COllaborative DEcision MAking (codema.in)

publicai.co and general position towards LLMs

Pirate Praveen Thu 26 Feb 2026 11:05PMPublicSeen by 70

Moving our position towards LLMs to its own topic. I initially proposed to support publicai.co as a reasonably good LLM we can recommend. But for now it is removed from the manifesto. So we can discuss it in more nuance here and come up with a position statement on LLMs.

About copyright violation aspect of LLMs in generating source code, this gives a good background https://lists.debian.org/debian-project/2024/05/msg00003.html

LLMs do enable people who do not have technology or specific language skills (especially English) to be able to do things they would not be able to do without an LLM. At the same time, this uses a large amount of computing power (and electricity, water and other resources) for training. But LLM are not really the only source of huge computing power usage, how about video streaming platforms? Is it ok for entertainment industry to use so much resources, but not for enabling people to do more? Where do we draw the line?

For me I'd like to focus on publicai.co as a starting point? Do we think it is solves some of the issues with respect to ownership, bias and copyrights? We may not be able to get to an ideal solution in one go, but I think we can at least have something that we can recommend in place of proprietary things like ChatGPT, Gemini or Grok.

Draft:

We support publicai.co as reasonably good compromise at this point. This is a compromise and we will continue to evaluate the options available and update this recommendation if required.

Deep PandyaFri 2 Oct 2026 5:24PM

@Pirate Praveen How about https://allenai.org/more-than-open ?

Pirate PraveenSun 4 Oct 2026 6:48PM

@Deep Pandya hearing about it for the first time, looks interesting.

Pirate PraveenSun 4 Oct 2026 6:56PM

https://allenai.org/about people who fund it makes me doubt its intentions. It is in contrast to https://publicai.co Allen AI seems to me more about charity and marketing than creating truly public goods like Public AI. But still the models seem good and their goals seems to be good, just their funding and political philosophy is a bit weak.

Pirate PraveenSun 19 Apr 2026 8:53AM

AkshaySun 19 Apr 2026 11:40AM

I've used a model called Big Pickle offered as part of Opencode Zen (for free now) to develop https://opendatakerala.org/KLA2026/

It enabled me to iterate faster and build the platform much faster than what I would have if I tried to do it by hand. In that process, the team that was scraping data, etc also got motivated to do more (because they were seeing quick turnaround for their feature requests, etc) and overall the platform became possible only because of the model assistance.

I tried to also use Antigravity, but the free edition was quickly hitting limits.

This experience has taught me a few things:

  1. The development speed / ease related improvements are real and the second order effects of these are also positive. (Third order effects? I don't know)

  2. It does become harder to switch back to manually writing code (mind gets used to delegating)

As for copyright, I'm in the group of "people who believe all software copyright is an ill-advised legal construction that limits people's freedom". (I've told that LLMs will probably lead to the decimation of copyright and that that is good as far back as June 2023).

I think overall, the debate that LLMs make us do is forcing us to find out what exactly free software was meant to achieve.

If free software was only about software freedom, LLMs are absolutely great!

If free software was about user freedom, then LLMs get complicated by questions of access/affordability, environmental impact, effect on economy and so on.

As for access/affordability I think publicai might be an answer.

As for environment, I wouldn't shout at anyone if they hoped that within maybe a few years there would be LLMs that are capable of running without much damage to the environment. I don't know if that's a form of wishful thinking, or whether it is a possibility. In any case, the comparison with other tools and products that cause damage to the environment is a fair comparison. I believe in degrowth (and the philosophy I recently came across -permacomputing - as shared by Badri). But I also believe that this particular dilemma is very tightly interconnected with our core politics/psychology/agency and the such. Are we better off if we go off to a cave and meditate? Or is there a quota of environmental damage each human is allowed to do? If so what's that quota? Is there a way to fairly distribute that quota? And if so, is there a possibility for each individual to choose what to do with their quota?

As for effect on job/skills/capitalism, I have a muddled understanding. I do hope that like copyright becomes obsolete, all bullshit jobs related to computers also become obsolete with AI. But is it possible? I don't know. When anyone can build any software, why would we still build large software companies? Does it make sense then to pay another company to give you what can be generated by you yourself? Does all the software SaaS/library kind of opportunities then get concentrated in the companies like OpenAI then? I don't know. It's complicated to think about. I don't really worry about skill. People who want to learn/acquire skill can continue to do so as always.

Pirate PraveenWed 10 Jun 2026 7:30PM

LLM Extractivism and the Politics of the Knowledge Commons

https://botpopuli.net/llm-extractivism-and-the-politics-of-the-knowledge-commons/

Aaruni KaushikThu 1 Oct 2026 10:33AM

I was redirected here from the FSCI MUC, and I here's some thoughts on public AI. I should preface by saying I am not a subject matter expert on LLMs, and not active in the research space. So, it is possible I get some details wrong, but the overall picture should still be close enough.

