I’m fine with machine learning, neutral networks, and even specialize tools for industries to use “AI”.

I’m just not fine with it replacing human culture (art, music, memes ect), and misinforming people on basic topics.

We should differentiate between Generative AI (LLM, Video and image Generation) and those old school, valid and useful “AI” systems.

A bit of clarification for people who are OotL (simplified from a few courses on ML, but maybe helpful):

The reason generative AI has exploded is because of the discovery of the transformer architecture in 2017 which itself was an innovation on deep learning. Transformers are special because they enable extreme parallelization of tasks that used to be heavily serialized (namely ones that needed recurrent neural networks (RNNs) or their more advanced variants like LSTMs).

We aren’t starting to mix this up with “classical machine learning”, because classical ML isn’t about the task; it’s about how you’re performing it. ICR, voice recognition, etc. in 2026 are largely done using transformer models. Even those ML tasks that aren’t using transformers (which do have their own downsides) are usually still using a deep learning model (for example, KDE’s Spectacle screenshot tool has OCR, and I’m pretty sure they went with an LSTM model). And it’s also why separating it into “GPU or no GPU” is a wild oversimplification (classical ML models, such as SVMs, can also take advantage of parallelization in training).

Classical ML (supervised and unsupervised) are techniques and architectures like SVM, k-means, decision trees, etc. Deep learning incorporates these classical ML concepts (for example, the most basic type of deep learning model, a deep neural net (DNN), is also a “multilayer perceptron” – where a single-layer perceptron is a bog-standard linear classifier and the granddaddy of ML models). But still, “classical ML” in the field is “machine learning that isn’t deep learning”.

• If you want to learn how these models work and don’t want to be smacked in the face with textbook-style math that expects you to already know what it’s talking about, the YouTube channel StatQuest gets a lot of students through their ML courses.
• If you do want the textbook math, the MIT press published a book in 2016 which is free online by the authors, but ironically, it being 2016 means zero mention of transformers. (It does, however, still talk about generative AI in the last chapter, which is quaint nowadays.)
• And if you want to try hands-on, there’s an extremely user-friendly Python library called scikit-learn that’ll be more beginner-friendly than e.g. PyTorch or TensorFlow.

To expand on your post:

AI is an umbrella term for three different levels of intelligence:

Weak AI or ANI (Artificial Narrow Intelligence): This is where we currently are, and where we actually already have been for decades. It’s a system that has specialized knowledge in some narrow domain. That chess program you played against 20 years ago, your cars built in driving security systems, your assistant on your phone (even the old ones pre Gen-AI), and that school project you’ve coded that consists of 87 hand-rolled if-conditions (which you now know is called a rule engine), and every LLM or LLM-based product is an ANI. An ANI could compete with a human in its specialized domain, but it’s not able to transfer knowledge from one domain to another.

Strong AI or AGI (Artificial General Intelligence): An AGI does now have specialized knowledge in a single domain, but all human knowledge in all domains, and could perform comparably to a qualified human in anything you task it to do. This is science fiction, no matter how hard OpenAI wants to push its IPO. There isn’t even a definitive confirmation that we will ever achieve AGI, it’s “only” likely to happen at some point. This point in time would be the so-called singularity.

ASI (Artificial super intelligence): Basically the same as an AGI, just that it’s better in everything. Even more science fiction, even less to say about it. Another purely theoretical construct, unlikely to be ever achieved.

What’s important to mention: Any LLM is by definition only an ANI, and it can never be an AGI. LLMs have a single domain: stochastic text proposals. They have no intelligence, emotion, knowledge or transfer capabilities. An LLM cannot be an AGI - ever. Whether it enables us to create it at some point - maybe. It seems intelligent and knowledgeable, because modern hardware, clouds and some algorithmic advances enabled a system which is (very simplified) based on the same algorithms like T9-typing on your old nokia, but has an absolutely unimaginable amount of data fed in. It seems emotional, because it’s based on data produced by humans, and we all are emotional beings, and also because it’s a system designed to be manipulative. You will only continue your LLM subscription when your “LLM-Buddy” is helpful, not only in providing good answers, but also packaging them in text which creates an emotional response in you when you read it.

Agreed it’s important to seperate. It is hard explaining to people though, they still just call it AI in general

I have a tragically nuanced opinion that the technology is neat it’s just being used by evil men for evil things. A small locally hosted llm trained with legally obtained data would be all right.

Don’t get me wrong sew Aultman into a canvas sack full of rats.

One of the best AIs I have is in my thermostat for floor heating. Adaptive heating changes the time it should start the heating so the floor will be at the target temperature when a temperature schedule starts.

Machines should not be allowed to learn, it will only lead to evil.