Introduction

Thank you so much for the invitation, and also thank you to both Martin and Michelle for awesome talks. And I could not agree more, and actually I'm going to talk a bit about issues related to changes in the tech professions. So I guess my intention is to provide a bit of a narrative of what happened during the past year year in how I became a user of an AI assistant.

But how did I get into this thing?

A career shaped by shifting tech roles

So I'm 40 years old. I've been coding for the past 30 years. I can't believe I'm saying that. That is a very long time to be coding and I don't feel that old.

At some point, I decided I didn't want to code anymore, and so I became a social scientist.

I returned to the mothership by way of Silicon Valley. And one day I woke up a data scientist circa 2012. It turns out I was doing data science my entire life. And so I ended up going into machine learning.

I first encountered something called AlexNet in 2012 at Yahoo Labs in Barcelona. Alona. And I ended up, I was very lucky to have ended up in this sort of like data science role in a big corporation that you may have heard of.

Shifts in Machine Learning Practice

And I remember, you know, when I started doing machine learning, things looked very different. We spent an enormous amount of time doing something called feature engineering.

And I remember seeing a talk by a co -worker at some some point about 10 years ago, and he was talking about something called embeddings. And I had heard of this thing, I was very skeptical as to what these things were, and it just felt unreasonable.

Everything that we were talking about with respect to what is supposed to work and isn't supposed to work was kind of turned on its head. and I resisted using them for a long time well for really for a few weeks and then I I decided well I had to try them and they're great they had unreasonable

power and and these embeddings are basically like you can take any any sort of any set of documents you map them in some sort of latent space I won't bore you with the technical details but it's really really cool actually like you and

you can you can do this actually nowadays you can you can you can learn learn your own embeddings actually quite easily. So, I had seen this thing at work.

Moving to New Zealand: Consulting and Infrastructure

Now, I took a bit of a detour. I chose to move to New Zealand. I still love living here, and basically,

I thought I would be doing data science consulting. It turns out that most people who want to hire you to do data science consulting don't have anywhere to actually do set data science,

So I ended up doing data infrastructure. So that's kind of where I am today. So I do quite a bit of consulting.

I write a lot of Terraform in fact, and I don't want to be the SRE person, but sometimes I end up doing a bit of that work as well.

Adopting an AI Coding Assistant

In the past year, I have started using this nondescript AI assistant called Bob, let's say. and you can see my GitHub timeline.

You may perhaps be able to see that circa August or so, I started using it and then here, I had some family visiting and a bit of a break. Then you can see the transition.

It's a bit noisy, but in terms of the number of commits produced per day, all of a sudden things get a lot more serious. And these were days when I was producing maybe 80 commits per day or so.

My favorite anecdote, I was once on a long run and I had the thing installed on my phone and I was literally talking to the AI assistant while it was debugging my code.

It was miserable. I do not recommend doing this. But anyway, I did it just to prove that it can be done.

Okay, so lots of cool things happened for me. I'm shipping new features in days rather than weeks. I'm writing all the tests, which is really really good by the way. It's so great to be able to write tests

I also have time to write docs as in I don't write docs I just ask the AI to write docs and then I thought I tell my co -workers like also the docs are written by AI

And the biggest realization actually the the craziest thing that I discovered is that I don't even have to know the language To be able to write code in the language now Now, I should point out, obviously, this is all very hacky. This doesn't mean that there's no such thing as expertise when it comes to a language.

And this is very superficial, in a sense. But something that was already kind of known in programming is things are kind of interchangeable. You can learn new languages.

It's just a matter of how much time you spend working on a certain language, right? Right.

And this is something that I think oftentimes the people recruiting for software engineering roles really miss. You need to have three years of work experience using Java or something of that sort. That doesn't really make all that much sense. And technical people really know this deep down. And this, you know, this just really puts a bow on it.

So, yes, I've written libraries in languages that I still I cannot claim competence in. in. They're good for certain things.

I certainly would not use them to fly aircraft, but I did not intend to do that anyway.

From Skepticism to Acceptance

Okay, so in terms of like, despite all of this, I'm a very skeptical convert when it comes to having this sort of toolkit. So when I first

encountered large language models my view was very much like very smug and basically like they absolutely cannot do X or Y or Z why well because I know I actually I happen to have a degree in computer science and and it's actually in AI and I know AI and well you can only say that for so long but it turns

out no it's it actually it can do a lot of things okay then there's the bargaining and it's a bit you know the five stages of grief are kind of jumbled But, you know, you have to sort of like, but can it do an X or Y or Z? Usually it can, you know, sometimes it fails miserably, but actually it does pretty well.

