{"slug": "building-a-laser-with-a-chainsaw", "title": "Building a Laser with a Chainsaw", "summary": "AI systems consume enormous resources, and their widespread adoption is accelerating environmental damage without delivering commensurate benefits, according to a critical essay. The author argues that current large language models are not a path to superintelligence and that corporations are prioritizing profit over sustainability and accountability.", "body_md": "AI costs a lot to run. Like an extreme amount.\n\nThe dollars spent aren't really the issue though.\n\nThe issue is we have a theoretically finite timeline as a society if we don't manage or limit certain activities.\n\nMoving forward, I'll be using terms AI and LLM interchangeably and maybe people will take issue with that. I'll be doing the same with AGI and superintelligence. I think this approach is consistent with the major corporations and voices in the space.\n\nAn argument for the benefits of AI is that it has helped make some radical gains in certain verifiable domains. Mathematics is an example. I take issue with this because a lot of radical gains in verifiable domains do not feed into outcomes that can help make the resource drain worth it. Disproving an arbitrary conjecture rarely leads to concrete progress. For example, unless AI can resolve environmental or climate issues, the fact that its widespread adoption and consumption are speeding them up is a problem.\n\nNow, there are some domains where we do indeed get breakthroughs like material science. This is evidently not the focus of these corporations. The mass marketing and pushing of widespread use are not in line with the generating cost efficient breakthroughs. If the models were provided only to perform research into better solar panels or carbon-capture materials I might feel differently. But an LLM is pushed on me in every Google search and a bunch of the apps I use (sometimes unwillingly), and while I can sometimes opt out, they are enabled by default.\n\nThe idea that educating a human engineer takes two decades and therefore AI is more resource-efficient is such a crazy thing to believe because I feel like for most people, the cost of that human is non-negotiable anyway. What? Are you going to start shooting babies because it's more efficient to run an AI system?\n\nRegardless of the so-called efficiency or scalability, the AI won't actually help if it's pointed in the wrong direction. As it is, a small subset of humanity are deciding the direction. If the rest of us think we should be moving towards a more sustainable society, they don't care and in fact are directly making that infinitely harder. (I'm not saying majority do but it makes the point clear.)\n\nAs far as I can tell, current AI corporations are not pointed in the direction that would result in superintelligence. The AI systems they are building and scaling now will not necessarily lead to superintelligence; in fact I think conceptually the transformer architecture and LLMs are definitionally not even a step on the path to super intelligence. At most it's a small step, the translation of language into information for the theoretical superintelligence to parse. And if superintelligence does get produced, surely it would find a better more cost efficient way to do that step anyway?\n\nYet somehow we're committed to scaling and building data centers and pouring money and natural resources into maximizing this glorified autocomplete. Yes I know it's more complicated than autocomplete but a statistical token pattern matcher is essentially just a fancy autocomplete or flow chart. The interpolation and randomness does make it output things that weren't directly written in its training data but it's not going to output new novel thoughts or ways of thinking. It's also worth noting the biases introduced by those choosing the training data.\n\nAnother selling point of these systems is widespread accessible expertise. A global pandemic? No problem, here's 10000 epidemiologist bots. However, the so called expertise they want to rely on isn't even expertise, let alone reliable. Yes, humans also make mistakes, but when a human doctor misdiagnoses you there's accountability. If an AI model \"accidentally\" poisons hundreds of people due to a training error or hallucination, will these corporations be held accountable? I somehow doubt it.\n\nThey'll profit regardless of our outcomes. Step right up to Dr. Gachapon.\n\nIf we can't trust what the AI is saying then how can we use it to solve a crisis? Don't forget the fact that relying on AI will lead to having less experts to validate the output. Human expertise comes from working on something, making mistakes and learning. If we substitute people with AI (which they will definitely want to do), we may end up with a single world-class epidemiologist whose job was to validate all AI epidemiology output. What if they fall sick or die?\n\nI'd also like to point out that powerful and smart are not the same things. A chainsaw is powerful, but you wouldn't use one to change a lightbulb.\n\nThe way the big shops are currently making it appear like LLMs are thinking is by adding layers to the process. Your prompt does not simply hit an LLM and get turned into an output. It can go through multiple different systems, layers and layers of differently trained LLMs even. Depending on the domain and the model/service you are using it may even be integrating a deterministic executor to perform validation on the output. In this case, the model produces code that essentially models the problem and passes it to a code executor which checks that the code runs and/or that it outputs the right result.\n\nWhat if your query isn't within the domain of things that can be converted into a symbolic problem (basically math)? Unlucky. Non-verifiable? You're getting a random result. A problem the LLM was not able to map onto the existing patterns it knows? Ditto.\n\nI'll note that adding the deterministic executor adds an extra step and a non-deterministic loop to the process. If the LLM is unable to come up with a solution that works it will probably keep searching until it hits whatever hard-coded iteration limit it has. So, the costs of a single query become ambiguously large.\n\nYann LeCun is an AI figure who I find myself agreeing with to some extent. He has used the words \"If you are interested in human-level AI, don't work on LLMs.\" He believes that LLMs are limited in reaching AGI or superintelligence and researches in world-model architecture, his particular project is called JEPA. His approach involves training a model using real world information like video, which would hopefully provide a model with actual understanding of the real world rather than the illusion of understanding via language tricks.\n\nTo be honest, and this might just be because I'm a simple software engineer and lack domain knowledge, I'm not clear on how JEPA could possibly lead to superintelligence. I suppose it sounds like reinforcement learning patterns used in game AI systems occasionally. But it's just taking an LLM and replacing the L. The language one not the large one. I feel like any system that cannot have and apply fundamental beliefs or rules cannot be relied upon to perform in the real world.\n\nFor example, the idea that if you cut a short-grain A3 paper in half you get long-grain A4 is a simple idea that I have seen an LLM fail to grasp. It requires the ability to understand the physical world and apply logic to it, which at this moment I have very little concept on how we could digitize. We understand these things intuitively. This skill develops in living things since we were babies to varying degrees and are core to our approaching the physical world. This is the basis for our logical problem solving, don't even get me started on morals and empathy which are arguably more important and aren't even consistent between humans (e.g. vegans).\n\nI have said all of this, and yet I must either be in the minority or extremely misinformed. I see the cracks forming around us and can't tell if I'm just seeing things. I worry that the breaking point will be when either our society falls apart or the world gets too damaged for us to fix it. I would hope if the things I said are true, people would be rioting and objecting. Instead it seems like everything is more or less business as usual.", "url": "https://wpnews.pro/news/building-a-laser-with-a-chainsaw", "canonical_source": "https://thunkpoorly.bearblog.dev/building-a-laser-with-a-chainsaw/", "published_at": "2026-08-17 17:16:00+00:00", "updated_at": "2026-08-17 17:42:26.667899+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-ethics", "ai-policy"], "entities": [], "alternates": {"html": "https://wpnews.pro/news/building-a-laser-with-a-chainsaw", "markdown": "https://wpnews.pro/news/building-a-laser-with-a-chainsaw.md", "text": "https://wpnews.pro/news/building-a-laser-with-a-chainsaw.txt", "jsonld": "https://wpnews.pro/news/building-a-laser-with-a-chainsaw.jsonld"}}