This post is strictly cathartic. It will be ranty, and you may disagree with much or all of this post if you are an “AI maximilist.” Consider yourself warned.
We have not achieved a singularity
The technological singularity occurs when technology can self-propagate improvement, and our current stochastic parrots can’t improve themselves. Instead, LLMs continue to improve via parasitic absorption of material made by humans.
LLM programming does not “resemble C”
I read a lot that transitioning to vibecoding is similar to the transition from assembly to C, and pushing back on vibecoding is akin to engineers in the dinosaur era that resisted programming in C. This analogy grinds my gears.
Of course, most of the people that parrot this have never written a line of C code in their lives, let alone assembly. C and assembly are both layers of abstraction in a hierarchy. What’s useful is the ability to reach into any layer of abstraction as needed, depending on the work that is needed.
I still consistently look at disassembly to understand why code is behaving or performing in a particular way. If I want to examine a crash dump and discern why the crash occurred, the disassembly gives me clues. Even if I didn’t ever write assembly directly, it’s a useful mental model to have. For engineers that don’t operate as close to the machine, you could drop down to this level of abstraction if needed, even if not necessary the majority of the time. Even if you never do this, you are guaranteed to rely on code authored and maintained by someone who does.
LLMs are not “a layer of abstraction.” You can’t peer into an LLM to gain any insight as to why something is happening. LLM inference is not compilation, and LLM code tends to obfuscate understanding, unlike C (and its successors) which improved comprehensibility of what the machine will do.
Quantity over quality
LLMs let people produce features and products that don’t matter. Maybe the feature makes for a nice demo or line item in a slide deck, but I am increasingly interacting with features I don’t want and cannot turn off.
I was on the phone with an overseas staff member who spoke with an AI-voice filter to attempt to obscure the accent. I don’t want this. I found the voice jarring, hard to understand, and frankly, dehumanizing to the other party.
Youtube offers canned AI responses to every comment posted on my channel. I don’t want this. If I’m going to reply, I’m going to take the time to read what was written, and reply with my own voice and ideas. This example dehumanizes me.
Everywhere, there are AI “features” that try to summarize or otherwise remix content into a form I don’t want. I honestly believe that much of the enthusiasm for vibecoding exists because a large amount of what is worked on did not matter and continues to not matter. Instead of working on meaningful products or product improvements, companies insist on making shiny trinkets that don’t make our lives better.
To hell with the environment
Moore’s law existed for a time, but there was always a hidden Anti-Moore’s law in place. As long as hardware got more powerful, engineering practices continued to get shittier. As Moore’s law leveled off, Anti-Moore’s law kept getting stronger, and we’re now at the point where the latter has fully overtaken the former.
LLM training energy costs are beyond unreasonable and each model iteration is more expensive than the last. Inference is being used with reckless abandon for meaningless tasks, and LLM-generated code is itself often wasteful. We are using more energy to generate more code that will, in turn, waste more energy.
AI stands for anti-intelligent
There is nothing intelligent about stochastic parrots, and there is no pathway for LLMs to get there. That won’t stop LARPing AI industry leaders from pretending there is though, or at least speaking with enough ambiguity to avoid outright denying the claim. The rhetoric about AI is so jumbled that Wall Street is thoroughly confused. That, or Wall Street knows full well that there’s no intelligence to be found here, but it’s more convenient financially to pretend otherwise.
LLMs don’t make us smarter. They make us lazier and dumber while somehow making us believe we’re smarter. Real engineers know that without the friction of failure, there is no real learning. This is true in math, physics, or literally any field. Skill acquisition is labor-intensive, but LLMs “solve” this problem by removing labor and providing a fascimile of skill acquisition.
To hell with copyright also
LLMs are trained entirely unethically. They should have been regulated out of existence, but instead, accelerated their theft with the blessing of nothing but economics and momentum.
People that say “well the cat’s out of the bag” or “there’s nothing you can do to stop it” have no fucking backbone. Just because other people have broken into a store doesn’t mean I have to break-and-enter along with them. The industry writ large has deemed that the use of AI is necessary, ethics notwithstanding, and anyone who disagrees are branded as heretics.
No clothes on this emperor
Companies are built on yes-men and yes-women, and the CEOs at the top are often the most enthusiastic yes-people of all. Can you imagine a pitch meeting where VCs ask about a company’s AI strategy and company leadership says “AI doesn’t actually factor into our plans at the moment.”? I can’t either.
Collectively, the tech world decided at some point that everything was going to be about decentralized ownership. Then, it was the metaverse. Now, it’s AI this-and-that. The industry is driven by FOMO masquerading as “thought leadership.”
Results are more important than anything
The hilarious swindle is AI companies producing giant reports showing that LLMs disproved or proved some unsolved math conjecture, or other reports demonstrating that LLMs identified however many millions of security vulnerabilities.
Nevermind that when playing theorem-roulette, you need an actual expert to determine if your latest spin actually landed on anything useful. Nevermind that we’ve always had static analyzers that required experts to discern false flags from real ones. Billing LLMs as being useful primarily to people that are already experts doesn’t sell quite as well though.
To hell with junior engineers also
Companies are routinely laying off humans with considerable expertise. Why bother training up a new generation of talent if we’re going to make misguided business decisions and lay them off anyways? Training up the next generation of engineers is someone else’s problem.
We have not achieved a singularity
We’ve instead achieved the exact opposite.
Content can and will be cannibalized, and the incentives to create new content and ideas is continually being diminished.
New code and content are being produced at tremendous cost at lower quality.
The workforce is continuously deskilling itself, and there are fewer pathways to sythesize new skills than ever.
Anti-AI heretics will be burned at the stake. AI must be the future, despite all signs that say otherwise.
Oh btw, fuck the environment and fuck copyright. AI will help us fix the environment problems we create (no, no it will not).
Oh also, how dare you steal MY model. I stole the data used to train that model first!
No, your non-AI product, feature, improvement, will not be funded. Think of the shareholders!
Short of achieving the singularity, all we’ve really done is achieved a world where everyone wastes time squinting to see if what they’re seeing is slop or not.