I think the idea behind energy shaming is that energy price rarely includes its externalities (probably resources depletion and pollution). "If You buy it, You do what You want with it" is of course legally correct. But I think the argument they're trying to make subconsciously in is that there's someone else who will have to pay part of the price for it.

Water use is not an externality. All water used is paid for.

[citation needed] Nuclear Power plants who use river water in their cooling cycle and pump the (now heated) water back to the river don't pay for that.

That's not the kind of water use the article is complaining about.

The cooling system illustrated in the article is very very similar to a typical nuclear power station cooling system. From the heat exchanger onwards, the secondary loop is functionally identical, the main difference is the primary loop working fluid (and that loop operates at a far lower temperature and pressure).

All water use includes that one right?

The argument is I think that all unnecessary energy expense should be shamed equally, or not at all.

Can be true, but it applies to Disneyland as well.

But if you're a conservative cynic about Disneyland, people will just label you as some old sad geezer with outdated opinions yelling at kids playing in the garden. You have to hate on something new and trendy to keep up your consumer focus group engagement metrics these days.

Which makes you seem insincere.

Energy shaming isn’t a conservative schtik. The OP is right, why are we picking and choosing? Shame them all. Disney, Six Flags, Universal Studios, Sea World, Bush Gardens

Then where are all the articles bashing them? Only targeting the trendy bait makes them seem insincere, just like the other commenter said.

I would presume, because it is the largest, and most famous of them all… Not for any nefarious reason

It is by far neither the the most famous nor the largest once you account for the whole population and not some bubble.

That's a great argument! Assuming every word generated by ChatGPT is a word which would otherwise have been typed by a human writer.

Is every word generated by ChatGPT a word which would otherwise have been typed by a human writer?

Why would that matter? Every system has inefficiencies. For example: Person sitting in office, trained for years, kept alive and breathing, being paid for by their employer — that is not writing that mail their boss is waiting for, but instead arguing about the resources used by ChatGPT.

Why wouldn't it matter? If ChatGPT produces words at roughly the same energy efficiency as human writers, but most of those words are unnecessary and would not have been produced if it wasn't for ChatGPT, it's worthwhile to discuss the energy cost.

ChatGPT is not housing anyone, nor is it providing office space. Even Disneyland is at least providing joy for children and adults.

I agree - it has no direct use to the wellbeing of humans - Id say if the indirect use (e.g. humans saving time [which itself is probably the most energy intensive thing, food, housing, entertainment], gaining knowledge has to surpass the costs in Order for the investment to make sense.

No one is 'gaining knowledge'. LLMs are just black-box tools that compute pseudo answers based upon a somewhat arbitrary set of training data. It's a crapshoot that looks good like ELIZA looks good.

Me: Please tell me an interesting fact I don't know.

GPT: Did you know that octopuses have three hearts and their blood is blue? Two of their hearts pump blood to the gills, while the third pumps it to the rest of the body. Interestingly, when an octopus swims, the heart that supplies blood to the body actually stops beating, which is one reason they prefer crawling to swimming—it’s less stressful on their system! Their blue blood is due to hemocyanin, a copper-based molecule that is more efficient than hemoglobin in cold, low-oxygen environments.

Me, after researching and corroborating each claim, which you should always do for any source: Wow, I didn't know some of that! Thanks for sharing!

---

This was no different than googling "interesting facts about octopuses" and skimming the first few links. And if I'd done that, you wouldn't be claiming that I didn't "gain knowledge". But by all means, commence the mental gymnastics that "prove" I haven't gained knowledge. Perhaps because I had to corroborate the information? No, you should always do that for anything. But I'm eager to hear your explanation for what really happened.

That's a lot of energy used to save you going through the octopus section of Wikipedia.

If only there was a test to find out if the person using it was a moron. Then it might be useful. Of course, Dunning-Kruger dictates that no one using such a system would ask such an important question. Also, the world is obviously well short of the necessary training data.

I've got yer real intelligence right here, dude. Ask away. I even know how to tell you that your question is a waste of time.

Amazing, you chose to move goalposts. No longer are you bothering to defend your original claim that no knowledge was gained. But now, your claim is "you could have just used another source"! That's great, thank you.

I can't tell if your deliberately pretending like I didn't give you a toy example or if you truly think I sit around querying animal facts all day.

> I even know how to tell you that your question is a waste of time.

The only waste of time has been attempting to engage with you intellectually, when all you're really interested in doing is "proving" that you're right. But if you're just going to keep moving goalposts we can just end this discussion here.