  1. thaura.ai

    Thaura is built on top of Qwen3.8, which is a chinese open weights model. The fact that it came out of china is not really important here. But being only open weight, and not having heard other positive things about its training process, I think it is safe to say this is not an ethical model at all. It is also safe to assume that they will want to upgrade to Qwen-latest-and-greatest whenever that is out, so we cannot argue that the training is one time sunk cost in terms of energy. As long as they keep updating their models, they keep creating demand for more unethical training in terms of both data and energy.

    They do try to make themselves as energy efficient as possible. But using the numbers from their own FAQ, the energy usage is still substantial, at around 0.5 kwh for every million tokens (million tokens sounds like a huge number, but a small code PR I am currently looking at as an example is already half a million tokens in its current state. This is after multiple rounds of revision, and not counting the tokens spent in "thinking". So, to merge a single PR, it takes somewhere around 5 to 6 million tokens, or 3 kwh of energy only in terms of LLM token usage. There is the additional overhead of wasting computer time on the client side, and wasting human resources time in terms of iterating over nonsense and wasting reviewer time. While this figure may be 10 times lower than other LLM providers, I would like to that this is no small amount of energy, and should not be championed as "the way forward" because it looks efficient in comparison to obviously worse options.

    They try to make themselves even more redeemable by saying "waste heat from LLM churnage is used to heat homes!". However, this much energy could be used in much more efficient form (heatpumps) to heat even more homes. (Reminder that heatpumps have an "efficiency" of over 100%, because they don't burn energy to produce heat, but instead use energy to move heat around. So they can produce more than 100 watts of effective heating when using 100 watts of power.)

    Even if you are willing to overlook the problems with unethical training and energy inefficiencies (but seriously how on earth will you manage to overlook such humongous problems!?), we run in the very next wall of deskilling and cognitive decline. It is clear that protracted LLM usage does lead to sharp cognitive decline. There is no "responsible usage pattern" around this. Human biology is hard coded to rely on external tools whenever possible in hopes of personal efficiency (which, because evolution isn't too sophisticated, just means least amount of energy used per task done "good enough" to not get us killed for now). And human brains also very much work on a "use it or lose it" model. The more things we offload to external tools, the less energy we use on it from a biological point of view, the happier our brains are. But those skills face very real atrophy very quickly. Also, evolution does not understand that offloading cognition from a 20 watt device (the brain) to a 1.5 kw device (estimated LLM usage) it not really efficient in the grand scheme of things.
    Then there is the problem of deskilling: as LLMs become a force for "productivity multiplication" for senior employees, organizations try to get rid of junior employees for short term gains. But the senior employee will at some point retire, and the junior employee -> senior employee pipeline has already been disrupted. So we will no longer have any competent employees!

    But why do we need competent employees when LLMs can do all the work? Well, by now it should be clear, LLMs need a very strong captain at the help to not produce slop. LLMs excel at producing "plausible" sounding answers. It then needs an expert to sort out the nonsense from the useful. But if LLM use kills off experts, how on earth is anyone supposed to "check" its work ?

    (BTW, this ~600 word piece was written by completely human intelligence, and according to some estimates, provided only 1 gram CO2 emission, v/s ~3 gram of CO2 emissions using similar estimates for LLM inference)

  2. euria.infomaniak.com

    This only appears to be a "run an AI model" service. Perhaps they also try to be efficient by using waste energy to heat homes or using only renewable energy. But as already explained above, I don't find it being "less bad" as a good reason to start recommending it.
    It allows you to choose which model you run, and all but the Apertus models are patently unethical. The Apertus model is discussed below.

  3. publicai.co

    I will use "Public AI" interchangeably with "Apertus", the LLM model they have generated / use to provide inference. As far as I can tell (from reading the paper they published about it, available at https://arxiv.org/pdf/2509.14233 ) it really is a truly open-source model (I have not read the paper with an aim of verifying every single claim and trying to poke holes. I am taking them at face value when they say some data set is publicly available and cite it.). This is impressive. However, on the energy front, things begin to look not as impressive. According to the paper, a "full pre-training run" is estimated to take 5GWh of power over 90 days. (5 billion units of electricity is no small amount. This could have completely covered the energy requirements of a thousand luxurious detached homes in Switzerland for the entire year). Note that this value is a lower bound on the actual power it took to produce this model (Model requires R&D, crawling, pre-training, failed attempts, post-training, benchmarking, post-training, and, and and).

    Coming to benchmark performance, it is impressive how close this model comes to other open-source and open-weight models. I will summarize some results from the benchmark section of the paper, pitching the best performance of Apertus v/s the best performance of everyone else, without really detailing who the competing model is in each case:

    General Language Understanding: 67% v/s 69%

    Factual Knowledge Acquisition: 58% v/s 72%

    Knowledge recall / Common sense reasoning: 63% v/s 68%

    Coding and math reasoning: 54% v/s 74%

    Reasoning and instruction following: 61% v/s 80%

    Cultural knowledge: 61% v/s 59%

    Test benchmarks: 51% v/s 65%

    RULER benchmark 64k: 56% v/s 89%

    As I said, it is impressive how close ethical training data can bring a model to other open-weight models. But it is also clear to see that in literally every case, this is the worse choice in terms of benchmark performance. Also they did not compare their model against the benchmark results of fully proprietary "frontier" models.

