If I tell people that I don’t to use AI tools for coding or my site, I see incredulous looks and get remarks from which I conclude that people see me in the backwards camp. I use the following rule that has proven itself in science, when I evaluate new tools and techniques: the more outrageous the claim, the more convincing the arguments for supporting it must be. Some simple tests should prove the claim or at least make the tool or technique look very promising.
Applying this rule did put me into the backwards camp sometimes, you know, the ’late adopters’. Yet the IT industry is an especially fertile ground for bullshit hypes. Often enough, when the dust settled and adoption ranged from pretty low to non-existent, I saved myself a lot of time and spared myself a lot of headaches (remember blockchain?).
I’ve always been interested in AI and its applications and check recent developments in that area from time to time. I’ve worked in logistics where I applied AI algorithms (Advanced Large Neighbourhood Search) to routing problems. I was never deeply involved in working with neural networks, but I do have a solid understanding of the main techniques used both in the symbolic and the stochastic camp in AI.
Regarding the current hype around LLM tools, I’m not impressed but disgusted.
Not only are the promises of an incredible increase in productivity nonsense. The side effects of using LLM tools for coding and writing in general are environmentally and socially harmful to an unacceptable degree.
When I program or write, I enjoy the process, which greatly supports focus. This is a re-enforcing loop that, given enough time, leads to a super-focused state known as ‘deep hack mode’. It would be stupid for me to choose a tool that interferes with my train of thought and my focus.
Yet this is exactly what LLM coding tools do constantly. LLM tools need constant baby sitting: prompt refining and if you use them for coding, there’s a whole range of other measures for baby-sitting, from providing instructions in markdown, providing documentation, code snippets and what not. Gone is promise that you can state what the code should do in an almost natural language, and you get the corresponding code, and also gone is your focus. What you get back more often than not is shitty results, and often it doesn’t even need closer inspection to see this. Now I’m not working on the problem, but I’m fixing the crap the tool created, which adds even more to the interference with my focus. Most times, by the time it took me to revise LLM generated code or writing in general I could as well have written it all by myself. Professional translators, who, like devs, have a profession in the domain of language, have raised similar concerns, both for the interference with their concentration at work and for the shitty results LLM translators produce.
So for the promised increase in productivity: I’m not focused on producing more in less time. My primary concern is not quantity, but quality. If a problem is not worth the time it takes to think about it thoroughly, then the problem is not worth being tackled at all. Fuck the busy-work mentality. Instead of solving problems it leads to patchwork that is creating more problems on top of the original ones. The whole LLM tool hype really caters to the busy-work mentality that has become so prevalent, and in IT, I see that it only exacerbates what we’ve been doing wrong for decades.
The vibes I’m getting here truly suck and generally, the term ‘vibe coding’ sucks monkeys ass. It’s fucking marketing speak. How come we hand the naming of what we devs do over to folks who come up with such stupid drivel?
This applies to almost all skills and abilities. If not put into use, they tend to decay. There was a time when didn’t drive a car for more than a year. I felt awkward when I drove for the first time after that period. Once I had a four months break from programming, I really felt clumsy during coding for the first few days after that break. And the code I produced was best described as clumsy. It was more difficult to even get started. I haven’t used LLM tools for a longer period of time, so I can’t talk from first-hand experience here. But I do suspect that extensive use would have a negative impact on my design and coding skills.
The same goes for writing in general. Thinking about how to write about a topic, mastering one’s native language – or a foreign one, are skills that have to be exercised regularly. Otherwise, such skills will decay. As I rarely write stuff like blog posts, I’m in a somewhat junior dev position here. I’m aware that I just don’t start out writing great stuff. Still, I’d rather just post the prompt instead of the output of a LLM. It would reveal much more about my intentions, and it would reveal much more precisely what I’m trying to convey than the stochastic, smoothly ironed dull output of a LLM.
This one is from the department of user experience: How come that no matter what kind of bullshit I feed into LLMs, they respond with “What a great idea!” or “Good point!”. I never got a reply like “What a fucking stupid question. Good you asked me first.”
My pronounced aversion to this behaviour may stem from that fact that I’m from Gen-X, the feral generation between Baby Boomers and Millennials (aka Gen-Y). While the sycophantic behaviour might cater to the narcissistic tendencies prevalent in today’s society, you can’t score with someone from Gen-X with that.
From low-code, no-code, ADF to outsourcing, from time to time crops up some “innovative” way to replace lots if not almost all of those pesky devs that take just too much time to churn out the stuff we want from them. Didn’t work out so well each and every time.
Now it’s the undergraduate students or junior devs who are targeted. It’s repeated over and over again that the code that LLM tools produce matches the quality of code that a junior dev authors. Some even go as far as to propose to work with a LLM coding tool like you’d work together with a junior dev. Both statements are an unjustified and pretty much moronic insult to junior devs in general.
