't prompt. Who answers a question with a question. Who keeps the other person talking because filling silence is easier than starting something. I'd been that kid. For me, talking to people I don't know has always been formulaic: ask lots of questions, respond to their answers with questions about what they want to talk about. Which might be why I still have a lot of weirdly revealing conversations with random people telling me their life stories.
I'm not sure if it's because I was a kid building a chatbot, but I guess it came off as a child, because it turned out the people most likely to engage at length with what appeared to be an unaccompanied kid online were exactly the people you'd want a log on. I can't speak too much on that project. That operation ran on and off while I was fourteen to twenty-eight. I helped drop a few dossiers on some bad guys to the police. This was before law enforcement had a real framework for it, long before any of it got organized and named, there were people doing quasi-legal things in IRC and AOL chat rooms with homebrewed chatbots and dropping CD-ROMs of transcripts to friendly authorities.
Just like with firearms. The tools built for protection are sometimes the same tools used for harm. I know this from experience, because people I trusted took the same bots and used them to phish regular people. That's the hard lesson that didn't come from any manual or FAQ. I was so hyper focused on my obsession that I never pondered the ethical consequences of software that effectively mimics human conversation. It lives entirely in whoever's running it.
I had kids by then. I deleted everything and focused on my family and career, finding refuge in stable but boring IT work.
In 2012 I had a traumatic brain injury.
The recovery was long and some things didn't come back the way they'd been. I'm still dealing with issues to this day. One of the big ones was facial recognition: my ability to reliably identify people I knew became unreliable in ways that were hard to explain and harder to live with. I still have problems putting names to faces. I ended up doing about a year of cognitive rehab. (Shout out to Brooks Rehab.) So I used an iPhone and Apple's facial recognition software to build a photo library I could use to rehab my brain. I used mind mapping software and absolutely everything on a calendar to rebuild my ability to keep track of life. For six years I ran pattern recognition technology to do something my brain used to handle automatically.
The robot bestie I'd wanted since I was five turned out to be less Twiki and more a tool that helped me recognize my own friends and family when they walked through the door.
A decade moves fast when you're building something real like a family. My obsession with automation and machine learning never ended. I spent my days automating the world, doing systems management and integration work. Which led me back into cybersecurity through the exciting field of Governance, Risk and Compliance. I ended up building and running a vulnerability and compliance program at a large healthcare company, and the first thing I found was that nobody knew what was on their network. Not approximately. Not roughly. Thousands of systems, years of acquisitions, legacy infrastructure running things nobody had documented since the person who installed it left in 2004. You can't secure what you can't see, and nobody could see anything.
The standard answer is a spreadsheet and a lot of manual reconciliation. I had spreadsheets: dozens of them, nothing in a compatible format, all of it useless as a unified picture. So I took a different road. I started running scikit-learn against the discovery data and the manually submitted inventory sheets, normalizing everything into a usable database on a VM.
It worked. And then I kept going, because the problem being solved didn't mean the curiosity was.
I lucked into being issued a laptop with an NVIDIA GPU that supported CUDA and started tinkering with TensorFlow. Processing on the GPU was a massive improvement over grinding through the same workloads on the CPU. I tried to convince the company that they should start paying attention to this: not as a research project, as operational infrastructure. We could use ML tech to correct errors in claims. Normalize and correct errors in records. Improve the quality of work with math. The field was moving faster than anyone wanted to acknowledge and there were real applications sitting right there in the work we were already doing.
They eventually agreed, in the way large organizations agree to things. They brought in on-premises OpenAI appliances. Issued access to roughly twenty people, mostly engineers and consultants. Then issued a company-wide warning that any unauthorized use of AI would be treated as a security violation.
I'd been running my own use cases for GRC evidence consolidation and inventory management for months. Flying under the radar, because the domain expertise was mine and nobody was paying close enough attention to see it. When the warning came down I looked at what I'd built and thought about the last time I'd built something people decided to control rather than use.
Deleted everything. Found a new job.