    The question then, is this, is 5GWh of power for pretraining, plus more power for other steps and overheads, plus more power for inference really worth it for benchmark results of ~58%, barely better than a coin toss?

AkshayThu 1 Oct 2026 11:35AM

@Aaruni Kaushik  I have not previously delved (I wrote delve, not an AI) into this mathematics, so my math may be wrong.

Also Thaura seems to have probably updated the number as you were typing the comment, because what I see in its FAQ now is:

Published industry and academic figures put a typical reply from a large AI chatbot at roughly 0.3 to 1 watt-hours, and several times more for long or reasoning-mode replies. We measured our own servers across the full production week to 1 October 2026: about 0.013 Wh of GPU energy per 1,000 tokens, roughly 0.03 Wh once server overhead and data-center cooling are included. Per token, that's more than 10x less energy than even the lowest published frontier-model estimates.

For million tokens then, thaura uses 0.03 kWh. 5 to 6 million tokens would be 0.15 to 0.18 kWh.

Just for comparison, my laptop maybe at least 60 W. If I take 1 hour for writing code or reviewing PR, it would be 60 WH, or 0.06 kWh.

If my math and assumptions on time, etc here hold, then AI reviewing is only 3 times more energy consuming.

Of course I've not factored in training costs into this. (There's a training cost for training a human being in coding too :) )

Can you check my math?

If that math is right, it might also be important to differentiate between multimillion token activities like coding and thousands of token activities like chat/query.

My opencode stats result says 334.5m tokens for the [Kerala election website](https://opendatakerala.org/KLA2026/). This translates to 10 kWh as per my math above. (Of course I didn't use thaura for this project, but we're in hypothetical sphere now). [This random website](https://henryman.net/2024/05/12/how-much-fuel-do-electric-cars-use/) says 1 litre of petrol is about 8.9kWh. Or why go there, I have an electric car that spends about 40 kWh for about 300km. And I drove about 150km today in that. That's 20 kWh right there.

I'm not saying 10 kWh of building a web dashboard is always a good idea or that it is ethical. I'm just asking - is my math right?

Also, can you convert the 5GWh to number of transatlantic flights? I can only see per passenger numbers mentioned everywhere. If one transatlantic flight takes about 500MWh, then, 10 transatlantic flights = 5GWh. And there are maybe 30 flights between just London and New York every day.

I'm not defending AI here, I'm just doing the math to compare AI training/querying costs compared to car, plane, etc so that I can better understand what actually is the energy cost (especially in context of boycotting one vs not boycotting another).

https://ntrs.nasa.gov/api/citations/20230006917/downloads/EATS%202023%20Presentation%20Slides%20(2023-05-03).pdf mentions 97,501GWh on page 10.

Aaruni KaushikThu 1 Oct 2026 12:38PM

@Akshay 

> Also Thaura seems to have probably updated the number as you were typing the comment

Indeed yes. When I looked up the numbers, the FAQ section had measurements from August. Looks like they have somehow managed to squeeze out just a bit more efficiency in their product line up since then.

Just for comparison, my laptop maybe at least 60 W. If I take 1 hour for writing code or reviewing PR, it would be 60 WH, or 0.06 kWh.

No, this is most likely not accurate. Your laptop maybe has a CPU with a listed TDP of 60 watts. This means that if you are using 100% cpu power across all cores and threads and coprocessors and iGPU, then you need to dissipate 60 watts of heat (which in turn should imply that the CPU is using 60 watts of energy, but the CPU can use a bit more during bursts, as long as it averages out around 60 watts). But when you are writing code for an hour, your CPU is not running full blast. Its probably sitting in some low power mode sipping 15 watts or so.

Think about it this way: your laptop probably has a 50wh battery. If you get 5 hours of battery life out of it, you are only consuming 10 watts of power at any given time (on average). My laptop (intel core i5-10210u) consumes around 8 watts of power with vscode open and youtube running in the background. The power consumption rises to around 10 watts when I am also typing in vscode. So, the act of typing something out only has an overhead of ~2 watts in terms of computer energy. So, if you say the computer is powered on but idle all the time anyway, and you only count the energy for typing, the right value is ~2 watts. (contrast this with if a server is powered on and idle, it also takes something under 100 watts of consumption. But if you use it for inference, it now takes the full advertised value of ~600 watts on the GPU (so, around a kw on the server as a whole, give or take). So, the power figure for inference is still around 0.75 KW, around 375 times more than typing.

And I drove about 150km today in that. That's 20 kWh right there.

Presumably, you had to do the 150km anyway, using whatever mode of travel or transport necessary to make it happen. And it is miraculous and mind blowing that it is possible to hurl a ~100kg load that distance in a single day using just 20KWh of energy. (OTOH, referring to the benchmarks above, the average gain we get for spending all of the LLM energy is 58% accuracy?)

