2026-08-06 · The real energy use of agentic AI

Transcript

Transcript. Paraphrased; sources in notes.md.

HostAugust 6th, 2026. Welcome back. I am Daniel, joined as always by Tom, our resident maintainer. Today we are looking at the massive energy cost of agentic AI, the new LLM governance policy over at the Rust project, and the strange, unopinionated aesthetic of modern AI interfaces. Tom, I was looking through the queue this morning and I saw the news from the Rust team. You have been in the trenches of open source for a long time. Does this feel like a turning point?

GuestIt feels like a long time coming, Daniel. The thing is, when you are maintaining a project, you are basically a curator of intent. You are trying to understand why someone built a thing, not just what the code does today. When you start getting these massive, hallucinated, or mass-produced pull requests from automated agents, you are not reviewing code anymore. You are just debugging an opaque pipe. It is exhausting. I think a lot of projects are about to find out that the sheer volume of bot-generated noise is going to force them to become a lot less welcoming, just to survive.

HostThat actually leads perfectly into our first story. There is a piece up today analyzing the energy cost of agentic AI, or AI that performs autonomous tasks rather than just answering questions. The author tracks their own usage over eight weeks and claims that these agentic workflows are eating up about 600 times more energy per task than a standard chatbot prompt, mostly because the agent has to constantly re-process context. It seems we are hitting a wall where the marketing says AI is cheap, but the reality is becoming a massive infrastructure tax.

GuestIt is the classic tech cycle, isn't it? Promise a miracle, scale it to the moon, and ignore the cooling bills until the planet starts sweating. When you use an agent to write code, you are not just getting a function back. You are feeding it your whole repo, or a chunk of it, over and over, while it dithers about. If you are a maintainer, you are the one who eventually pays the price for that compute in terms of the technical debt left behind. If I have to spend two hours fixing a subtle bug in a function an agent hallucinated, that is a massive human cost on top of the kilowatt-hours. It is a hollow efficiency.

HostLet us shift over to that Rust policy I mentioned. The project is formally discouraging the use of LLMs for generating pull requests or issues. They are essentially saying that code contribution is a social act. Over on Lobsters, which is where we pull these threads from, the community seems relieved. One user, JulianSildenLanglo, makes a sharp distinction. They say it is fine to use these tools to distill or refine, but not to create.

GuestI agree with that. The act of creating is where the understanding is forged. If I write a patch, I understand the failure states, the trade-offs, and the edge cases. If I just ask a model to vomit out a solution, I have outsourced my own competence. There was another great take in that same Lobsters thread from a user named sanxiyn, who pointed out that if we are moving toward a world of AI-translated or AI-generated contributions, we might as well just let people write in their native languages. Why demand a machine-translated, sanitized English PR that might be hallucinated, when you could just have the person who actually understands the problem write it in their own words?

HostIt is a strong point. It challenges the assumption that we need this standardized, bot-verified output for a project to be healthy. It brings us to the third topic, which is the visual language of these tools. There is a blog post floating around about the AI aesthetic, the streaming text, the sparkles, the thin, non-opinionated iconography. The author notes that these are becoming universal design idioms. But in the Lobsters thread, a user named jgb hits back pretty hard. They describe this as the aesthetic of un-opinionated good taste, comparing it to someone who is tone-deaf but keeps their singing on-pitch by staring at a tuner.

GuestThat is the most accurate description of modern UI I have read all year. You look at these interfaces and they are trying to be perfectly neutral, perfectly safe, perfectly unimpeachable. They lack the texture of a human designer saying, this is how it should be because I believe in this way of working. When I build a tool, I want it to be opinionated. I want the UI to tell you how to succeed because I have actually thought about your workflow. AI interfaces feel like they were designed by a committee that was terrified of having a personality. It is just clean, white, empty space, everywhere.

HostIt is interesting because that lack of opinion is exactly what makes them feel so disposable. If everything looks the same, nothing matters.

GuestExactly. You see it in the code too. A pull request that is just clean, generic, AI-generated nonsense. It has no style, no soul, and usually no clue about the real constraints of the project. It is like they are trying to fit the whole world into this standardized, bland box, and they are burning an incredible amount of electricity to do it.

HostWe are out of time for today, but I have to ask, Tom. Are you going to spend your weekend fighting AI-generated PRs or just hiding from the internet?

GuestProbably a bit of both. I have a lingering issue in one of my repositories that I have been meaning to patch, and I am certainly not using a model to do it. I think I will stick to my own terminal and my own opinions.

HostSounds like a plan. Thank you, Tom, for the perspective. And thank you to everyone over at Lobsters for providing the context and the debate. We will be back tomorrow. See you then.