Bullshit said by people who do not understand that LLMs are NOT AI

NOT AI, not even a little

The people who espouse such nonsense as “if you don’t use AI you’ll fall behind” believe that using AI will become an absolute necessity in the future, and that those who use AI now will have more expertise in using AI than those who don’t, giving them better prospects. Those same people end up saying “you’re just not prompting it right” when you point out flaws in the LLM’s responses.

From what I’ve experienced it’s mostly managers saying this. They’re afraid of competitors taking over the market because they can be more efficient with AI. They say this without any idea of how exactly that would work.

They think there’s some skill necessary to use LLMs for some reason

My experience of a decade in corporate US companies:

Quantity > Quality; no matter what the marketing line says. People who speak a lot in meetings and send a lot of emails are regarded as better at their job, substance is only somewhat relevant. And LLMs are great at producing lots and lots of well sounding slop.

They (pretend to) believe, either due to experience or marketing or a combination, that AI* will be such a ubiquitous tool within their lifetime that “not using AI” would be like “not using electricity”. There are people that do not use electricity, even when it is available, today, but they are broadly considered “behind” the times and out of touch with broader (including global) society. AI advocates are unlikely to always have the same idea about exactly what the advantages are, though speed is often among them.

The actual scientific studies we have done are limited, but so far do not back up these claims. For example, AI assisted programmers self-report being faster, but are actually slower at the same tasks than a similar cohort without AI.

I have a fairly strong anti-AI bias, but my limited experiences with them have been quite poor. Maybe I’m just asking the “wrong” questions, but my experiences have had them either give an response that is internally inconsistent, contains easily refuted factual inaccuracies, or (for tasks with no reality/logic testing) absolutely the least creative, most middle-of-the-road text. I’m also aware of the human rights abuses that Amnesty International has reported around generative AI, the copyright “minefield” around models where you do not know all of the training data, and how we’ve lost at least a decade of sustainability goals because of the desire for “more compute”.

So, keep my bias in mind when considering my reply. I might already be “falling behind”.

I can think of some ways to use an LLM that are interesting to me, but for now I’m staying away from the tech until the bubble bursts and the hype disappears. It’ll be easier to make informed decisions about what, if any, to use then.

(If it really does improve society, I’m sure I can be retrained then, or at least I can be taken care of one AI-incompatibility is recognized as a disability. /s)

AI: They probably can’t define this term, but likely to have a specific product in mind.

FOMO, and the temptation of being able to get the speed of AI without the drawbacks of “vibe coding”.

Reality is much more nuanced.

You can get a massive boost in your productivity if you use dependencies or just steal code, compared to writing everything from zero. I personally just reuse my own code snippets, and even started to put regularly reused code into their own libraries (next up is to make stuff for parsers).

I also seen cases, when “AI assisted” coding had the same amount of issues as fully vibe coded stuff, as the person really wanting to cram the chatbot into their workflow did that trusted Grok too much with understanding the documentations, so he just ended up trying to implement stuff based on Grok’s summary, which was wrong at places.

If you ask: artists also sometimes fall for this trap, though much less often.

I think a good alternative explanation is this: AI is here to stay, so knowing how to use it in whatever form it takes in your life as a tool is good knowledge to have.

Typewriters are close to extinction, but most everyone these days uses a computer to write documents. Those that refused to learn how to use a word processing application on a computer were “left behind” because now almost all documentation is written digitally.

im going to put it easy: You work doing some kind of research or whatever on daily basis. You used to do that task in maybe 5-6 hours, same as your colleagues. Now all your colleagues use AI and do the same task in about 20min (even if the result is crap it does have some good takes), so the CEO sees you are the only one who doesn’t perform equally because you choose not to, and you’re fired.

So yeah… it’s crap, it will make things worse, but for now you have to use it wheter you like it or not.

So realistically, it’s kinda BS pushed by FOMO farmers with hustle culture vibes.

The reality is that the acceleration of AI is reaching a speed where by the time you’d build your whole setup to work with AI there will be a new AI who can set the whole thing up in an afternoon.

There is a learning curve to things like how a model processes information and how to set them up for success. But it’s fairly easy to learn and likely to shift quickly.

The biggest gap that I see between people who have success with models and those who have a bad time is how you treat the model.

If you are a jerk, I hope you don’t like your production DB or home directory too much.

Things that seem crazy to people who aren’t following things, like giving the model breaks to avoid context burnout or encouragement vs yelling in all caps, go a long way.

TL;DR: The methodological best practices are likely to change so fast you won’t need to do much catching up, but the relational dynamics of working with AI will likely remain pertinent into the future.

Its people falling for propaganda.

The fact of the matter is, AI can be useful for minor coding help. The reality is its stealing advice from one or a few sources and spitting out an answer based on them, but it can be a quick delivery of advice.

The major problem is

1. AI (these LLMs) hallucinate ALL THE TIME. Relying on them to code for you is fucking moronic. Every coworker I’ve seen who does vibe coding is constantly getting dunked on, their pull requests are torn to pieces, and its baffling why they still get a paycheck.
2. AI can only, in the best circumstances, do exactly what you tell it to do. That is to say, you need to explicitly spell out what you want for it, including explicitly covering edge cases. If you don’t, you end up with situations like Microsoft making Notepad code executable (which actually fucking happened.

Promotion of AI is total horseshit by the fascist tech broligarchs being done so they can replace the worker and, and I’m serious, achieve immortality by transferring their consciousness to machine. That isn’t a fucking joke, that’s an actual belief they hold. The AI and general tech CEOs are all in a massive death cult. This is real.

It’s from the same people who tell you to “GET IN EARLY” for whatever hot trend is about to take off, usually involving markets and trade.

People who say that don’t mean anything. They are either trying to sell something or repeating bullshit they heard somewhere.

It really is that simple.

And remember, the AI salesmen will never, ever, define “AI”. They can’t, because the minute they do, their product sales will evaporate. If all you have is vaporware, you desperately need to pass it off as something else.

They’re very very very few tasks in which AI is great. Everything else is just a waste of money and unnecessary at the cost of destroying the planet.

I never understood that argument. The same idiots claim AI will democratize everything, it is inevitable. That one would only need to know natural language, “coding is over” or ‘x’ is over. If it does democratize ‘x’ for everyone with only requirement being natural language, who would even get left behind ??

I tried using a wrapper LLM for a specific software for a project, and sometimes it gives correct answers, sometimes gives the diffuclt way to implement something, sometimes it hallucinates. At first, I thought great, I wouldn’t have to sit through an entire playlist to get the gist of how to do something mildly similar. But recently it has been giving wrong answers. Today it gave answers beginning with “handoff - DENY}” and answers just kept cut off. Sometimes asking a different question, it just resumed the cut off answer from half an hour ago. The stupid piece of shit was wasting my time!! In the end, I just figured it out on my own.

I think NLP is amazing, a computer being able to understand what exactly I’m trying to do/look for is useful. But my work shouldn’t be dependent on an LLM’s “mood”.

Before anyone says just read documentation/tutorials, this is a niche vfx software which doesn’t have good docs and tutorials are mostly outdated for latest versions