If one transatlantic flight takes about 500MWh

I dont know about the energy costs of flying. Let us take this number to be true. Then we should still marvel at the fact that moving this quickly for legitimate use across the world is possible and this power figure is spread out across 100s of people on the flight. (Also, this is around the time the discussion around costs comes in again. The huge cost of you driving is paid by you, more or less unsubsidised. The huge cost of flying across the ocean is paid for, again, more or less unsubsidised, by the flight ticket. Who is paying the costs of AI usage?)

AkshayThu 1 Oct 2026 1:46PM

@Aaruni Kaushik thanks for checking my math. As far as the impact on environment is concerned which usage is legitimate/marvelous shall be judged by each person. I don't think transatlantic passenger flights have much legitimate needs.

If there's a discussion about ethics of LLM use and the argument against it is energy usahe, it necessarily have to be compared against other energy consuming human activities and proportionality has to be applied. It cannot be that transatlantic flights are ethical and LLMs are unethical. Maybe LLMs are X times as ethical/unethical (on the impact to environment dimension) as flights. The X has to be based on actuals. (Granted there are second order effects, excess consumption due to shift in habits, etc). But simply because flights have been burning planet forever and AI started burning it only recently, doesn't make one different from the other (in the environment dimension).

To be clear, I think both flights and LLMs (and any other energy wasting activities) have to be avoided as far as possible. But that also means I find situations where using them are acceptable. (And I believe each person will have to be judicious about their own total impact on environment). But consequently it also means I don't think a blanket boycott of LLMs alone is justified.

Aaruni KaushikThu 1 Oct 2026 2:01PM

I don't think transatlantic passenger flights have much legitimate needs.

Just to clarify, do you think people and places far enough away should just be sectioned off and no travel should be allowed? A transatlantic flight might take 500MWh. But if you look at efficiency, (from the same nasa presentation you have linked) a flight gives you 40 to 80 mpg/seat. That's between 17 and 30 kmpl of fuel efficiency. Does it then also follow that any means of transport that is less efficient should also be illegitimate ?

Just out of curiosity, I tried to find some number for cruise ships. Apparently, they use around 475 thousand liters of fuel every day, and travel at 29 knots (53 kmph). Lets say all 24 hours in the day are spent travelling at this speed, we get 1287 km per day. So, the average fuel efficiency is 0.0027 kmpl. (The real comparison value will be higher. This assume only one person is on the whole ship. But I don't know how many people travel on a typical cruise ship... Assuming there's 2 thousand passengers (no real reason for this number), you still only get efficiency of 5.4 kmpl)

If there's a discussion about ethics of LLM use and the argument against it is energy usage

The argument is cost v/s gain. Energy usage v/s what we get for it.

(As a reminder, the above arguments are only for the Apertus model, because all others are unethical to begin with, because of how they are trained.)


I accidentally posted this as a new reply. This was meant to be a reply to @Akshay

AkshayThu 1 Oct 2026 3:45PM

@Aaruni Kaushik 99% of humanity travels barely kilometres from where they are born. There's absolutely no survival benefit to traveling large distances.

The point I make anyhow is that while "cost" (environmental damage) can probably be calculated realistically, any measure of "gain" will be subjective here - and therefore any one person's subjectivity of what is gainful/legitimate cannot be applied to determine the "ethics" of damaging environment for the sake of producing software.

Even otherwise, there've been several examples given of how people are producing legitimate software with LLMs.

That does not mean that people should use LLMs indiscriminately. It only calls for nuance.

Aaruni KaushikThu 1 Oct 2026 3:53PM

@Akshay 

99% of humanity travels barely kilometres from where they are born. There's absolutely no survival benefit to traveling large distances.

As someone who is required to travel a lot, and lives over 6 thousand km away from "home", I would love to see some sources and statistics for this claim.

any measure of "gain" will be subjective here

Lets, for our purposes, call "gain" as "otherwise impossible to do with similar cost". I.e., (an example I just made up, I don't know if its true) spending energy on heatpumps is fine, because it is otherwise not possible to climate control to living conditions at similar energy cost. What does the calculus say now ?

AkshayThu 1 Oct 2026 3:57PM

@Aaruni Kaushik That statistic is made up. But, there are statistic like the proportion of passport holders in India. India has <10% passport holders. Which means, most 90% Indians will never travel that far.

Also, you have the option to stay near your "home" and avoid 6 thousand kilometre travel.

You might find excuses to say why all of that is unavoidable.

But those are subjective.

If you don't understand how that is subjective, there's not much point in this discussion to continue.

Badri SunderarajanThu 1 Oct 2026 4:45PM

@Akshay I would not stop at "cost vs. gain" but dig deeper to ask who gains and who pays the cost?

