Clojurians Log v2

I’m not worried about the job part. I’m worried about the collapse of the commons/community/theft part.

By a similar token, even if the jobs remain, LLM induced brainrot might make them qualitatively and quantitatively worse positions to fill

This is a Clojure/software forum, so let’s think about this as it applies in our industry. Think about it in this progression: • will Uber (centralization) replace jobs? No, but the cabbie jobs pay less with the pay difference, and centralization of control going to Uber • will Waymo (automation) replace jobs? Yes, and the rides cost just as much with the difference going to Google. • will agentic LLM harnesses (automation) replace software engineering jobs? If the LLM companies are to be financially successful - yes. LLMs are sold to CEOs as a way to decrease labor cost, with Anthropic/OpenAI and the CEO splitting the savings. If LLMs don’t decrease costs in some way, the CEO won’t buy them. If you coded in Java, you know how Intellij IDEA makes you more effective both in quality and speed. But it’s like pulling teeth to get it approved as an expense because from the bosses’ measurement, they’re paying for expensive you PLUS a tool. But, as in outsourcing, if they can eliminate the expensive part - you - they’ll happily do it. To paraphrase Doctorow: it’s more likely that the LLM companies will convince your boss to fire 9/10 of your coworkers, split the salary cost savings, and change your job to be constantly reviewing code that comes out of LLM prompts. Back to the article, which I largely agree with: • will LLMs replace most jobs? Probably not, for the reasons in the paper. Creative destruction (Schumpeterism) can be a positive force over the long run. But as John Maynard Kaynes said: “The long run is a misleading guide to current affairs. In the long run, we are all dead.”

I think increasing productivity from technological innovation is a driver of decreasing birthrate. So if the overall picture looks stable it’s only because these two effects have been able to counteract each other somewhat. IIRC Swedes (which the article draws its data from) have very generous incentives for new parents in part because of this. So in my opinion, yeah, AI will be a huge problem for existing economic systems.

There are a multitude of issues around code generation with LLM. • the security impact of having code you do not understand in the system • using code without attribution in chewed up by LLM trainging • energy requirement • sending out your code to 3rd party with a promise that they wont train on it • the copyright issues around LLM generated code

I'll share mine. My own strong skepticism against it is almost entirely for technical reasons. I believe it is the antithesis of simplicity and all the adjacent teachings of Rich, such as the hammock. I am honestly surprised to see such a strong opinion regarding LLMs "stealing" content. I did not expect intellectual property to be highly valued in an open-source community to this extent. Perhaps my view is too colored by the FSF and the likes.

People do care about doing the right thing. They might not want to pay for it, but attribution is minimal effort.

would it be useful to distinguish between the different types of code generation, e.g. I use an LLM to generate at least some code most days but I direct it in terms of design/modeling and heavily refactor afterwards.

Any subjective distinction is welcome. I want to know the (subjective?) reasons people have for opposing it, in any serious shape of form.

As you probably are aware, some would frown upon your use of LLM as well. I would love to hear their thoughts on why.

for me, LLM’s are great for non-production code (e.g. linting rules, tools, configuration) since the risk is generally low, the item can be rebuilt easily from scratch and you don’t really need to understand how they work.

In my experience, code, once written, becomes a liability, (hence the prevalance of the term “technical debt”). The only asset is the understanding held by the people who operate and maintain the code. Normal coding processes, creates both an liability and the asset. Usage of an LLM to generate code creates a liability (code), without an asset (understanding).

Craft, art, and invention all take a back seat (or were simply ejected altogether). And the very dubious nature of the legality of how these tools are/were trained.

There is too much money involved.

@danieleneal this is great. Have you ever thought about putting that in a blog post? I am always practicing talking about tech in financial terms in order to speak in front of an executive board. I have never come up with such a concise version

I am asking for a blog post because I would like to have something to quote when writing about this - don't want to pass this of as my idea

ahaha feel free to pass it off as your idea but yes I get what you mean about something to quote. It's nice to point at something. I do want to write something about this more in depth, purely for myself, if I do I'll share it here, but it might be part of something more mystical I'm afraid. I don't have a tech blog, but I write about music and magick sometimes. So it may not be good for sharing with finance types.

OK :) I will refer to Daniel, a fellow clojurian then:sweat_smile:

heheh yes

Thanks for the encouraging words though, definitely motivates me to put this onto paper

I was hesitant to add to the conversation as it's all anyone talks about

but I like the idea of using writing to figure out where I stand with it all

Is software that much better these days? Shouldn’t it be up to the Altmans of the world to prove that it is? I’m just applying Occam’s razor here. (Personally I do think that LLMs with a little evolution have a future, but that future is not necessarily compatible with current economic systems)

LLM are getting better at doing complicated work, i.e. on systems, like software, which are deterministic and predictable. Experts can do analysis and know with confidence that doing X will result in Y. The LLM can do some of this, but they are nowhere near as good as an experienced engineer. LLM’s need heavy guidance, hand holding and highly constrained environments. Realistically they are good a clear complicated work, where the solution is obvious/standardised or where the impact of a bug is low (e.g. code which doesn’t run in production). But an LLM can not do any complex work, i.e. on systems which are not deterministic or predictable, like systems containing people, our customers, competitors, markets etc. Here the past (LLM training data) is not a predictor of the future, so an LLM fails. It can’t think, let alone think laterally or creatively. It can not innovate. This requires a human and experimentation.

My biggest objection is that for being so expensive to use, LLMs are infuriating to work with. No consistency, no best practices, there’s troves of AI usage out there and I’m really not sure what I’m to make of all of it. I’m used to reading books, working on exercises, understanding what I’m writing down and how the program I write works. The best LLM users may be professional gamblers, but I hate slot machines. I’d rather not have a slot machine in my workplace, especially where I’d have to rely on it to produce anything of value.