IHeartBadCode

Then we are completely on the same page! I actually mentioned earlier in our thread that I hate the notion of calling these specific, highly useful operations ‘AI’ for exactly the reasons you just laid out.

My previous replies were just looking at it through a strict, literal computer science taxonomy lens. But my original comment literally points out the very things you bring up.

We’re fighting an uphill battle with the vast marketing hype and it’s good that you can understand the nuanced difference between all the things that have gotten rebranded AI. My point isn’t to try and use past technology as a way to validate what the techbros are doing but to highlight how muddy the waters have become.

I closed out my original comment with a hope that perhaps the more utility based work that’s had solid proof of usefulness doesn’t get tossed out with the bathwater so to say.

Again, I’m not apologetic of what has come from that hard work, more lamenting that everything has become the fuzzy mess it is.

By all means, head back to my original comment and let me know if there’s any part where I’ve come off indicating embracing this thing that I would say you and I agree with.

We are saying the same thing, just at different layers of the tech stack. Your description of an SGD-driven model running on an NPU is the precise low-level math and hardware that allows the Deep Reinforcement Learning (DRL) agent to function.

What’s happening here is that you’re falling into a trap many people now mistakenly believe that if software doesn’t generate text or use a Transformer, it isn’t “real AI.”

But that’s only because LLM have become the dominating conversation piece of AI.

Saying ‘that’s not AI, that’s an algorithm’ is a fundamental misunderstanding of computer science. All AI is built out of algorithms. Neural networks, stochastic gradient descent, and transformers are all algorithms.

The line between ‘traditional programming’ and ‘AI’ is machine learning the ability of an algorithm to optimize its own internal weights based on environmental feedback rather than relying on hard-coded rules written by a human.

Would you say when Google’s AlphaGo beat champions at Go that it wasn’t AI? Because it didn’t use language transformers either. By definition, a DRL agent that uses a Neural Processing Unit (NPU) to continuously calculate optimal radio frequencies via Stochastic Gradient Descent is text-book Machine Learning.

But the thing is I don’t blame you for the confusion. Marketing hype leads many to this same trap of prerequisites for particular transformation to qualify as quote/unquote AI. But technically that just isn’t true. I do enjoy this conversation we’re having as it does highlight common misconceptions.

Technically, yes, it’s an algorithm but all AI software is built out of algorithms. The critical difference is that traditional algorithms are fixed, static instructions written step-by-step by human engineers. Deep Reinforcement Learning (DRL) is a self-learning algorithm. Instead of a developer programming exactly how to handle every single wireless interference scenario, the DRL model acts like an AI agent. It continuously learns, adapts, and teaches itself the absolute best optimization paths purely through real-world trial and error.

Modern dense networks face a ton of unpredictable interference and variable traffic patterns. Wifi is a victim of it’s own success. It’s literally everywhere and thus all of these sources clobber the airwaves around them. This makes the traditional methods for traffic management and resource allocation of the airwaves too complex to fully implement.

However, your usage of LLM isn’t correct here. Wifi 7 doesn’t use a large language model, it uses what is called a Deep Reinforcement Learning (DRL) model. Wifi 7 isn’t trying to be generative, it’s being administrative. It’s looking at the airwaves as they are, and attempting to find an optimization for the current situation it is in.

In most cases the wifi coverage is not such that the NPU needs to step in. Traditional methods for transmission can be used, but in cases where you’re walking in a mall or in an apartment complex. You have tons of APs vying for the same resource. AI is used here to listen to what’s going on out in the world and come up with a method to target the highest bandwidth that can be achieved.

The intellectual poverty extends to the economics. Acemoglu has found that only 4.6 percent of tasks in the economy are currently cost-effective to automate with AI. His estimate for AI’s total productivity impact over the next decade: 0.66 percent. Goldman Sachs projected seven percent in 2023, before we began to see the shape of this thing. McKinsey projects between 0.5 and 3.5 percent annually.

Yeah this is the thing I keep saying and everyone on social media keeps saying “Boo, AI bad.”

The reality is that AI presents a very real, very useful thing. A combination of linear algebra, calculus, and probability. There’s lot of use cases for it. But putting it in all those use cases is not undoing the entire economy. We don’t need anywhere near the AI data centers that techbros keep saying we need.

Things like Bayes theorem have uses, I hate that it’s gotten lumped into AI. Optimization via loss function is incredibly useful in a lot of domains. But none of them can one-to-one replace human beings and it’s wild watching all these “captains of industry” lose their collective shit.

