llm

So its been a few years since I put my opinion piece on generative AI.

I continue to do work that is free from the outputs of large language models and still hold the view that these statistical models, whilst most definitely artificial are not intelligent. It looks intelligent because the viewer has never seen this content before, and is interpreting it from the view point of wanting it to be intelligent. The output would have existed somewhere. And in there, lies the danger from a legal standpoint: if that somewhere was a place the author very much wanted to stay private, there’s potential for copyright infringement and misappropriation of others’ intellectual property. Amnesty International calls it unlawful by design.

Much of this material is scraped with no regard to resource limits at the source, sometimes causing denial-of-service for legitimate users. In one case, Anthropic’s ClaudeBot found a mirror of the Linux kernel tree on infrastructure I run, and decided to scrape that. It hit the VM so hard, it triggered a high-temperature alarm on the server concerned! Now there are two points I’d make here:

1. There copies of this same source code on “faster” servers than mine, like kernel.org itself or any of its mirrors.
2. They could just use git clone to grab their own copy of the code, and scrape that instead of asking my server to spend CPU cycles fetching the source, the HTML pretty-printing the source, only for them to extract the source code they came to scrape.

The latter point shows a total disregard for doing things efficiently. The former point shows just how little intelligence these systems actually have… a few pages ought to have been enough to say “I’ve seen this before, nothing new here, move on”… but no, they kept hammering.

Much of this raw material then has to be categorised and sorted before it can be fed into the training algorithms. More often than not, slave labour does this. This isn’t talked about much, but cheap labour from second/third world countries is the principle source for this labour.

The larger-scale systems continue to have an appalling environmental record, both in terms of their running costs, but also in their construction. Companies want to build “AI” data centres here in Australia… gigawatt data centres. South Australia’s peak demand hit 3.274GW in 2025, and there are proposals to build data centres that draw a third of this peak, all the time. It’s good that the Australian Government demand that such facilities build “new renewable generation“, however that consumed energy must go somewhere, and it’ll be into the atmosphere around the facility. The environmental impacts will likely be enormous.

This is just running costs, there is also hardware replacement costs, because this stuff works hard, wears out quickly, and is obsoleted even faster. Cory Doctorow points out that this equipment is allegedly replaced every 2 years, with the new replacements requiring a full datacentre refit as the new parts are not backward compatible with old installations. Semiconductor production requires a lot of water, a lot of very pure water, and high energy consumption. This is hardly environmentally friendly!

As such, I have big ethical objections to using this stuff. I also have big legal objections to working with the output of these systems. Using a library that is coded using these tools, sometimes you can be lucky and things work… but if you ever have to debug that code or write a pull request, you’ve then got to dig into the slop. Get your hands dirty with it… figure out exactly what it’s doing, perhaps what you or it is doing wrong. You become tainted by that slop.

I raised these issues with my workplace last year… their response was two fold:

1. they put out an “AI policy” dictating how it should be used
2. they subscribed to Anthropic’s Claude service

Now my work colleagues, instead of writing code, are cosplaying as IT managers micromanaging virtual junior developers. There is slop everywhere. Not only did management pretend they could “policy” their way out of a situation that even the High Court of Australia has not yet decided on… but they’ve actively encouraged everyone to join everyone on the slop train, because otherwise they’ll be “left behind”.

Funnily enough, our very own code-of-conduct has words about this:

Protect the environment by making environmentally friendly choices

It would seem they do not honour their own code-of-conduct. A facility that needs energy equivalent to a third of an entire state’s peak demand doesn’t sound like an “environmentally friendly” choice to me.

And while I do value co-workers contributions, the ones they themselves produce… that value is not required to extend to contributions made by external parties. Anthropic Claude is not a human, and as such in Australian law, cannot be held accountable for its actions or enter into contracts. Therefore it cannot be considered an “employee” or co-worker. I will not value its input here.

Even if LLMs did not have these issues, there is also the matter of cost. My workplace frequently cried poor when I asked for some equipment to do our jobs properly. This meant we were guessing with high-stakes projects, and this stress was causing me burn-out. Automated systems badgering me to submit timesheets (I always did, the “reminder” often arrived 3 days after I had submitted completed timesheets), or enable unwanted LLM features further intensified this burn-out.

The reality of my work had changed too… I’m mostly a firmware developer these days as JavaScript’s idiosyncrasies drive me up the wall, and the project I do this for, is a mature product. There isn’t that much work to do, just occasional maintenance of the code base. It didn’t justify 5 days a week, and I was stressing out trying to figure out how to fill-out a 38 hour week. So I dropped back to 3 days part-time… 22½ hours a week.

That worked well, there was a project to port this firmware over to a new hardware platform, that kept me busy… until late last year. I started my own little project to make a test rig for this firmware project, so we could have automated testing of firmware images — something I’ve wanted to do since about 2019. I got a prototype going, and had boards on order for me to build: the order got cancelled, because they were “not going ahead” with that project.