I have been thinking about it, and my two main deal-breakers for LLM solutions have been environmental damage and lack of respect for artists/writers/other creators. Both cases have an asymmetry under this lens:

  • Environment: LLM companies gain, local communities and the rest of the world including future generations pay the cost

  • Respect: LLM companies gain, artists/writers/other creators pay the cost (and society as a whole, though maybe to a lesser extent)

(Both are also things I feel viscerally as I have seen the trivialisation of works like mine and my friends and am currently having to carry bucketfuls of water to my house every day due to a depleted water table. Both were happening before LLMs, but we were supposed to try and make the situation better, not worse.)

Production and querying of LLMs has other drawbacks too, but most of those are not deal-breakers to me in the sense that (a) I will not object to or try to discourage other people from using them in a blanket sense, and (b) I see them more as problems that we can work together to solve rather than things that should be avoided entirely.

For example, I may not use LLMs to generate code for me, and I may discourage its use in projects I care about, but if other people use it I would be fine with that—either the code turns out bad (or untrustworthy, which is also an issue if the stakes are high) in which case I don't use it, or they seem to be doing it sensibly in which case good for them. At this point, my viewpoint is becoming more like those who are currently sceptical of LLMs but focus on strict review requirements as the solution.

As another example, an LLM may come up that's very useful for generating certain kinds of code, in which case I would choose to use it based on cost much like I choose VPSes or "rent a virtual Macbook" services today (I do not rent virtual Macbooks) and perhaps with a focus on things like solutions that work on sustainable energy.

I think the main reason this cost/gain asymmetry came up is due to rich people pouring all their money into this technology. This subsidisation distorts the way people choose to use LLMs, since they have access to burn resources on trivial tasks without paying an immediate, direct price for it. If it wasn't for these subsidies, people would be paying something closer to the actual costs of LLM production and usage in a financial sense which means that for example they may look to other solutions than using up 5GWh of power for a mediocre text generator (or, conversely, the community would flock together to sustain 5GWh of power for the tool that they know is going to be super useful for whatever thing they are doing, the utility of which would be amply clear because if it weren't they wouldn't have started doing it).

Conversely, I wonder if things will stabilise when the bubble bursts and we end up with people only using this technology sensibly where it's appropriate (we'd likely have access to powerful models too, albeit at great cost).

Similarly: coal and gas being heavily subsidised make people more reluctant to switch to less polluting alternatives; reversing that to subsidise solar power and other more sustainable solutions speeds up the process of adopting those instead. (Of course, this does not address the real solution which is for us to learn to use less of those things. As the person whose birth anniversary we are about to celebrate pointed out, if everyone in this country decided to consume as much as we see our Western counterparts doing, we would strip the world bare like locusts.)

Regarding the other discussions regarding flights, heating, an more, I think they all fall at various points of the "it's bad but I can/can't do without it spectrum" for different people. Personally, I avoid flights, cars, air conditioning, and even having too many lightbulbs on at the same time—but then I have the luxury of being comfortable with multiday train journeys, being prone to carsickness, finding air conditioning physically uncomfortable, and evidently being more sensitive to light than average. We all need to do what we can, and keep pushing the boundaries of "can".

LLMs are a low-handing fruit because they are new enough that people are in many cases not completely hooked onto using them, and they are useless enough in most cases that we do not lose much by avoiding them.

AkshayThu 1 Oct 2026 5:26PM

@Badri Sunderarajan The only real point of debate (for me) here is whether LLM is exceptionally bad compared to other ills of capitalism (like bitcoin, instagram/tiktok, flights, Amazon - the consumerist platform, etc) OR whether it is just worse.

The difference between those two conclusions is the difference between extremist attitude towards LLMs alone (calling for boycott of projects which use or have neutral stance towards LLM, etc) and cautious/skeptical pragmatist attitude towards LLM (discerning between different kinds of LLM deployment, etc).

If LLMs are exceptionally bad - then the debate will have to be centered on proving that they're exceptional - and that would necessarily require comparing them with all the other forms of capitalism, technology, exploitation, etc.

If they're not exceptionally bad, but are just bad/worse/on the same spectrum, then singling out LLMs doesn't make much sense to me.

You are saying, one reason it could be singled out is that it is not yet a habit and that people don't lose much by avoiding them. I vaguely agree that it is relatively new on the scene and sounds like it is easier to take a stance against it. But I don't really feel that argument that well. It has diffused into use very widely. And it does enable people to do things that weren't possible for them in the past. Whether they feel like a loss when they stop using it is not for me to decide.

Another way of singling out LLMs is by saying that on the whole LLMs perform worse in all dimensions - environment/respect/labor market/concentration of wealth/humanity... And by being consistently worse in all spheres, it is singularly bad. That is probably true. In 2023 when chatgpt etc came out, there was a recognition that this is not just another passing fad, it is life changing. It did come with a lot of power and a lot of damage.

But all of these are, in my opinion, relative/subjective. In that even with all of these arguments there's possibility of good uses of LLMs (increasing accessibility, making coding, etc possible for people who weren't able to do this earlier, making it more easier to gain information for a lot of people, getting structured data out of unstructured data that can then be used for research, solving unsolved math problems?). So, it can always be argued that LLMs, if used judiciously, can lead to good things. Of course, practically it will never be used in that way and capitalism will destroy the world. But that theoretical possibility exists and in pockets of goodness it can be put to good use.

