And yet, China is using AI.
…I… don’t know what to think about that.
…I really don’t.
Because it seems that AI is just a scam.
It may “exist” but what it can do is a scam.
Maybe China thinks we have to use it just to “keep up” with the Western powers, but I dunno.
Anyway, interesting discussion with Adam Conover and Ed Zitron. It’s long, but you can listen to it while doing other things. And the comments are interesting too, but then again, there are also trolls in the comments as well (AI supporters here and there).
Frankly, though? I oppose AI. I’m anti-AI. I’m anti-AI in China and anti-AI in America and anti-AI in the whole damn planet.
I think this is a perfect illustration of how technology ends up being applied under different social and economic systems. The reason AI is problematic in the west is due to the fact that it’s applied towards finding ways to increase the wealth of the oligarchs. On the other hand, China is using AI for stuff like industry automation, optimizing government workflow, and so on. AI is just a tool for automating work, there’s nothing inherently bad about it. The question is how this tool is applied and to what purposes.
I’m not so sure about that. Your analysis correctly identifies that it is being used in the West for nefarious purposes, but I honestly think even on the technical merits it is a flawed technology and a waste. DeepSeek is more efficient, yes, but it is still a flawed technology that I do not believe they should be using
I wouldn’t say AI (or pattern-replicating models resembling AI) is flawed. It’s a great tool for saving time and automating certain processes.
The problem is the myriad of grifters who appeared, mostly in the West, trying to sell it as a cure-all snake oil.
For instance, there’s a massive push in the EU to insert AI in education, but with little regard or planning on how to do it effectively. It would be a great tool if we were to feed AI with our curriculi, then ask it to update it to current knowledge (e.g. in science), come up with suggestions for better delivery of certain topics, eliminate time wasted on erroneous, repeating, or useless topics and improve our schedules for other topics (e.g. teaching Romeo and Juliet in Languages, and at the same time go through the history of 1400s Venice in History). These things could be done using commitees over a 5 year period. Or they could be done by AI in a day. Instead though, we get to have handsomely-paid private contractors organize days-long training sessions over how to use AI to draw a picture, because it might make a presentation to students slightly more exciting.
Honestly even your idea of having an LLM “update” a curriculum just makes me annoyed. Why does everyone automatically give authority to an LLM on perhaps one of the most important societal functions, instead of trusting teachers to do their job, with the decades of experience that they have in teaching?
Is this what we want? AI generated slop for teaching the next generation because it’ll get it done in a day?
I find it works well for many purposes, particularly R1 variant. I’ve been using it for lots of stuff and it saves me time. I don’t think it’s flawed technology at all, you just have to understand where and how to use it effectively just like any tool.
I would argue that if your goal is to get an output that is the statistical mean for a given input, then sure an LLM will generate a set of outputs that statistically go together. It just happens that you throw enough data at it and waste a small country’s annual energy consumption then of course you’ll get something statistically similar. Congrats. You did it.
The energy consumption has already been reduced drastically by reinforcement learning, mixture of agents, quantizing, and other techniques. We’re literally just starting to optimize this tech, and there’s already been huge progress in that regard. Second, it’s already quite good at doing real world tasks, and saves me a ton of time writing boilerplate when coding.
So, that’s another thing that I wonder about. All these LLMs are doing is automating boilerplate code, and frankly that’s not really innovative. “Programming via stack overflow” was a joke that has been in use for nearly two decades now (shudder) and all the LLM is doing is saving you the ALT+TAB between SO and your text editor, no?
If you’re doing a TODO app in Angular or NextJS I’m sure you get tons of boilerplate.
But what about when it comes to novel, original work? How much does that help? I mean really how much savings do you get, and how useful was it?
The reality is that most of programming isn’t all that innovative. Most work in general isn’t innovative. Automating the boring stuff is literally the whole point. Meanwhile, it’s far more sophisticated than copy pasting from StackOverflow. It can come up with solutions in a context of a specific problem you give it, and the chain of reasoning DeepSeek R1 produces is actually interesting in itself, as it reads like a chain of thought.
This itself can actually be useful for doing novel and original work because it stimulates your thinking. Sometimes you see something that triggers an idea you wouldn’t have had otherwise, and you can pull on this thread. I find it literally saves me hours of work, and it is very useful.
For example, just the other day I used it to come up with a SQL table schema based on some sample input JSON data. Figuring out the relationships would’ve taken me a little while, and then typing it all up even longer. It did exactly what I needed, and let me focus on the problem I wanted to solve. I also find it can be useful for analyzing code which is great for getting introduced to a codebase you’re not familiar with, or finding a specific part of the code that might be of interest.
It’s also able to find places in code that can be optimized and even write the optimizations itself https://github.com/ggml-org/llama.cpp/pull/11453
Based on my experience, I can definitively says that his is a genuinely useful too for software development.
How likely is it that this JSON structure and corresponding database schema is somewhere in the (immense) training data. Was it novel? New?
Like I just have continual questions about an LLM doing 3NF a priori on novel data.
Like if we just outright say that LLMs are just a better Google or a better IntelliSense that can fetch you existing data that it has seen (which, given that it’s basically the entire Internet, across probably the entire existence of the Internet that has been crawled by crawlers and the Internet archive, which is a boggling amount) instead of dressing it up as coming up with NEW AND ENTIRELY NOVEL code like the hype keeps saying, then I’d be less of a detractor
You seem to think that the way these things work is by just simply pulling up chunks of existing code from a big db. That’s not actually what’s happening. It’s building a solution for a particular context based on its training data. This is not fundamentally different from how a human developer solves problems either. You learn through experience and once you’ve solved many problems, you recognize patterns and apply solutions you learned previously in a new context. It is writing new and novel code when it produces solutions.
I think the difference between generative AI and AI as a whole needs to be made here. DeepSeek is generative AI, along with all the issues that goes along with that but there are very applicable AI based systems for professional, industrial, and scientific uses. The level of data that machine learning systems can analyze provides a usage far beyond the problems that are certainly inherent to generative AI.
Does it? I would not consider LLMs to be AI and it’s unfortunate that a marketing gimmick has been turned into a colloquialism to the point where they have stolen the term AI and renamed the original pre-OpenAI concept to “AGI”.
Regardless, the point I am making is that the current hype surrounding LLMs is unwarranted.
But here’s where I disagree. I think Machine Learning at least has a relatively modest marketing pitch, where you feed it data and based on some training and inputs (yes this is indeed similar to an LLM) and you will get a reasonable estimation of whatever you are seeking as an output, based on historical data. Nobody is running around claiming that my monitoring system that has some ML in it for CPU and temperature and load averages is suddenly going to turn into God like Sam Altman and all these other wackos want to happen with LLMs.
Of course, but I don’t think we should discard LLM based AI either. It’s a piece of a bigger picture. One of my favorite examples of it being used effectively are neurosymbolic systems where deep neural networks are used to classify noisy input data, and then a symbolic logic system is used to reason about classified data.