Someone is catastrophically wrong, and the people spending the money are not the ones with the Nobel Prize.

Yeah, it’s the techbros. They’re taking really useful mathematical operations and functions and doing neat albeit useless tricks with it. Anyone who understands the actual fundamental math behind these models and doesn’t get too carried away with it will tell you, the reason…

Over ninety percent of firms surveyed in 2025 reported no measurable impact on employment or productivity despite a quarter-trillion dollars in AI investment.

is because people are being handed something that they have no idea how to use and the way they’re being told to use it, is actually wrong.

Engineers who did the spec for 802.11be (wifi 7) understood the nature of a channel matrix operation in MU-MIMO. Singular value decomposition benefits from vector dot multiplication and gradient descent minimization. This is absolutely perfect for AI, it’s dang near what you’d want to use it for. And that’s why Wifi 7 routers come with an embedded model and NPU to run the model onboard. 4096-QAM benefits from linear transformations through a Euclidean space. Mass matrix operations to perform those transformations are ideal with AI.

Which is why I hate this notion that we’ve called these specific, highly useful operations, AI. Because they have way more application than neural networks, but since you have the hardware, Wifi 7 LDPC uses GAANs, because you’ve got the hardware. MLO and studying the interference in a particular space are also perfect for neural networks.

There are all these uses and honestly it’s crazy watching this insanity that is people like OpenAI, Claude, and so on. There’s no way they’re going to make good on their promises of being able to fire everyone. It just doesn’t make any logical sense when you look at the various domains of math that underpin AI.

And maybe that’s because, the people who fly off the handle with AI, are people who take this math and see the human brain in those formulas. I think that’s wild take, but my understanding of biology is limited. But I feel our brains are bit more complex than a two year study in College Calculus and Linear Algebra. But that’s just my, not very well studied in biology, opinion. But I think that’s where these people fly off the handle, they see activation equations, ANN layer transformation equations, and what not and think “human brain”. And it’s that thinking that’s drove them to this insanity.

There’s no way the AI industry as it is can keep up. It is bound for collapse. But in all of that, the underlying math is still very important and very useful, and maybe that will get relabeled to neural networking or linear optimization? But what we are seeing is a party trick that can be done with these equations and it’s apparently a trillion dollar party trick.

The issue is, I don’t think the generative is here to stay. But it’s clear the industrial stuff has solidified. Wifi 7 has AI on an NPU as part of the spec. Because when it can learn the radio signals in the air and the interference it needs to avoid, you get vastly better wifi. BGP is a great protocol for routes between ASes, but it’s dumb protocol, it relies on static rules and metrics. AI overlays are making better routing choices based on learned patterns of traffic, ISPs have seen gains by better optimization.

The industrial grade AI has an objectively proven track record. People can downvote me all they want, none of that matters in the light of fact. AI is in a lot of programming. The stuff that’s proven is the boilerplate. The industrial AI. The generative AI where you ask a few words and get a wbesite, yeah that’s smoke and mirrors. Bridges aren’t useful while they’re being built, they’re only useful after they’re built.

Just like we saw wizards to churn code out back in the day, we’re going to see that with AI in coding. Is the AI going to code at least 50% of the program? Not likely. But having a ban on even 1% AI in code is just unrealistic. One, it denies the reality that we’re already using some AI in tech and coding. Two, you better believe that bad actors are going to be using AI to punch holes in software. And three, it’s completely unenforceable. Flathub lacks the staff to actually police that policy and so it’s going to devolve into Flathub chancing rumors and “hints” on which program has AI in it or not.

And it’s silly because when we have tools and use them correctly, they make our lives easier. Is the 100% generative AI garbage at coding, absolutely. But things like technical documentation, generating API docs, commenting DDL, and so on. Things that we programmers aren’t paid enough for. We talk about commenting our code, who here has time to do that properly? We keep trying to invent all kinds of new ways to “auto-doc”. But now we have a generic documentation generator.

That is an absolutely terrible standard. In fact, that’s technically not even a standard. The “I know it when I see it” measure is literally the logic used in censorship. It allows cognitive biases to seep in with no check. A lack of hard metrics means that there’ zero ways there can be any objective consistency. And finally, this kind of rationale makes the biggest sin that I can think of, “non-falsifiability”.

Whatever people’s opinion on AI are or are not. This logic should wholly be rejected in every instance it is brought forth. It is literally the antithesis of rational thought.