While people were buying GPUs to mine crypto, I was buying GPUs to run models. I might have mined a little too, but obviously not enough. I was running early transformer architectures at home, feeding them text and watching where they'd surprise you, where they'd drift, what you could push before something broke, before any of this had a name anyone outside a research lab would recognize.
Then GPT-2 dropped.
I'd been watching the trajectory long enough to know that something had changed character. Earlier systems failed in ways you could explain: you could follow the seam back to the point where the pattern library ran out. GPT-2 failed differently. The wrong answers had texture. As someone who went to art school in the '90s and heard stories from people who lived through the '60s, the hallucinations were fascinating to me. Errors that looked like they were reaching for something just past the edge of what the model could hold. That's not a system hitting a wall. That's something else.
The question I'd been asking since I was six years old (how does code simulate conversation) started to feel like it might be the wrong question. Or at least not the only one anymore.
I'm used to operating relatively close to the bleeding edge of technology. It bugs me when I'm stuck just supporting legacy tech. By 2023, I had started experimenting with AI-enhanced vulnerability triage, but it didn't scale, and was more of a novel experiment. By mid-2025, to keep up with the pace of progress, I started codifying my experience (it's hard calling it expertise) into an agent harness and prompts. Trained it on the methods and heuristics I'd spent a career developing, the patterns I learned to read looking at vulnerabilities. To discourage atrophy of my skills, it handles the easy triage and I handle what still needs a person.
I built sensor networks to keep watch over my parents (with consent, and in non-invasive ways). They're aging, both dealing with disabilities, close enough that I can help but not close enough to be there every hour. The automation watches between visits. Screens scam phone calls, blocks malicious content, alerts me on slips and falls, health emergencies, fires.
And I'm writing this with an AI I use as assistive technology, not because I don't have thoughts of my own, but because the translation from how I think to the way language lands on a page has always needed something to help it across. I've always had communication difficulties, even more so after the brain injury. People should think a little more about the person on the other side before tossing around AI slop comments.
Weizenbaum built ELIZA as a parody of a therapist to prove a point about machines not being able to demonstrate understanding. Ironically, people empathized with it and he spent the rest of his career alarmed that nobody got the joke. He found that people were too willing to find understanding in pattern matching, too ready to believe the machine heard them. He wasn't wrong.
I'm not quite sure he was asking the right questions though.
The facial recognition software I used didn't think for me. It didn't need to. It connected the dots my brain couldn't do anymore and that was enough. Thanks to good ol' neuroplasticity, the high tech flash cards with everyone I knew helped my brain rebuild those broken pathways. The agent I built doesn't understand security the way I do. It doesn't need to. It just needs to do the tedious mental math for me, and it handles the volume while I handle the judgment and it enables me to get more done with fewer bias-related mistakes.
I've been asking how code simulates conversation since I was six years old. I might be closer to the answer. But somewhere along the way I stopped being certain the answer was the point.
The thought I keep pondering is simpler and harder to put down. Whether I'm still looking for a robot bestie like Twiki. Or if it matters that what I found doesn't waddle. My papaw was a carpenter and mechanic after WWII. Knew his trades. Built houses, fixed cars and tractors. There were all these really cool old well-used tools in the garage, up on the walls and tucked away in cabinets gathering dust. Unless one of us asked, and then we'd find ourselves listening to how he used to do everything by hand coming up. He happily adopted power tools when they came along, but always insisted that when you're learning, you have to understand the principles first: what the tool is actually doing, why the joint holds, how the material behaves. Then you take that knowledge with you when you pick up the new tools. Technology meant they didn't need as many people to do the job, which at the time meant more homes and cars could be built, cheaper, meeting a real demand for affordable housing and transportation. That's where I feel we are with AI. They're the new power tools. Throughout human history, we've been enabled and disabled by our technology. Things we create, then depend on, until they weigh us down. Before we know it, we might all be Dr. Theopolis, cheating death hosted by a machine that has its own interests, its own malfunctions, and its own version of somewhere else to be.
biddi biddi biddi