For that reason, objectively stating LLMs singularly are evil/unethical, etc doesn't sound right to me. There are two positions I can agree with easily.

  1. Stating that most things capitalism are evil/unethical and needs to be boycotted.

  2. Stating that we need to be judicious in using LLM.

If we indeed think in these ways, we will be able to come up with much more practical solutions, I think.

For example, imagine a browser add-on that shows how much energy gets consumed every time someone makes a chatgpt query. (There already are, but has any been widely shared in our circles?). This would single-handedly raise consciousness of people about the environment impact.

We could increase the awareness of people on how bad these are by really breaking down the cost and presenting it to people. Look at how many people have, without any knowledge, joined trends like 1980s style photos.

There are many ways people can be judicious. A whole classroom can use one LLM chat together instead of each student doing their own conversation. There could be collaborative wikis where LLM output could be shared and saved (like caching the result so another person won't have to run that query again). ...

People can really make that choice for themselves, is what I feel.

Badri SunderarajanFri 2 Oct 2026 7:43AM

@Akshay I agree with most of what you said, including the two positions you listed:

  1. Stating that most things capitalism are evil/unethical and needs to be boycotted.

  2. Stating that we need to be judicious in using LLM.

I don't think LLMs are exceptionally bad compared to other ills of capitalism like bitcoin, Amazon, and the rest (although I think they do lie more on the "worse" side of the spectrum than those things). And like you said, I don't see any particular reason to single out LLMs. The only reason we are having this debate is because "everyone" is talking about them and using them, so we are expected to either use them too or explain why we aren't.

This is partly why, while drafting the FSCI manifesto, I wasn't enthusiastic in making a special point for LLMs. We already have a list of general points and these can be applied to anything whether it be emulators, desktop environments, ovens, printer drivers, or, yes, LLMs. (My initial reason for not using ChatGPT was not specific to LLMs but just part of my general policy of not signing up to proprietary and centralised services.)

Now, coming to your suggestion of helping to ensure people use LLMs judiciously. I think this is a good idea. Where we stand on specific use-cases would still be subjective as different people would have different cost/benefit analyses and lie in different places on the hedonistic/altruistic spectrum, but at least we can get rid of the binary "all or nothing" arguments and have discussion more grounded in specifics.

Is there a reason to single out LLMs for "more judicious usage"? I would say "yes and no". We still need to keep focus on other things like sustainability, urban mobility, and so on. However, there are already many existing initiatives focused on those problems. Since LLMs are the hot topic today, it is important to ensure that we as a society walk into adopting LLMs (or not) with our eyes open, just as we have tried to do (with various degrees of success) for other technologies and developments.

More specifically: an enormous amount of resources (money, media, marketing) has been poured into making people believe that LLMs are the solution to everything. Therefore, I think I can risk making the generalisation that if LLMs are brought up as a solution to something, they are more likely than not to be an inappropriate/suboptimal solution. So, if someone were proposing an LLM based solution to me, I would assume it's inappropriate and put the onus on them to prove that it is actually appropriate. Through similar reasoning, in projects that I am involved in, I would ban usage of LLMs by default but open the door to people making a convincing argument to the contrary.

Basically, we are being intellectually DDoSed with "LLMs are the solution" so the immediate available defence is to block all "LLMs are the solution" claims and then sit down to figure out which ones we actually want to allow in. (The second part is as important as the first.)

I specified "projects that I am involved in" because as you may have pointed out I'm likely not the best person to judge appropriate level of usage for some project that I'm not involved in. (I can't comment more here since I'm not aware of the specifics but I'm going based on the fact that you mentioned seeing people utilising LLM powered tools to do things that they weren't able to do earlier.)

On a different note, I think LLMs were singled out for writing a policy about in Prav because it was a heated debate. In practice, we also discourage the use of proprietary software and centralised platforms, for example by discouraging Figma usage in favour of Inkscape or Penpot. We did not end up having to write a policy for that because it was noncontroversial within the community. But I personally think it's better to have one in place to inform people outside the community (including potential new contributors) and also in the interest on not just singling out one specific technology. To this end, I'm planning to propose a policy similar to EthicalAI but aimed at all proprietary software and centralised services.

(A note about the LLM policy I drafted for Prav vs. the FSCI one. People may be wondering why I was opposed to "just use PublicAI/these well rated LLMs" in Prav while I am open to the suggestion for FSCI. The difference is that Prav is a specific project, and that too has nothing much to do with LLMs. So, it makes more sense to forbid usage by default and stop the DDoS attack. FSCI, on the other hand, consists of various people working on various projects, so providing a universal guideline is the wrong approach: instead, the aim here is to provide information and let each project make informed choices on its own policies just as we at Prav decided ours. Finally, just noting that the above statements are with my community hat on as on a personal level I'd be rather more LLM-avoidant in terms of choosing which projects to depend on; that is again a personal choice like avoiding flights and air conditioning and not something by which to inform policy except perhaps as a single data point among many.)