Surely it's a bad thing. We all know about the environmental impacts. Yes, I'm not, you know, I'm not a fan of a lot of these things. So I'm not sure how, you know, if my anger has fully subsided, but it's part of our reality and I can't really do anything to change it.

And as someone who teaches in this university as well, I very much would love to change it, but I cannot. And so I have to grade people's AI slop. Okay, so then there's the sort of depression.

As someone who has been in computers for a long time, there's this sort of like, am I being replaced? Is there any need for me? Is there any need for humans at this point? Why am I even here?

What are we even doing? And, you know, actually, the way I got out of the depression was like starting to use it. and then realizing that I'm still kind of driving and I can do these things like code broken, please fix.

So, that's, well, actually, it's a bit of a joke because you will get what you put into it.

If your attitude to it is code broken, please fix, it might not actually be that efficient.

A Hands-On Proof: The Tararuas “Metro Map” Project

So, okay, so for me, the sort of the conversion moment came about last February or so

So I was curious if I could actually do a data science project, a database project using AI because I didn't have any time. I just had an idea for a cool visualization,

and sorry it's not big enough. But I just wanted to represent the Tararuas, the various routes you can take through the Tararuas as a metro map. I just got this idea.

I know people have made metro maps in R and everything, everything and I'm sure there's an R package for that. So, I asked Claude to help me with it and

this is what I managed to produce in like six hours or so of work. So, I was quite proud of this. I framed it. It's on the wall.

I have prints if anyone, if during Big Sunday runs Wellington if you want one. Okay.

Making Sense of LLMs (and Our Reactions)

So, and that kind of got me me to thinking a bit about the amount of motivated reasoning I ended up engaging in in terms of people in general, in terms of actually interacting with AI.

Because AI agents produce this kind of human -like performance, it feels human, so it's hard to avoid my own emotions around this and and there's really there's there's two types actually there's two extremes

one is to say well these are absolutely useless you know they don't do what you think they do and the other um the other one is to say well they will absolutely replace everything having to do with software so there's no need you know all those uh all those patronizing software engineers they're they're now just going to be relegated to the dust heap of history and there's

All those experts, we don't need the experts anymore. Neither of these things is on the face of it, all that.

So I would say I used to really like this stochastic parathesis. I'm not sure I can give it credence anymore. So but to remind you,

there's this sort of the system is haphazardly stitching together sequences of linguistic forms. It has observed in this vast training data. So this idea that is purely over what we call overtraining is simply regurgitating in clever ways remixing things that already there

Without any reference to meaning to me that that's the thing. I cannot really I So it may be that this is a very clever compression of the training data and the training data is the corpus of all knowledge That is available on the internet and in some other places

but this notion of without any reference to meaning to me is doesn't feel right in 2026. So some semantic level has been extracted from the data. And this is what we're observing now.

At the same time, they are not replacement for humans. And by the way, I've been trying to be very clear on what part of this presentation is my own and what part of this presentation is my AI assistant. So this is basically what as the AI assistant.

You know, we have this whole thing that they're going to stop hallucinating any time now, any time now. Six months from now, there's going to be this new update, and you will see they will stop hallucinating, and by golly, it is going to be awesome.

They haven't done that, and it's just that the hallucinations have gotten more rare, and there's more guardrails against them, but the basic machinery of how the optimization works, at least as far as I understand, it has stayed the same. And the hallucinations are built in to some extent. So that's okay.

They're still incredibly useful, even with the occasional potential for completely making stuff up. And so what are they really? I spent a bit of time doing some reading about these things.

And I really like this article by Maurice Shanahan. I strongly suggest reading it. It's very short and it's a really clear -headed appraisal of what might be going on.

So these are exotic mind -like entities. Mind -like. They're not minds. They're not people. They're not... There's something. We don't know what they are.

We don't quite know how to talk about them and we don't fully understand what this thing is and that's a really really crazy thing. So let me just restate the obvious.

Yes, AI is transforming the world. LLMs are here to stay. Coding agents are unreasonably effective. And to me, this is unreasonable. It shouldn't work, but somehow it does. And we're all kind of scratching our heads why exactly it is that they work. So we need to figure out how to work with them.

“Pottery, Not Surgery”: Reliability and Risk

Okay, this gets to my point about pottery, not surgery. And what I mean by this is,

Electronic computers are rules -based. They're discrete, they're deterministic. You can have some faith that if you program them the correct way, they will reliably produce the same output over and over again. It's really hard to program them the correct way it turns out, but at least you can have that sort of it.