My entire life trajectory has changed because of how LLMs have expanded my ability to write, think and reason. Have fun at Disneyland, though.

If you've experienced this life altering change and it's reproducible, you could show up a lot of AI critics with specifics. How have LLMs expanded your ability to reason, for example?

At the risk of sounding negative... The thing is I have tried interfacing with skeptics, and most of them are just run-of-the-mill closeminded, have their minds made up before you even begin and will just argue in circles until it simply isn't worth it.

Meanwhile, a decent portion of my own peers echo my sentiment, mainly programmers. I don't have to argue with them. Instead I can analyze and improve methodologies with them, collaborate, and generally have positive, stimulating and thought-provoking conversations. You can imagine which of the two kinds of people I typically choose to engage with. Especially on a forum like HN, where meandering circular arguments are discouraged. Maybe the happy medium is just blogging about it. I really do want to blog and I'm trying to learn about it and make more time to write, but HN does already scratch a lot of that itch for me.

All in all... the only self-proclaimed AI critics left at this point are people who have either chosen to be deliberately ignorant about the benefits of this technology, or who lack the depth of skill or interest to find useful things for these models to do for them.

I've taken on increasingly ambitious projects because I now have little programmable assistants embedded into the UX of various tools that I work with. They're not perfect, sometimes we argue, sometimes I just have to turn them off for a few hours, roll up my sleeves and work through the muck. But at a high level, the benefits are huge. Modern adequately-trained LLMs are amazing at project scaffolding, architecture, design, documentation and generally being a backboard for my thoughts when I have no one else around to converse with. My typical day is highly varied and while I have specialist friends, none of them match the generalist abilities of some LLMs.

I am generally skilled enough with informatics to know when I need to corroborate or discard information (basic research skills learned independently of LLMS) and so "what if it's wrong / how will you know?" has always been an irrelevant, false paradox parroted by people lacking those basic skills which should have been taught in school. These same people already struggle with finding accurate information with search engines or libraries, it's no wonder they struggle with something like an LLM which can be very confidently wrong. With programming, it's easier. You can keep the model focused, break down tasks into tiny pieces, solve and write tests for them one by one. It encourages you to write clean, modular, self-documenting code.

With other topics like science, art and mathematics, I'm still consistently impressed by 4o and o1's capabilities, despite glaring shortcomings. I know you want very specific instances but I'm really ADHD and so I am not exaggerating when I say my conversations with LLMs are highly multidisciplinary and before a just a couple years ago, I had a hard time organizing some of these more esoteric and complex systems in my head. My typical conversations would probably be of little use to someone.

I'm now working in parallel on several scientific and mathematical inquiries, as well as a few ambitious engineering projects. My time spent researching has greatly reduced, as I can get up to speed very quickly by dumping some articles into an LLM conversation and asking it detailed questions, asking it to provide thought models, etc., getting maximal value out of the information by immediately dialing into areas of interest.

Essentially... I've always felt naked without a phone or computer around to google the random questions and ideas I have throughout the day, and I consider the internet and search engines integral to who I am and the knowledge and skill I've attained. Now, instead of a search engine, I feel naked if I don't have a well-trained chat model around. I have developed a similar dependence as I have to search engines (I panic when I have a question, slap my leg and realize my phone isn't on my person) but the tradeoff is that I feel my thinking and doing has been augmented.

And even if transformers/LLMs hit a dead end in our lifetime... after all, many of these techniques are rooted in ideas from the 60s-90s which simply "didn't work" at the time, only to become relevant again after sufficient compute is accessible... I am eager to see what the next 40-odd years of technology brings us.

Regarding reproducibility, I think it would be extremely beneficial for people to be shown how AI can help them with day-to-day things, but I have a hard time recommending LLMs to people who I know aren't equipped with enough research skill to avoid harming or hindering themselves from incorrect advice. And my experience with these models is mainly academic and related to engineering, so I just don't have much to offer normies, but I'm sure other people have made lots of ground on that.

Personally I think chat models in their current form just aren't natural or functional enough to appeal to the average person. I think the average person will see the biggest benefits from tooling built around modern models. I'm working on one right now, a grid-based component system which essentially lets you create tiny little tools that can speak with each other, and orchestrate them into one or more domain-specific UIs. Users can share and adapt apps or individual components, or create production-ready in-house creative or business tooling. Social productivity programs like this, as well as calendars, sheets, etc (all buildable within my app in minutes with a handful of stock components, with built-in multi-user support) will unlock the full power of advanced natural language models. Hoping to debut it sometime next year.