I’ve seen quite some code of junior devs that blows LLM generated code out of the water. Junior devs do not hallucinate randomly (unless under the influence of drugs). They do need guidance now and then, but that’s neglectable compared to the baby-sitting a LLM tool needs. Yet, on average, they still produce better results than a LLM tool. Plus you employ a junior dev –one of the enthusiastic type, who didn’t just finish a coding bootcamp but who has been properly trained for the job– for some months and she or he will not only get better at programming but will also pick up domain knowledge that is specific to your business and that is way beyond of what you can expect from a LLM.
Whoever comes up with the argument that LLM tools produce code much faster: yes, they do produce shitty code much faster. Producing lots of shitty code in short time has always been the main contributing factor for projects becoming a train wreck that everybody hates working on and that ultimately fail in the long run.
So it’s just plain stupid trying to replace junior devs with LLMs and expect senior devs fixing the crap that LLMs produce.
Now that’s a good one: If you don’t adapt and use it, you’ll be out of the game. Wait a minute, where did I see this before? Oh yeah, advertisements. That kind of bullshit ad companies crap into peoples brains in order to instill fear. Fear of losing your job, social and/or financial status, fear of loosing it with your preferred sex, fear of … whatever. Now go and use our product, and you’ll be safe. Well, not exactly safe, but you won’t be left behind tomorrow. The day after tomorrow? Subscribe.
Two major players in the AI field, Meta and Google, at first, are no IT companies, but ad agencies. Another major player, Microsoft, is known more for shady business practices than for innovation. These companies are known for screwing their users and customers big time. The companies working on LLM tools have blown vast amounts of money for quite some time and the returns have not been particularly high so far. No wonder you’re told you’ll be losing your job if you don’t use their products.
Like a one or two decades or what? More often than not, bold claims made in AI range from the hilarious to complete bullshit. Within AI always were some dazzling figures that made bold claims of what AI holds for the near future of mankind. Pretty much of it turned out to be nothing but hot air.
In 1970, Marvin Minsky predicted that within three to eight years, machines would reach the intelligence of an average human being.
In 2005, Ray Kurzweil, principal researcher and “AI visionary” at Google, predicted the same in his book The Singularity Is Near, but now for 2030. On top of it, he predicted that by 2030 human brains can be directly connected to the cloud via nanotechnology, enabling us to do funky stuff like reading and writing emails just with our minds. He also stated we would become god-like, “We’re going to be funnier. We’re going to be sexier. We’re going to be better at expressing loving sentiment”, as Kurzweil put it.
Yet folks like Kurzweil forget to ask if we even want that. Given the inherent security problems in IT systems, I do not think that this brain-cloud-connection thing would be a particularly good idea, if it would work at all. But hey, never mind, I don’t see either of those things coming within the next five years. And not within the next 50 years.
In the last decade, there have been amazing achievements in the field of AI, like robots doing a flip in the air and other acrobatic stuff. The people involved just build the stuff, show it off, and the crowds goes like “holy shit!”. And the people involved just go on, building new stuff rather than pester their environment with ridiculous claims.
The sad thing is that hypes like current one around LLM tools can damage the reputation of this technique or AI in general.
In 2024, Eric Schmidt held a talk to Stanford University graduates who listened in awe as he talked about economic and social disruptions that AI will hold for the near future. Compared to Minsky’s and Kurzweil’s ideas, he toned it down quite a bit, but I still don’t buy into it.
This is not what we need: An old guy like Eric Schmidt telling young folks how this incredibly ingenious new technology will be disruptive (or, in less marketing speak: fucks up everything they know), then telling them how to avoid getting trampled over or left aside when the stampede is on.
Rule of the thumb: disruptions come unannounced.
The content that is scraped by AI companies to train LLMs, from source code, blog articles, newsletters and whatever, comes from people that put some effort into authoring that content. Most of those authors are not happy about their content being scraped and used for LLM training. They didn’t consent to this. And this shouldn’t be an opt-out option, with software to block those scrapers being the only existing opt-out method, but an opt-in option.
The AI companies can’t just say “Well, you put it on the internet, so your content is meant to be used by others.” Yes, it is meant to be used by others, but by ‘others’, I mean people, and I mean people using my content directly, not by some company mixing it with other bazillion sources, mangling it up and then selling it repackaged to others.
You don’t just copy content and use it in your own work without citing the source. You don’t just copy content and monetize on it before you got the original author’s permission and without compensating her or him. But that’s exactly what’s happening on a large scale right now.
If I’d use LLM tools, no matter whether it’s a free or paid service, I’d quietly consent to this. Given the recent surge of software to block LLM scrapers, I guess I’m not the only one who is pissed about it.
And there are not only ethical but also legal issues: How can you make sure if you put code into your codebase that has been produced by a LLM, that this code is appropriately used, so that you do not commit any copyright infringement? You can’t. You’re walking in a grey zone, eventually crossing the line. Maybe the providers of these tools think “let’s just go, we figure that out later”. This might not be an option for you or your company.