So that killed the only thing that has been keeping me busy the past few months. I’ve been thumb-twiddling the past few weeks since returning from long-service leave. I feel my head is on the chopping block, and I’m just waiting for the executioner to swing the axe. My thought is the sooner this happens, the better!

I have big concerns about the financial viability of my workplace if they start building their whole workflow on these LLMs. As Cory points out in that video, the financial situation of these tools is not great. Anthropic for example, is estimated to have earned US$22 billion, but have spent US$331.1 billion… a loss of US$309.1 billion… in the time they’ve been operating. Nikkei Asia believes may of these leading LLM providers are hiding as much as US$1.65 trillion in debt. Oracle has apparently over-invested so hard in “AI” that S&P downrated their credit rating to “just above junk”.

This is a financial house-of-cards waiting to be blown apart by the slightest puff of breeze. The financiers are waking up to this reality, and are now starting to ask hard questions of company boards. Much of LLM subscriptions are subsidised by venture capital, and when this funding is pulled, LLM subscription costs will skyrocket. Businesses that are dependent on these tools will be sent to the wall with crippling bills if they don’t reign in their usage.

Back at the tail end of 2010, I was hired as a contractor for the organisation that would, in 2022, itself be acquired by a larger organisation. I was a contractor until they hired me full-time-permanent in 2013. I had offers from more prestigious places, but this was a “devil I know” situation, and at the time, I was comfortable where I was.

A fatal accident would forever change the culture of the place, but we largely stayed “the same” up until 2020. Just prior to that fateful day, a colleague and I were working on a re-vamp of the office network. A project spearheadded and lead by the person who sadly lost their life that day. I poured my heart and soul into that project. When we did switch over, we had something that would be the envy of organisations today: all our data was hosted in-house, email, calendars… working in any standards-compliant client. A network where the client OS largely didn’t matter, you could use the tools that suited what you did.

After that event, I wanted to quit, but I kept going because at that point, projects that we collaborated on now had a bus-factor of 1… me. So I stayed, and kept going. A couple of years later, the firmware project that would become basically my sole work started… and that, as frustrating as it was at times… was engaging and rewarding. That too, I poured my heart and soul into.

2020 hit, and lo and behold, we’re locked down. This was a big shift, and it got frustrating at times. But I stuck with it. Sale of the business was also a big topic… with various organisations looking at acquiring us. I had read the horror stories of what can happen when a little firm gets swallowed up, but I was assured “nothing would change”. When the purchase finally did happen in 2022, I was apprehensive, but I made a point of “giving it a go” first. After all, maybe this would be different?

One of their first acts was to rip out the infrastructure I had poured my heart and soul into setting up… in favour of Office365. When this was announced, I couldn’t help this was merely being done for the sake of corporate fashion, much the same reason LLMs are being adopted today.

I thought, okay, Microsoft has changed their stance, maybe they’ve improved and their mail system isn’t the standards pariah it was in the past… wrong… we wound up with a mail system that couldn’t even deliver a email without subtly molesting it (enough to make PGP signatures fail). Later they’d manage to break sending on my account, so I couldn’t send an email from my work account, to anyone. Then they broke inbound email so I couldn’t email people at work from my home email address. I reported this, but yeah… years later, the problem persists. I’m fed up with it. I can talk to other Microsoft Office365 users from my home email server, just not my workplace. I can only see this as a deliberate decision they made to configure their server this way.

In 2023, my 10 years full-time service came up. I was starting to show signs of burn-out, but I figured if I quit at this point I’d never reach this milestone. It was a struggle to finally make it to that point, but I finally got there. A small amount of stress relief. With this, came some long-service leave. I also had a stack of annual leave I hadn’t yet used. After this milestone, the decision to drop to part-time permanent was made in an effort to recoup some work-life balance.

Now in 2026, I think my time there is done. The writing had been on the wall for some years. Spoken words might say one thing, but it was clear the direction was more towards front-end applications, which I can do, but I find the most frustrating. The incursion of LLMs into my work was the final straw. A year ago I delivered an ultimatum, it’s me or the LLM Late last year I warned that these tools would ultimately make my position untenable for legal and ethical reasons.

Looking around the office in mid 2026, hearing the chatter… I see they’ve chosen the LLM: I am done. Obsolete, and ready for the scrap heap.

Work-wise, I’m an oddball: while I have a preference for back-end work, I am a full-stack developer with talents that extend right down to circuit design. An “overqualified rejection” magnet. I also am a Linux specialist that’s starting to dabble in BSD and has absolutely NO interest in Microsoft, not even their online services. You can see my CV here.

That said, I’m not in a hurry and do not have a position lined up. I expect to take a break for the first bit of 2027 so I can process what was the last 16 years of my life… and figure out where I can go from here. Long-term employment might be my future, but we’ll see. I get the feeling yanking cables through conduits is going to be a better bet than software in 2027/2028 because it’s very hard to get a LLM to do that.