On The Merits Of LLMs

[nephele]

@nipos

Unfortunately no photovoltaic here because I don’t own the house I live in,so that’s not my choice.

If you are a tennant in DE you can use a Balkonkraftwerk, the House owner is not allowed to prevent you from doing this

[suhr]

You make a good point, but what If we make the primary focus on us using open-source language models like BLOOM to reject the closed-source monopoly.

Picking one of the most useless models ever made is not exactly an argument in favor of open-source models. There are way better open source models.

[Philipingram25]

@nephele

You can’t “cancel out” bad effects like this. Sure you may have a solar panel on your roof, but that didn’t stop AI data centers from re-starting already discontinued coal plants, and that completely ignores the cost of components, cooling (water) etc.

I know we can’t cancel out the bad effects it going to take year to do that but I feel like if we push eco friendly everywhere we still doing something and I feel like that better than doing nothing mainly but we can still make a tiny impact and it can help a lot

And I wasn’t trying to say eco friendly magical erases harm caused to the environment but it more that Eco friendly used as a collective causes less harm to the environment than the search engine and ai we are currently using right now

@nephele

the economic costs that these Data centers cannot ever recoup the investment, they are build to become e-waste in 2/3 years… so you have the enviromental impact of recycling to contend with aswell

And I know this but OceanHero managed to help the ocean by a lot and I feel like using this search engine, for example OceanHero the community removes more than 825,000 plastic bottles from oceans and coastlines globally over a single 30-day window. This may not fix the E waste but it helps the Ocean.

BLOOM has A Carbon footprint of 50.5 tonnes is remarkably low compared to commercial AI models of similar scope

With this point I was just trying to say making more action and coming together to make Haiku eco friendly since it make a difference as we are doing something then sitting around and being aware of it which is why we can push for more of a eco friendly comprise.

@suhr

Picking one of the most useless models ever made is not exactly an argument in favor of open-source models. There are way better open source models.

I know Bloom is not the best Open source model but it only I was mainly doing research on, and I will look at more open source model so I can make a more valid argument.

The point I’m trying to make is Haiku is used by million or more since it very undocumented but if an operating system used by millions builds eco-friendly and resource-efficient habits directly into its core design, it massively does far more collective good than just sitting back and doing nothing.

[nipos]

If it makes sense in the building,true.
Unfortunately I have the sloping roof at the side where the sun shines most of the day and the balcony where I could place one is at the side where the sun reaches it only for a few hours in the evening.

[PulkoMandy]

@Andrea

Is your car electric? Does your home run on diesel or natural gas for heating? Does your electricity provider, which powers your computer, guarantee that the energy comes 100% from renewable sources? Do you produce the energy used to write and read your posts on this forum via photovoltaic systems? If so, then we can discuss the ethical issue of resources. But I don’t think that’s the point of this forum.

I don’t have a car. I have a bicycle built in 1998. And travel by train in cases where cycling is not practical.

My appartment is heated by a wood chip/pellet burner (shared with several buildings) fed from local renewable sources. Electricity is mostly nuclear and renewable. The building also is equipped with solar panels, although the electricity is sold to the power company, but in practice a lot of the electrons I use actually come from there. I think I am doing reasonably well in that regard.

I think I did my best on these aspects.

And I also think keeping Haiku usable and developpable on old machines (32 bit support, low RAM requirements, optimization of the system) also fits in that category and is a somewhat important part of the identity of the project. How important exactly, depends on who you ask.

I agree that this has not been written down in the core values of the project until now, so, that is just my own opinion so far. The reason it had not been written down is that this impact was almost none as long as we only had an svn server and a website, or even all our infra running on a single machine. It has gotten a tiny bit larger with continuous integration. Intensive use of LLMs at least associates it with digital computing resources of a mach larger scale. The practical difference from Haiku users and developers (maybe a hundred people in total?) using LLMs is probably not that large in that aspect. So it’s more about sending a “we care about that” message than actually reducing our emissions. It is about showing an example of what a small team with low resources can do.

[Andrea]

@nipos

My home uses natural gas for heating,but I don’t need like 25C all the time,18-19C is enough and I don’t keep the windows open longer than needed.
My energy is from 100% renewable sources,yes,when comparing prices renewable energy is often cheaper than mixed energy here in Germany anyway.
Unfortunately no photovoltaic here because I don’t own the house I live in,so that’s not my choice.