Yes, I know the political correct term is ‘developing nations’. Fuck it. You don’t change the condition by rebranding it.
After all kinds of content has been aggressively scavenged from all kinds of sources, the content has to be classified to make it digestible for the LLMs. This is primarily the job of so-called click workers from –using another politically correct term– ’emerging markets’ like Kenia, Madagascar or Uganda. These click workers form a cheap throw-away-after-use labour force that forms the back-bone of LLM model training by classifying whatever content.
The classification is a tedious task, it is poorly paid and the working conditions are terrible, like not being allowed leaving your desk during work.
The content to classify ranges from ordinary everyday situations to disturbing material with extremely violent content: images, videos or detailed descriptions of murder, suicide, torture, self-harm and sexual violence. These click workers have no feelgood-manager to turn to, let alone access to help with mental health issues should they arise.
Some tech companies black-list countries like Madagascar, yet it’s well known that accounts for click-working are available everywhere through middlemen.
I want to see the IT industry working on getting rid of the exploitation in their supply chains, instead of extending and deepening it. Once again, regulations currently in effect in this area have proven to be not effective enough.
It’s the individual decision of each and every potential user or customer, that, added up to a large-scale market, decides how much revenue a corporation has – which directly translates into power. That’s why those corporations need the brainwashing I write about above so badly.
There are several factors to be considered for evaluating the environmental impact of LLM applications:
The setup and running of LLM based applications needs extremely high amounts of electric energy and water (for cooling). The inference phase, in which the model is trained, is heavily compute-intensive. During deployment, it’s the massive scale of requests that amounts to high energy consumption.
Here are some estimates of a LLMs power consumption, compared to some other applications:
| application | electric energy | carbon footprint |
|---|---|---|
| Laptop, 1 hour | 0.03-0.09kWh | 0.01-0.05g / 0.13 (1) |
| 1 search engine search (without LLM) | 0.0003kWh | 0.2g |
| 1 LLM request | 0.0003kWh | 0.06g / 2-4g (2) |
| 1 hour TV (3) | 0.03-0.15kWh | 3g |
| 1 hour Netflix | 0.12-0.24kWh | 55 - 110g |
(1) includes carbon footprint for manufacturing, with the laptop being in service for 6 years.
(2) includes carbon footprint for LLM training. Does not include carbon footprint for hardware manufacturing.
(3) standard terrestrial broadcast, LED or LCD TV.
Compared to TV or even Netflix, LLMs don’t look so bad. Compared to a search engine search or one hour of laptop use, LLMs look terrible. Another resource munching application on top of those we already have is not what we need.
We are beyond the point where the power concentration in megalomaniac companies is destroying the very fabric of our societies. Big IT corporations have run unchecked and wreaked havoc for long enough to prove they better stay in check. It’s obvious that regulations alone won’t do. We better check ourselves what tools we use –and for what purpose– instead of letting it being shoved down our throats by a fake consensus that this is the new cool thing or this is how it has to be done.
Power grows where attention goes. I’m past the point of being interested in the latest innovations from FAANG or MANGO companies (Microsoft, Anthropic, Nvidia, Google DeepMind, and OpenAI). Now, I just briefly note what kind of crap they’ve come up with this time. Not only can we do without their products and services, but we are better off not using them.
Once again: It’s the individual decision of each and every potential user or customer, that, added up to a large-scale market, decides how much revenue a corporation has – that is, how much power a company has. The road to hell is paved with convenience. A big share of this convenience is advertised by mega-corporations that feed of laziness in thought and action.
Hypes are a recurring theme, they come in cycles. I’ve seen the hype of an AGI (Artificial General Intelligence) being announced as just being around the corner before. Which obviously wasn’t true. I’ve seen the recurring theme of developers becoming obsolete. Which also was obviously not true. But after a few years, people forgot how the last hype had vanished into thin air. What was left at best were some tiny improvements. From a business perspective, it totally makes sense to make outlandish assertions in order to secure funding. But what’s good for a business, is not necessarily good for its customers or the rest of those involved.
Remember the saga of the ‘prompt engineers’ with insane salaries? The party was over before it even began.
And yes, this is just another hype. The real ‘advancement’ is in the destructiveness of the cons, which heavily outweigh the pros.
I didn’t check on this thoroughly, but I wouldn’t be surprised if that whole LLM tool thing crashes. There are insane amounts of money shoved into the companies that work on it, but the tools don’t deliver and the returns so far have been low. The whole thing reminds me –partially– of the blockchain/cryptocurrency story.
Noema, Ray Kurzweil: In The 2030s, Nanobots In Our Brains Will Make Us ‘Godlike’
BLOOM: A 176B-Parameter Open-Access Multilingual Language Model
Journal of Machine Learning Research 24 (2023) 1-15, Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model
The Guardian, Meet Mercy and Anita – the African workers driving the AI revolution, for just over a dollar an hour