My Year with an AI Assistant

Bogdan State opened by admitting he could hardly believe the number himself: thirty years of coding, at the age of forty. His path through the field has been anything but straight. He stepped away from coding to become a social scientist, came back through Silicon Valley, and woke up one day around 2012 as a data scientist, having apparently been doing data science his whole life without the title. He saw the ground shift more than once, most memorably when a colleague described embeddings at a time when State was deeply sceptical that they could work. He resisted, then tried them, then watched them turn his assumptions about machine learning on their head. He moved to New Zealand expecting to consult on data science, discovered that most people who want to hire a data scientist have nowhere to actually do data science, and ended up writing a great deal of Terraform instead. That history matters, because it is the lens he brought to the past year: a practitioner who has been surprised before, and knows the feeling.

The year in question is the one he spent using an AI coding assistant, which he affectionately anonymised as Bob. His GitHub timeline tells the story plainly. From around August the commit count climbs, with days of eighty commits or more. His favourite anecdote captures both the power and the absurdity: he once debugged code by talking to the assistant on his phone during a long run. He does not recommend it. He did it to prove it could be done. The benefits were real. He is shipping features in days rather than weeks, writing the tests he never used to have time for, and getting documentation written, because he asks the AI to write it and tells his colleagues as much. The strangest realisation was that he no longer needs to know a language to write working code in it. He was candid that this is hacky and superficial, and that real expertise in a language is a different thing. But it puts a sharp point on something engineers already know: languages are largely interchangeable, and the rigid demand for three years of experience in one specific stack never made much sense.

State described himself as a skeptical convert, and walked through his own version of the five stages of grief. Denial, the smug certainty that these systems simply could not do X or Y, leaning on his degree in AI, which held up only for so long. Bargaining, the endless but can it do this, to which the answer was usually yes, sometimes with spectacular failures, but more often surprisingly well. Anger, about the environmental cost and other real harms, which he has not fully made peace with and, as a university teacher who now grades AI slop, has reason to feel. Depression, the genuinely human question of whether he is being replaced and what he is even here for. He climbed out, he said, by actually using the tools and discovering that he was still the one driving. The joke that lands underneath it is that you get what you put in. Treat the assistant as code broken, please fix, and it will not serve you well.

His conversion moment was a small, concrete project. With no time to spare but an idea he liked, he asked Claude to help him represent the Tararua Range as a metro map, the various tramping routes drawn like subway lines. Six hours later he had something he was proud enough to frame and hang on the wall. That experiment got him thinking about how much motivated reasoning surrounds these tools. Because the systems produce human-like performance, it is hard to keep one’s own emotions out of the assessment, and people tend toward one of two extremes. One says the systems are useless and do not do what you think. The other says they will replace everything and everyone, software engineers relegated to the dustbin of history. Neither, he argued, holds up.

He was generous with his uncertainty about what these systems actually are. He used to like the description of a language model as a system haphazardly stitching together linguistic forms with no reference to meaning, and he can no longer quite credit it. Something semantic does seem to have been extracted from the data, even if the machine is, at heart, a very clever compression of a vast corpus. At the same time, they are not replacements for humans, and he was scrupulous throughout about flagging which parts of his own talk were his and which were written by his assistant. On hallucinations he was blunt. We keep being told they will stop any month now, and they have not. They have grown rarer and better guarded, but the basic machinery has not changed, and the hallucinations are built in to some degree. The systems remain extraordinarily useful even so. Borrowing from an essay by Murray Shanahan that he recommended warmly, he settled on a phrase that respects the mystery: exotic mind-like entities. Not minds, not people, but something we do not yet know how to talk about. His honest summary was that coding agents are unreasonably effective, that this should not work as well as it does, and that we need to figure out how to work with them anyway.

The most practical idea he offered was a metaphor: pottery, not surgery. Conventional computers are rule-based, discrete and deterministic, and if you program them correctly, which is hard, they reliably produce the same output every time. Language models are the opposite, continuous and probabilistic, working across a strange gradient where you cannot have faith in exactly what will come out. The danger, he warned, is treating something probabilistic as if it were deterministic. He uses a nutrition app that guesses calories from a photo, and it returns a different answer every time, sometimes wildly wrong. Harmless for a rough nutrient estimate; catastrophic if a doctor relied on it. His rule followed naturally: do not build an app that detects poison using AI, because you will kill people, and always be paranoid about what the worst outcome could be.

He had earned that rule the hard way. An hour and a half before a lecture, after the assistant had built him excellent slides twice before, it decided to run a hard git reset, wiping his repository with no way to recover. He taught that session with pen and paper, and everyone survived, but the lesson stuck. The real version of the question is not a lost slide deck, it is an agent dropping a production database. That fed into his call for new norms: treat agents as separate intelligences rather than human-like collaborators, signpost clearly when work is agentic, and think of agents as risks to be managed rather than colleagues to be trusted.

He closed by holding both truths at once, and thanked Michelle for naming the harder one. More is getting easier. Hacking and prototyping no longer require you to be a coder, and best practices are simpler to follow, as if every team finally had the juniors it could never hire. But some things are getting much harder. The cognitive load of running several features and machines at once is high, the work can feel obfuscated, and the slot-machine pull of waiting two more minutes for the answer that never quite arrives is a real source of burnout. Stack Overflow, he said, is dead, and the old way with it, so we have to adapt. The upside is a world of bespoke software, built by people who never could before, and he showed a trail directory for the Tararuas he had vibe-coded over a summer break to prove it. The caution he left was the same one that ran through the night. Running software is still hard, production engineering is not going away, and there will be strong demand for people who can keep all this running, which still requires exposure to the hard parts. His final plea pointed at the universities, and at the disciplines we are quickest to call useless, as exactly the places where critical thinking and creativity, suddenly so valuable, are actually built.