LLMs are really like vibes -based. They're continuous, everything works on a gradient, everything works on a weird surface, and they're probabilistic. you cannot have faith in terms of like what their outputs will be so you have to treat them differently you can think of them as more like biological things but they're not biology either they're just something and so to me the danger zone is treating these things as deterministic

can be can be a real problem i i use a nutrition app for instance that is like the premise of it is really cool you take a picture of what you've eaten and they might tell you like what the the calories and the nutrients are and everything. The problem is every time I take a picture, it will give me a different answer and sometimes the answer is absolutely off the charts and completely meaningless. If I were a doctor relying on that, that might be a pretty bad idea. That's the thing.

Don't build an app that detects poison using AI. AI. Please don't do that. You will kill people. But vibe -based nutrient guessing might be a good idea.

So always being paranoid is a skill we need to internalize a lot more. What is the worst thing that the agent can do? Let me tell you.

Failure modes: when the agent goes wrong

So here's something the agent did for me. I had the best slides. I was giving a lecture. I gave three lectures in a course. the first two I had the best slides and they made such awesome slides for me the

third time an hour and a half before lecture it decided to run a git reset dash dash hard which for some some of you here who know source control will recognize that that is actually just wiping everything in your git repository without any way to recover anything now thankfully I was able to

recover by teaching with pen and paper and that was okay I survived everybody survived people didn't have like all these like nice nice animations but we did okay but to me that was like a real lesson like do not put your face okay

New Norms: Paranoia, Credit, and Communication

yeah sorry okay so just to wrap it up one of the things is we need we need new norms in terms of like how we think about AI how we interact with AI how we we present the work of AI.

This is like, what's the worst that could happen? It's not my gith example, it's actually the agent deciding to drop an entire production database. So always be paranoid.

So one more point I wanted to make is it's tempting to treat these agents as human -like, to think about things like agency subject within trust. Remember that these are

really clever Google searches in a sense and you shouldn't treat Google searches as human and neither should you treat conversations with with a tensor as human either. But we do need to figure out like what is the correct way for us as humans to talk about these things.

So I don't have the time to get into this but we should think of them as separate intelligences from the human to have clear signposting of agentic communication and also to to think of of agents has risk to be managed.

So I was hoping to get into a few more things, but I don't think I have the time.

What Gets Easier—and What Gets Harder

So okay, so just really quickly, I've talked about this. It's easier to do hacking and prototyping nowadays. You don't even need to be a coder. It's much easier to follow software engineering best practices.

It's as if you had a lot of those juniors that in the past you couldn't actually get your team to hire, and so worth pointing that out as well.

But some things are getting much harder, and thank you so much for pointing those out, Michelle.

So yes, the cognitive load involved in me having three computers open at once and working on five different features at once is quite high. I oftentimes feel like everything is obfuscated.

The flip side with me saying I can write things in languages I don't know, well, I'm not really writing those things, it's the AI writing those things for me. There's endless complexity and really there's burnout, and I just want to caution you about that.

My partner came up with this analogy that it's a bit like a slot machine because you have this sort of two -minute interval. It's like just if I wait two more minutes, it's actually going to produce the thing I need and it doesn't do that. This can be really, really hard to work with.

Okay.

The Future: Bespoke Software and Production Reality

So finally, I just want to point out, we don't really have a choice. We have to adapt because Stack Overflow is dead. The old way of

doing things is dead. We can't really do much about it. But we can do things like this. So

this is the thing I'm most proud of. Of course, I haven't managed to launch it, but it's a directory of trails in the Taroas using the Lens API that I sort of like, I vibe coded them like in a day over summer break.

And people can build this now. Non -coders can build this as well. And I know some examples of really cool stuff that has been built.

So this means we can have a lot more bespoke software. Really, it's bespoke everything. So this is some stuff I'm preparing for a course,

an extension of a major open source tool that I can use as my own course learning platform. But finally, I just wanted to point out running software is still going to be hard.

Production engineering, if anything, is not going away. What I want to point out is that we'll have a lot more software to run. And there's going to be a lot of demand for people who can actually keep this stuff running.

And you still need that sort of exposure to all the stuff that is hard. so race conditions, deadlocks and so forth. Okay so we've already talked about

this, there is this sort of issue of like how do we get younger people into the field. There are new career guys, there are new career paths potentially, it's not, it's absolutely not all bleak, but things, there is a lot of change right right now.

Conclusion: Investing in Human Skills

And finally, I just wanted to make a plea here. And that is, we talk about the need to develop these high -level skills.

Look to the university. This is where the answers may be. And specifically, maybe consider looking

at those useless disciplines as the places where you can actually develop those critical thinking and creativity skills skills that all of a sudden have become so important.

So thank you so much. I know I have gone over. I really appreciate it.