Fortunately, throughout the European community, it is possible to install tools similar to this one: TRONIC® Balkonkraftwerk 860 Wp / 800 W TOPCon »TBKT 800 A1« | Lidl

Of course, it doesn’t work miracles, but it’s something—especially for offsetting your PC’s energy usage

The damage caused by your gas boiler is far worse than that of a heat pump, but I’m not here to argue with you just because you heat your home using such outdated and highly unethical methods

Riepilogo

Feliĉe, tra la tuta Eŭropa Komunumo eblas instali ilojn similajn al ĉi tiu: …
Kompreneble, ĝi ne faras miraklojn, sed ĝi estas io, precipe por kompensi la energion por via komputilo
La damaĝo, kiun vi kaŭzas per via metana kaldrono, estas ekstreme pli malutila ol varmopumpilo, sed mi ne polemikas kun vi, eĉ se vi varmigas vin per ĉi tiuj ekstreme maletikaj kaj antikvaj metodoj

[suhr]

Fun fact: the word “riepilogo” does not exist. It is completely made up by an LLM.

[Andrea]

Fun fact: Nephele can post comments without the 30-minute delay that we poor, mere mortals have to deal with.

[nipos]

It seems the delay has been removed from the topic.
I don’t see the message at the bottom anymore.

[Zardshard]

@axeld

Translation between German and English usually works well with mechanical tools. From my own experience, this is not always true for other languages like, for instance, Russian, though.

Ah, you’re right. Let me try some Japanese and Korean. I was going to try some Chinese, but it doesn’t appear Google lets me choose between classic and LLM modes, so I don’t know which method it uses.

Example 1 - Japanese

Original (Source):

滋賀県に住んでいた頃に不思議な飲食店に行ったことがあり、探しています

Google Translate classic:

I went to a strange restaurant when I lived in Shiga Prefecture, and I’m looking for it.

Google Translate llm:

I once visited a mysterious restaurant back when I lived in Shiga Prefecture, and I am currently trying to find it.

Example 2 - Japanese

Original (Source):

ドゥカティとかBMWのバイクって日本だけ異常に安くないですか？
S1000RRが300万円
パニガーレV4Sが420万円

Google Translate classic:

Aren’t Ducati and BMW motorcycles unusually cheap only in Japan?

S1000RR: 3 million yen

Panigale V4S: 4.2 million yen

Google Translate LLM:

Aren’t Ducati and BMW motorcycles—specifically in Japan—abnormally cheap?
The S1000RR is 3 million yen.
The Panigale V4 S is 4.2 million yen.

Example 3 - Korean

Original (Source):

아놀드 대변인은 "우리는 이 도구를 개발하고 배포하는 과정에서 잠재적인 개인정보 위험을 신중하게 고려했다"며 "관련 법규를 준수하기 위해 최선을 다하고 있다"고 덧붙였다.

Google Translate classic:

“We carefully considered potential privacy risks in the process of developing and distributing this tool,” Spokesperson Arnold added, “and we are doing our best to comply with relevant regulations.”

Google Translate llm:

Arnold stated, “We carefully considered potential privacy risks during the process of developing and deploying this tool,” adding, “We are doing our utmost to comply with relevant laws and regulations.”

Example 4 - Korean

Original (Source):

개인정보보호위원회(위원장 송경희, 이하 ‘개인정보위’)는 5월 27일(수) 제10회 전체회의를 열고, 네이버 주식회사(이하 ‘네이버’)의 검색 인공지능(AI) 에이전트 서비스(이하 ‘AI Tab’)에 대한 사전적정성 검토 결과를 심의·의결하였다.

Google Translate classic:

The Personal Information Protection Commission (Chairperson Song Kyung-hee, hereinafter the ‘PIP Commission’) held its 10th plenary meeting on Wednesday, May 27, and deliberated and approved the results of the preliminary adequacy review of Naver Corporation’s (hereinafter ‘Naver’) search artificial intelligence (AI) agent service (hereinafter ‘AI Tab’).

Google Translate LLM:

The Personal Information Protection Commission (Chairperson Song Kyung-hee; hereinafter referred to as the “PIPC”) convened its 10th Plenary Meeting on Wednesday, May 27, to deliberate on and adopt a decision regarding the results of a preliminary appropriateness review of Naver Corporation’s (hereinafter referred to as “Naver”) search AI agent service (hereinafter referred to as “AI Tab”).

Personal opinion: They’re pretty similar. The LLM sounds slightly better, though. Of course, I can’t be sure if it just took unprofessional writing and made it slightly more professional in English or not. And I have no idea what is going on in example 4

[Andrea]

The summary is written in Esperanto.

[us3r1d]

@suhr

Fun fact: the word “riepilogo” does not exist. It is completely made up by an LLM.

DuckDuckGo finds a lot of translation dictionary sites that disagree with you on that:

riepilogo at DuckDuckGo

@Andrea

The summary is written in Esperanto.

OK, I laughed when I read that. It would have been funnier if it were true, though; the translation dictionary sites at the above link all say it’s legit Italian, though.