Anyway, coming back to FSCI, I like the idea of rating LLMs on different scales so that people and projects can make an informed decision on which ones to use, assuming they have decided they need to use one at all, without having to conduct deep research. This is similar to the "anti-features" option in F-Droid which lets people make informed choices when installing apps.

I also feel that it is not enough because by the time you begin choosing LLMs based on ratings, you are assuming that there is a need to use LLMs at all. There can be good uses (as you have discussed) but there can also be silly uses for cases where a handwritten filetype detector tool (sorry couldn't find the source) or a GIMP brush pattern can work just as well if not better, and it is worth spending some time to make people aware of the other options they have. I see most LLM usage as a bruteforce method, and bruteforce is something you only use when you've given up on everything else; this is something worth explaining in any guideline on LLM usage given how much LLMs as a solution has been magnified over everything else. I am not sure if what I just proposed is workable at all or what would be the best way to go about this, but if I'm not wrong this is the main thing we need to start discussing next.

AkshayFri 2 Oct 2026 7:51AM

@Badri Sunderarajan as someone who opposes even software in general (the non-LLM ones) when being suggested to "solve" the wrong problem, I wholeheartedly like your comment and agree with it fully.

Pirate PraveenSun 4 Oct 2026 7:09PM

@Badri Sunderarajan I think another consideration is differentiating energy usage for training and inference. The energy usage for training remains static and energy usage for inference is smaller. So unless we completely stop training, the energy usage is much smaller after the training is done, it does not matter much if 1 query or 1000 queries. ie, you don't save the energy already spent on training by not using it.

This also depends on the type of queries as well. Code generation may take more energy but language translations may not. Also energy usage of servers is shared across its users.

Aaruni KaushikTue 13 Oct 2026 1:23AM

@Badri Sunderarajan

I would not stop at "cost vs. gain"

I would clarify that "cost vs. gain" is still a big argument, even if you want to go beyond it. In my example above, the (additional) cost of writing an email by hand is 0.5 Wh (2 watts sustained over 15 minutes, say) v/s using an LLM is 25 Wh (750 watts sustained over 2 minutes, say). That's 50 times higher cost for the same gain.

I will not object to or try to discourage other people from using them in a blanket sense

Even beyond the energy cost and the environmental cost and the social/respect cost, LLMs have an additional cognitive cost: avid users of this technology report starting to have to depend on it, and their cognitive skills begin to atrophy. (This is only relevant in principle: this cost is totally borne by people who use these LLMs, and in practice, it doesn't affect people who avoid it. But this is one reason to try to discourage other people from using them, at least if you care about these people.)

@Akshay

The only real point of debate (for me) here is whether LLM is exceptionally bad compared to other ills of capitalism (like bitcoin, instagram/tiktok, flights, Amazon - the consumerist platform, etc) OR whether it is just worse.

I would say they are not always exceptionally bad, but they are out there among the worst. (Bitcoin, for example, burns energy and drives up prices just like LLMs, but doesn't suffer from a wider ethical concern in terms of infringing upon copyright as part of the creation process. instagram/tiktok, in my opinion, cause similar cognitive decline, but doesn't use energy at quite the same scale, etc).

solving unsolved math problems

Let me briefly comment on this, since this touches upon my domain more directly.

  1. People in the domain are generally not that excited by LLMs solving unsolved math problems. LLM use is causing huge problems/disruptions in the field because it produces results, even sometimes correct results, which are then absolutely useless. The system is not really built to deal with "correct but useless" results.
    Publishing something new in the domain of math is associated with an increased understanding, a conceptual breakthrough, development of new tools, etc, as part of the new solution. That is, the journey is absolutely just as, if not more important, than the destination. LLMs have been able to (in very few cases, to be clear) short circuit this journey and provide a result which is provably true, but which doesn't advance the subject very much at all, because nobody understands the result, or how it was arrived at.

  2. The results which these LLMs produce are also questionable, in terms of the authenticity of work. For example, the recent large result from September was the Navier Stokes thing. Looking into it, you find that two human researchers were already investigating this (with help from LLMs), and the LLM company (OpenAI) more or less stole this work, and attempted to have the co-author removed from authorship of the paper (because the co-author worked in a competing LLM company) (Some links : https://mastodon.social/@tristanbuckmaster/117233413705701198 , https://mathoverflow.net/questions/502120/examples-for-the-use-of-ai-and-especially-llms-in-major-mathematical-development/515023#515023)

So I would not attribute "solving unsolved math problems" as a positive selling point for LLMs.
LLMs also appear to be stealing work and marketing it as their own in other domains : https://front-end.social/@SaraSoueidan/117347353081013904

@Pirate Praveen

I think another consideration is differentiating energy usage for training and inference

I would not worry too much about separating out these costs. As long as people keep using LLMs for inference, there will be incentive for people to train a bigger-better model.

