Discussion
I've seen this pointed out in an internal discussion and it's absolutely fascinating: it seems that the output of Anthropic's LLMs is becoming increasingly verbose and convoluted the larger the model gets. This has gotten to the point that it's becoming grating to users.
We've seen a huge amount of ways in which LLMs can fail, but "the computer sounds like a obnoxious toxic co-worker" is a brand new one.
@gabrielesvelto when I still had mandated AI use at a job (and during a recent experiment too), I found that I'm angry at the LLM and telling it in all caps to finally start to do some work, because all it does is "think" and spiral, and then generate some more fake thoughts. Before you know it 15 minutes have elapsed and nothing has happened.
@device and several kWh have been burned, most likely
@gabrielesvelto did you notice opus5 is at the top of all the graph but last on the last one?
@gkrnours yes! It's using more words but shorter ones which means that the output is more generic on average. This could also be a side-effect of Anthropic watermarking their outputs. You can defintely hide more data into a longer and more repetitive output.
@gabrielesvelto
theses are really good pointsc
I'm not a fan of how the information in these graph is presented. Like the bar chart repeat the ordering by the position of the bars, the length and the color. It's kinda obscuring that labels order is not stable. A bit naively, I think matching each label with a color or using a bold font for opus5 might help readability.
On first quick read, I thought latest models are using longer phrase with more complex words, sounding pedantic
@gabrielesvelto is there a comparison for these values to say "average text" and academic papers? Would be good too have a baseline to compare to
@gabrielesvelto sounds like a ploy to sell more tokens
Reminder that the fossil fuel industry funds AI.
Each iteration of a design or report or refinement in a query burns more fossil fuel & inflates wasteful energy usage.
The inefficiency & inaccuracy is deliberate. So is the verbosity.
Anything to prolong an energy intensive interaction with a data center so it burns more fuel.
https://www.wired.com/story/us-data-centers-drive-gas-power-expansion/
https://www.nytimes.com/2026/08/08/climate/amazon-data-center-texas-pollution.html
https://futurism.com/artificial-intelligence/pollution-ai-data-centers-severe
https://fortune.com/2026/08/18/data-center-gas-plants-to-boost-u-s-power-emissions-by-20/
https://www.nytimes.com/2026/08/25/climate/epa-data-centers-public-comment.html
@gabrielesvelto might this be to get through more tokens quicker, so people have to pay more? Shrinkflation without the shrink... Flateflation? >.<
@leberschnitzel @gabrielesvelto
AI is designed to be verbose to use more fossil fuel.
https://billmckibben.substack.com/p/hey-grok-whats-a-waste-of-energy
https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/
The Jevons Paradox
https://www.npr.org/sections/planet-money/2025/02/04/g-s1-46018/ai-deepseek-economics-jevons-paradox
@Npars01 @leberschnitzel yes, absolutely. These companies are in a flagrant conflict of interest: larger models cost more, users are billed by input and output tokens and the conversational nature of chatbots turns outputs into inputs. They have all the incentive to make the output as verbose as possible to milk their users. I wouldn't be surprised if that's why this is happening.
I also find fascinating the way it's framed by users in that GitHub issue. "Its anwsers are hard to parse", "it's argumentative", "it's hard to understand its answers", etc...
People, this is a machine. It does not understand nor does it convey meaning. It does not know things and it cannot answer questions. Your brain is trying to put meaning where there is none. You're reading the statistical output of a very large matrix multiplication and assuming it makes sense. It does not. By design.
And because of #enshittification , the output will keep getting worse and more expensive.
@gabrielesvelto Not making sense but pretending it does is certainly a trait of an obnoxious toxic co-worker, so the fact that the LLM sounds like one seems inevitable. 😅
Install bonfire.cafe
Get the full app experience