[VoloDroid]

@axeld

Translation between German and English usually works well with mechanical tools.

Funnily enough, I have an anecdotal evidence where classic non-LLM Google Translate keeps hallucinating for German-to-English translation, but Gemini, ChatGPT or CoPilot produce much better and accurate translation. I found it like 2 months ago, and it’s still there today, so at least it confirms that NMT are quite deterministic

Try to translate in Chrome or on Google Translate site (using the Classic mode) the following text from the article in German Wikipedia about the Invalidenstraße in Berlin:

German text

Die Straße wurde im 13. Jahrhundert als Spandauer Heerweg angelegt. Der spätere Name der Straße geht auf das Invalidenhaus zurück, das Friedrich II. 1748 zur Versorgung der Kriegsversehrten aus dem Ersten und Zweiten Schlesischen Krieg errichten ließ. Der Name Invalidenstraße erschien um 1800 auf den Berliner Stadtplänen, in den Adressbüchern trug das Invalidenhaus noch bis zum Jahr 1828 die Adresse Vor dem Oranienburger Thore. Er findet sich erstmals 1830 mit diesen Details: „Invalidenhaus, vor dem Oranienburger Th., in der Invalidenstraße“.

The translated text looks like:

English translation

The street was originally laid out in the 13th century as a military road through Spandau. Its later name derives from the Invalidenhaus (veterans’ home), which Frederick II had built in 1748 to care for the wounded soldiers of the First and Second Silesian Wars. The name Invalidenstraße (Vultures’ Street) first appeared on Berlin city maps around 1800, but in address books, the Invalidenhaus was still listed as being “Vor dem Oranienburger Thore” (Before the Oranienburg Gate) until 1828. It is first recorded in 1830 with the following details: “Invalidenhaus, vor dem Oranienburger Th., in der Invalidenstraße” (Vultures’ Home, before the Oranienburg Gate, in Invalidenstraße).

For some reason, Google Translate decided to provide the translations of places and streets in parentheses, and that’s where the translation went from not so great (first time “Invalidenhaus” was translated as “veterans’ home”, which is not 100% correct but is still somewhat close contextually), to a completely wrong one, by translating “Invalidenstraße” as “Vultures’ Street” and “Invalidenhaus” as “Vultures’ Home”. If someone doesn’t remember what a vulture is (I didn’t), here’s the definition straight from the Cambridge dictionary: “vulture is a large bird with almost no feathers on its head or neck, that eats the flesh of dead animals”. That’s a far cry from “Invaliden”, which means “people with disability” (apparently, it was often used in the past for describing injured veterans, if I’m wrong please correct me on that).

If you start to remove some sentences from the original text, “Vultures’ Home” will transform into “Victims’ Home”, then “Veterans’ Home” and finally to “Valvaries’ Home” which is not even a word.

[Andrea]

@us3r1d

OK, I laughed when I read that. It would have been funnier if it were true, though; the translation dictionary sites at the above link all say it’s legit Italian, though.

Oh, I’m sorry, I didn’t understand. Yes, ‘Riepilogo’ is Italian and it means summary.

[us3r1d]

On the “merits” question, by the way: have y’all seen the blowup over rsync that started yesterday?

As near as I can tell, the maintainer seems to have full-on YOLOd a Claude rewrite of the test suite from bash to Python (among other things). The sysadmin community is … somewhat irate … at the number of new breaking bugs that shipped on what almost all of us thought of as fundamentally finished software.

This is not to say that all use of LLM-backed coding assistants will necessarily result in a fiasco like this, but it is absolutely the sort of thing that will happen sometimes.

And I realize a lot of folks response is “he was doing it wrong”, and that is true as well; but some people absolutely are doing it this way. And now we have a pretty solid real-world example of it going very wrong.

This sort of indeterminacy in the results is why I personally avoid this stuff.

[nipos]

That was a good reminder why having alternatives and choices can sometimes be a good thing.
A alternative implementation made by the OpenBSD team has existed for years,so it’s mature enough to be a drop-in replacement in most cases.

GitHub - kristapsdz/openrsync: BSD-licensed implementation of rsync

MacOS already ships openrsync and FreeBSD,NetBSD and some Linux distributions have it in their repository.

[us3r1d]

@nipos

MacOS already ships openrsync and FreeBSD,NetBSD and some Linux distributions have it in their repository.

haikuports, on the other hand, uses this original rsync; the version there (3.2.7) is before the rewrites started (which folks who looked through the repo said is version 3.4).

[jackburton]

Any rewrite inevitably introduce new bugs: that’s why big software companies almost never rewrite anything.

[jackburton]

Or, better said, smart big companies never rewrites anything unless it’s absolutely necessary