This also depends on the type of queries as well. Code generation may take more energy but language translations may not

I think this is confusing what "AI" means? I am not against purpose built machine learning models, for example, for language translation, or voice-to-text, etc. These consume less energy, because these are not LLMs. Or was there a development I missed, where LLMs are now natively performing language translations (in which case, do I really trust a generative model to translate into a language I don't understand and not change its meaning?)?

Also energy usage of servers is shared across its users.

I don't think we need to care much about this metric either. If I understand correctly based on my experiments, one query more or less requires the entire computer while it is being processed. This is roughly a kilowatt per node, for the amount of time it is in use. The fact that it can time share multiple users doesn't really make much difference on the fact that when it is in use, its more or less 100% in use.

We could also completely forget about servers, and just focus on the energy consumption cost we know: around 300 watt-hours per million tokens on the thaura setup, which is "10 times more efficient" than commercial providers, so, around 3 kilowatt-hours per million tokens outside of the thaura setup.

Badri SunderarajanTue 13 Oct 2026 9:05AM

@Aaruni Kaushik responding to this:

I would clarify that "cost vs. gain" is still a big argument, even if you want to go beyond it. In my example above, the (additional) cost of writing an email by hand is 0.5 Wh (2 watts sustained over 15 minutes, say) v/s using an LLM is 25 Wh (750 watts sustained over 2 minutes, say). That's 50 times higher cost for the same gain.

Fair enough as far as numbers go, but from a hedonistic perspective, the cost to me of writing an email by hand is 0.5 Wh and 15 minutes, whereas the cost to me of using an LLM is 0.0 Wh and 2 minutes. This is the "cost" that most people are using to decide whether to use an LLM for a particular task. While that doesn't nullify your argument, it's important at the very least to specify which costs and gains you are including.

On your other comment:

Even beyond the energy cost and the environmental cost and the social/respect cost, LLMs have an additional cognitive cost: avid users of this technology report starting to have to depend on it, and their cognitive skills begin to atrophy. (This is only relevant in principle: this cost is totally borne by people who use these LLMs, and in practice, it doesn't affect people who avoid it. But this is one reason to try to discourage other people from using them, at least if you care about these people.)

Those are some of the "other drawbacks" I mentioned, yes. Quoting myself,

Production and querying of LLMs has other drawbacks too, but most of those are not deal-breakers to me in the sense that (a) I will not object to or try to discourage other people from using them in a blanket sense, and (b) I see them more as problems that we can work together to solve rather than things that should be avoided entirely.

An interpretation of the above statement when applied to your comment: Besides the energy, environmental, and social/respect costs, production and querying of LLMs has other drawbacks such as an additional cognitive cost: avid users of this technology report starting to have to depend on it, and their cognitive skills begin to atrophy. (a) While I will not attempt to push for total bans on LLMs for these other reasons, I may object to or try to discourage people I care about from using them, and (b) despite not being as pressing issues as the energy, environmental, and social costs that I highlighted earlier, they are definitely problems that should be addressed once the aforementioned higher priority problems are resolved.

Agreed with your other responses, although I'll reserve judgement on the maths one because a lot of people here at Chennai Mathematical Institute seem freaked out and what the models have recently helped to compute. Anyway, I think "not very effective at their work" is a relatively low priority issue; I have no problem with people spending time on ineffective tools if the rest of us don't end up paying the costs for it 🙃

Pirate PraveenMon 16 Nov 2026 11:28AM

@fugata mentioned this in xmpp group,

"Still ecologically awful, still built by breaking the free software contract, still complicit in destroying the web, still a tool for deskilling."

  1. We could run this on renewable energy, use heat generated to warm homes like Infomaniak's Euria.

  2. I think a weakening on copyright is good for functional knowledge, I have a concern about knowledge getting locked into a few big llm companies, but things like open weight models and public/open source ai can mitigate that. It was similar for search engines, it also needed huge computing resources to index the whole web, but we did not give up on search because of it.

  3. I share this concern, I think things like Common Crawl can mitigate that if people don't have to crawl multiple times.

  4. I understand this concern, but it is a very narrow view thinking there is only one way to use LLMs. Even a good skilled programmer can use LLM to work on a different programming language they don't know the syntax well enough, but can understand the logic. A sys admin could use LLM to generate code for some situations etc. Not everyone need to be skilled in every programming language. LLMs can help fix specific issues in complex software without depending on the original developers.

fugataWed 18 Nov 2026 7:50AM

I'm going to be highly skeptical of any environmental claims without deep scrutiny. With the hype, there's a pressure to greenwash as well as a pressure to create greenwashed alternatives.

I'll talk about #2 in a separate reply.

LLMs can help fix specific issues in complex software without depending on the original developers.

...by moving the dependency to probabilistic automation companies instead. No, I would not recommend anyone do that. My recommendation will always be to do the right thing, learn things the proper way, and to improve one's skills and knowledge.

Also, "knowing the logic" isn't enough. If you don't know the syntax and semantics of the language, you are not qualified to review the LLM output.