Disneyland also involves a lot of plane flights and driving to get to it. If you add all that up it's a lot more.

Yeah, I'm personally not a huge fan, but it's not nearly as bad as what any gambling destination is. That is nearly pure waste, with the possibility that people are seriously damaging themselves with addiction and alcohol.

There is a force inside us that seeks to convince us to undermine our own selves.

Sure, but thinking of all data centres as gambling is just going to lead to poor conclusions, no?

ChatGPT saved 3 hours of my time today.

You are far more energically efficient than a data center.

Wetware >>> software

I really am not. I just drank 1 liter of water. ChatGPT consumed several orders of magnitude less doing much more than me. I also just heated up my meal, with meat and fries. ChatGPT didn't need that.

You're not counting the training of the LLM.

I am. That's amortized over millions of users and billions of invocations.

But who cares? I'd gladly pay for something to do in minutes what would've taken me hours. It doesn't need to be as efficient as my body in the same time frame.

chat gpt doesn't enjoy anything for itself, it provides value to people. i don't use it much, but on occasion it's been helpful to me.

So do coal-fired power plants. So does our endless production of plastic waste and our ever-growing landfills. So does all our fossil-fuel use.

Where is it all taking us, friend? Most people are not honest enough with themself to understand the bigger picture, nor are they selfless enough to give a darn to sacrifice something of their own benefit to help the whole.

Yes, let's all kill ourselves and save resources!

Yeah, having less data centers will cost lives; it won't be lack of drinking water or living in an overheated environment!

My mistake is mistaking high karma for intelligence. Wrong again!

This anti-compute crusade you're on is very strange. Data centers are not why the planet is burning up. Maybe you should be less patronizing of others' intelligence and do a little research of your own so that your arguments will be more rooted in fact and not feelings.

~1-3% of global energy consumption is a paltry price to pay for what digital technology can offer us. Do we waste a lot of it clicking ads and watching Netflix? Yeah, a lot of people do, and I take issue with that, not... data centers.

...You know there's some options between "capitalism optimising for profit margins leads to unchecked consumption of resources" and "everyone has to die", right?

Yes, in the way that there are differences in pollution between using coal-fired plants and using LLMs.

do you live in the forest in peace with nature?

The question is how much value. Sure, learning about entropy & unblocking students is huge value - producing clickbait content does way less so imo

You don't have to click on clickbait. If you voted with your clicks, it wouldn't make them a profit.

You know what’s also not housing someone? Office space. This feels like the ultimate waste - massive empty buildings in prime central locations with plentiful public transit designed to house humans during the day in addition to their own houses - which are being pushed further and further away from the central locations because of the office space - that they’re forced to commute long distances against their will to occupy simply because management wants them to. I’d rather have a data center.

I’ve used ChatGPT to teach myself several programming languages, guided me through learning how to design circuits for my own projects, among many other uses. Professionally I use LLMs to scale processes for detecting fraud, financial crimes such as money laundering and sanctions evasion, amongst other things at a much greater precision and recall than other models and human investigators at a much larger scale (we still use those other vectors as well but the major break through for quality and scale was LLM hierarchical agents). We’ve done the analysis and it’s demonstrably improved our business and made the financial system in our corner safer and more stable. So, I dunno my friend. I think you’ve perhaps got a bias based on your inexperience ?

I think the question is how efficient is ChatGPT in providing economic value. I don't want to get in a game of anecdotes because I don't want to deny that there are use cases well-suited for LLMs and you might be in such a field. The efficiency question is thus: of all the queries LLMs serve, how many of those are hallucinations (which then need to be reprompted), how many of those are simple queries that could've been solved with more traditional search methods, and how many of those are queries which only LLMs could've solved?

I don't really have an answer to my own efficiency question and will likely never will. But for every company like yours, there are companies who blindly introduce LLMs into their product for no other reason than hype. Spotify, your recommendations ("DJ") are no better because you used LLMs; collaborative filtering worked just fine for music. Meta/FB, your value prop to me as a social network hits rock bottom if my friends are not even sharing real-life experiences. Google, you're already on thin ice with all the SEO spam but your value decreases even more if you give me hallucinations for summaries. Hence, sorry but my outlook isn't rosy.

The thing is, at a certain point, economics is a bitch. Either efficiency improves or those uses stop. I’d note that inference is much cheaper than training and has become radically more efficient rapidly. There’s good reason to believe the current wild eyed training boom we see now will collapse on itself in terms of cost to benefit, or will curtail to the economic value.

However you need to understand that what you said is wrong. Many of these things can’t be solved with a simple query because a query returns a blob of data and the human has to search for the information in a set of documents possibly containing it.

