[E-impact] Re: Gen AI vs AI usage - looking for a taxonomy or guide to understand the different kinds of AI used, and their relative resource consumption

[Sébastien Rumley <[e-mail]>]

Hi Chris,

We are running the swiss ai center, devoted to help SMEs to adopt AI 
(there are 600k SMEs in Switzerland and 2/3 of the swiss population 
works for them - we are a nation of SMEs). 
(https://app.swiss-ai-center.ch/ - I just realized how outdated is our 
website, as usual...)

We clearly see two classes of "clients" :

- One class has some data accumulated, like a few years of logs, several 
days of measurements of vibrations of machines, etc. In that case it is 
often possible to start a "good-old" ML project to do predictive 
maintenance, for instance. With about 15-30kCHF (part of which is 
covered by the company, the rest by the center) we can establish the 
viability of training a model on purpose, or fine-tuning an existing 
one. Sometimes the company follows-up with a larger project resulting 
into an operational mechanism in production. In some way it is a bit 
"technology push" because we get the data first and see afterwards how 
we can exploit them, but when it works, it is meaningful. We had one 
project in the power sector you mentioned, and we prove that we could 
find a system that would "fill the blanks" rather efficiently (a blank 
is when the monitoring system is down for a few hours).

- Another class has no real data at hand. They want AI to automated time 
consuming tasks, like pre-processing incoming mail, pre-filling forms, 
etc. They have read the news and they are concerned about 
competitiveness. For those guys, we often end up assembling an LLMs with 
a RAG system, sometimes a few more logic boxes.. and we cannot really 
prove that it is useful because of the absence of test data.  Even tough 
it is a bit "market pull" in the sense that the company declares its 
pains and we try to work on it, we often end up pouring all the 
"description of the pain" inside an LLM hoping that a relief will come out.

I started to mention a disclaimer to the second class, I tell them "sure 
we can try to do something for you but be aware that most likely in 2-3 
years microsoft, oracle and the likes will have advanced enough their AI 
tools that you might be able to buy your functionality on the shelf".


On a more philosophical level, the "Gen-AI" term only put the emphasis 
on the output (generating the response), but analysis the input is also 
crucial. We have to reckon that language (+ quotes, points, italic 
text...) is very good at encoding concepts and information (no surprise, 
it results from thousands of years of evolution). Now that machines can 
finally parse this quite information dense (but often ambiguous) flow, 
we can insert easily and quite rapidly information inside a computer. 
Any information. No more need to structure it (with expensive UI, for 
instance), the system will do it for you (just hope it'll do it 
correctly). This makes LLMs incredibly generic because they can deal 
with whatever input (including wonky one with single quote instead of 
double). And the chatbot UI is also incredibly generic.

This is why we see many many cases of "prompt engineering" where 
essentially you send a big text with all kind of instructions to the 
LLMs. Instead of opening your calculator to perform 13 * 32 (or, would 
say my father, instead of doing the maths yourself to stay mentally 
fit...), you just ask "What makes 13 * 32 ?"... and you burn 100'000 
more energy for this. But it is generic, always open, and cheap. And the 
chatbot sometimes is nice enough by asking if you also want it to 
calculate 13 * 33... finally someone nice in this ugly world...


So, to conclude, I believe the main differentiator is the genericity. 
"traditional AI" is trained for a specific problem, needs a specific 
structure, and, when done correctly is generally pretty good and 
efficient at solving the problem. "(LLM-based)-Generative-AI" is trained 
for any problem provided that it is expressed in text, it takes 
unstructured (in the computing sense, not linguistic sense) data in, 
always provide an answer... and the quality is questionable.. the 
efficiency... very poor. But you don't know about it because energy is 
too cheap. And because BigTech want to establish a monopoly on (alleged) 
problem solving business so they sell it at incredibly discounted rates.

But rest assure, when most of the people will have lost their own 
problem solving capabilities, discounts will disappear...

Does that help ?

Sébastien


On 16.06.2025 11:52, Chris Adams wrote:
>
> 	
> You don't often get email from [e-mail]. Learn why this is 
> important <https://aka.ms/LearnAboutSenderIdentification>
> 	
>
> Hi folks,
>
> I’m trying to make sense of the way different kinds of AI are 
> currently portrayed in public discourse, and in particular how GenAI 
> and “traditional AI” after conflated on climate grounds to justify 
> support for massive digital infrastructure build out. I wondered if a 
> kind soul might help me with this question:
>
> /Have any of you come across any meaningful taxonomy of AI techniques 
> to help understand the differences between techniques we associate 
> with more "established ML / traditional AI" that we see generating 
> value in the power sector for example, compared to the more splashy 
> "Generative AI" that generates a growing amount of coverage in the media?/
>
> *A little more context on where I’ve been looking so far*
>
> I’ve written a bit more in this blog post, showing three examples 
> taking about this /Gen AI / trad AI/ split, from:
>
> a) academics I class as experts in the field
> b) a well regarded industry report by one of the largest suppliers of 
> electrical equipment to datacentres, and
> c) another framing from a well known tech journalist that found helpful
>
> I’d welcome pointers to help people have somewhat more nuanced 
> discussions about the different kinds of AI used. This is because “AI" 
> keeps getting invoked in public discourse around national 
> competitiveness, technological progress and so on, like it’s all one 
> monolithic thing, and I don’t think that’s helpful.
>
> https://rtl.chrisadams.me.uk/2025/05/different-kinds-of-ai-in-the-climate-context/
>
> *Further context*
> *
> *
> I've also written a follow on post, looking at one taxonomy that 
> /might/ be helpful
>
> https://rtl.chrisadams.me.uk/2025/06/is-the-problem-just-genai/
>
> Finally to be super, /possibly unnecessarily/ clear, I’m only talking 
> about the direct environmental impacts right now, (so the X axis on 
> this simplified 2 by 2 in this blog post below). I’m not really 
> talking about what you use AI /for, /for this query.
>
> https://rtl.chrisadams.me.uk/2025/02/sustainability-via-tech-vs-sustainability-of-tech/
>
> I’d welcome more pointers, hence the email here.
>
> Thanks in advance!
>
>
>
>
> Chris Adams
>
> Director of Technology and Policy
>
> w: thegreenwebfoundation.org <http://thegreenwebfoundation.org/>
> e: [e-mail]
> t: @mrchrisadams
>
> German Office
> Naunynstrasse 40
> 10999 Berlin
> Germany
>
> See our contact page for more details
> https://www.thegreenwebfoundation.org/contact/
>
> Book a short call with me to discuss something.
> https://cal.com/mrchrisadams
>
-- 
Dr. Sébastien RUMLEY, Professor
iCoSys Institute, part of
HEIA-FR - School of Engineering and Architecture - Fribourg, part of
HES-SO - University of Applied Sciences and Arts Western Switzerland