The UX of “I have a question” and the answer of “ok here’s a likely direct answer that’s comprehensive and comprehensible to which you can ask any follow up” in unbeatable. It’s what AskJeaves.com was as one of the first search engines but couldn’t achieve because NLP was so shitty. The fact it might hallucinate - which is very unlikely in a frontier model on a pedestrian question about common knowledge that can likely be answered by a search query since it’s training set likely contains a lot of examples of the answer and it can supplement with a basic RAG against a traditional search engine - is not even that relevant because unless you’ve not used modern search engines recently it’s almost impossible to get a straight answer from search due to SEO.

So the reality to your rhetorical question is “every single one of those questions only an LLM could have solved.” Because they’re the first real NLP system we’ve invented that can give coherent answers to the question.

Very few search engine queries are seeking documents. Most are seeking an answer to a question. A search engine has been the best we could do until a few years ago to answer questions and the user experience is fairly shit and rapidly deteriorating due to economics. It’s like asking a professor a question and being answered with a bibliography but over time realizing the professor is being paid under the table to stuff the bibliography with advertisements and the rest of the bibliography was replaced by other ads pretending to be source material and the professor just looked at the title to select them. Not only is the answer not an answer - it’s a list of stuff to read which might contain the answer if you read it all - it’s increasingly unlikely the answer is even there at all. You can argue “but the professor answering with a bibliography ensures you learn a lot!” Which is fair but people usually just want a direct answer and would sooner go to the cool professor down the hall who answers your question and gives the bibliography as citations (as modern frontier LLMs like ChatGPT and Claude do).

The fact the cook professor dabbles in mushrooms too much during office hours is unfortunate and the school administration is working on that. But it’s useful enough people literally don’t care.

For use cases where the cost >> value like collaborative filtering - well, fine. LLMs won’t be used there because the value doesn’t hold. We are only two or three years into this and there’s a lot of just dumb stuff because no one knows what will stick. And LLMs will not be the answer to everything. But I think they’re going to be increasingly more powerful simply because we will defer the right algorithm to the problem and the LLMs will be the glue of language and abductive reasoning that makes something useful, such as information retrieval, usable. And that alone is more valuable than Google’s search engine. And that is not the only use - as I’ve seen directly. So yes LLMs won’t be grand masters at chess beating alpha blue. They don’t have to be. They can be the user interface to Alpha blue and alpha blue becomes more usable immediately and much more valuable.

The hardest problem in CS isn’t cache coherency or naming things or off by one errors it’s making useful things usable, and among all the other useful things LLMs have done, they have absolutely solved the hardest problem of all - making all the useful things usable.

But this article wants to talk about unnecessary office cooling without talking about unnecessary offices. If this is such a big deal, then why aren’t office parks springing up in the arctic?

Agreed. If it's useful, why not scale up the electricity and water supply, and make the latter sustainable.

Sometimes the reactionary rules make sense to try scaling down: electronics on standby can’t use more than 1W, by law in EU and UK. Consumers were right to wonder: what was my TV doing that needed more than 1W?

If "energy-shaming" is a bad practice, why start by mentioning energy consumed by the author of the article?

I think we can leave Disneyland aside and talk about the content of the article, and whether AI companies should receive the gentle treatment they have received: premium access to water, power, and money. Sure, they can buy it, but what about the people who have to absorb the externalities?

So how do you suggest we act in a world where energy usage ultimately has to be limited? Just let the market regulate that?

Then you end up with luxury villas with swimming pools in the desert while farm land dries out and people die of thirst. If you grow a system against stiff bounds you will get hard saturation effects. And that means you can't have it all in such an environment.

Now using that drought-example I think it is reasonable to say your pools have to stay empty, priority 1 is drinking water and priority 2 is farm land. And if you are the type that would like to fill a pool in that hypothetical situation you probably deserve to be shamed.

Claude estimates the average American consumes about 3800 liters of water per day counting all the food inputs (Claude only drank half a cup for this answer).

You did not count the gathering of data and then training the LLM.

Inference consumes far more total compute than training the model.

By 'inference', do you mean using the model to compute a specific series of results?

Yes, training is a tiny amount of usage, this is well-known.

Well, if that's the case, then the situation is worse than I imagined. I thought that training the model would be far more expensive computationally than using it to compute an 'answer'.

Thanks.

Training is much more computationally intensive than computing a _single_ answer. But not compared to running the model over time at scale for millions of users.

Thanks. That is precisely how I intuited the situation. 40+ years of programming FTW!