@osma@sigmoid.social

· 5mo ago

Tämä on ehkä väärä some tähän, mutta yritetään.

Ei löytyisi Helsingistä / pk-seudulta ihmistä, jolla on jonkinlaiset sukellusvarusteet ja taito sekä hinku käyttää niitä?

Kävi niin ikävästi, että puhelimeni putosi mereen Koivusaaren rannassa. Ihan laiturin viereen, ehkä 3 metriä syvää, pohja voi olla mutainen ja vesi sameaa. Itse laitteella ei niin väliä, mutta muistikortilla oli ... muistoja.

Palkkion lupaisin, jos joku sen onkisi ylös ennen kuin pilaantuu. YV.

#sukellus #sukeltaminen

sigmoid.social

13

16

57

0

Osma Suominen @osma@sigmoid.social

· 9mo ago

Replying to

@Linebyline yes, that's a difficult com-putin-g exercise... @gabrielesvelto

18

0

Osma Suominen @osma@sigmoid.social

· 5mo ago

Sitten joskus kun saan aikaiseksi, voisin perustaa viivyttelijöiden tukiyhdistyksen. Sen nimeksi tulisi Pro Krastinaatio ry.

4

0

1

0

Osma Suominen @osma@sigmoid.social

· 5mo ago

Replying to

@reimamak En ole tuota listaa enää ylläpitänyt, kun alkoi käydä raskaaksi - ja alun perinkin olin asettanut ylläpidolle määräajan, joka ylittyi. Jos joku haluaa jatkaa niin tämän saa mielellään ottaa haltuun, vaikka forkata (githubissa) tai siirtää jollekin fiksummalle alustalle! @srholop @hepo

3

0

1

0

Osma Suominen @osma@sigmoid.social

· 5mo ago

@kaspian@mastodontti.fi Tosiaan, kirjoitettu tanska ja norja (no, ainakin bokmål) ovat ruotsia sujuvasti lukevalle käytännössä sama kieli.

Tästä tulee muistoja omista printterikokeiluista. Esim. kun kytkin TI-85:n matriisiprintteriin (vaati aika paljon elektroniikkaa väliin!) tai kun vielä aiemmin tein GWBASICilla ohjelman, joka teki ESC-koodeilla "kaunokirjoitusta", jonka olin itse suunnitellut ruutupaperilla. Ja vielä Linux-käytön alkuvuosina oli ihan normaalia tuupata printterille suoraan tekstitiedostoja.

2

0

Osma Suominen @osma@sigmoid.social

· 5mo ago

Replying to

@rubdos@mastodon.social Nothing major I guess... I use SailOTP for one time password generation, GPSInfo to wake up my GPS in case I want to use location in Android apps, and some really basic apps like Messages, Notes and Calculator. Oh, and SailTuner for tuning musical instruments!

2

1

0

Osma Suominen @osma@sigmoid.social

· 5mo ago

Replying to

@rubdos@mastodon.social Thanks for the writeup and all the app tips! I've also used Sailfish as my main phone OS for over ten years (and N9 before that). I use much fewer apps than you do and learned a lot from your post. The Firefox news sounds great!

2

11

0

Osma Suominen @osma@sigmoid.social

· 11mo ago

Replying to

@dgar@aus.social Amazing, that's the e-mail account for my luggage!

5

0

Osma Suominen @osma@sigmoid.social

· 10mo ago

Replying to

@amberage The key words "MUST (BUT WE KNOW YOU WON'T)", "SHOULD CONSIDER", "REALLY SHOULD NOT", "OUGHT TO", "WOULD PROBABLY", "MAY WISH TO", "COULD", "POSSIBLE", and "MIGHT" in this document are to be interpreted as described in RFC 6919. https://www.rfc-editor.org/rfc/rfc6919.html

rfc-editor.org

4

0

1

0

Osma Suominen @osma@sigmoid.social

· 5mo ago

Replying to

@timolaine @stuartyeates Not sure how much I can contribute here - I know a lot about Linked Data and Wikidata, not so much about LCCNs or Koha. I do have colleagues who work on Koha, but they are not on the Fediverse AFAIK.

1

0

Osma Suominen @osma@sigmoid.social

· 5mo ago

Replying to

@jabsonik @nen @saaste @Grade Ei sillä toki minulle väliä ole, mutta en nyt oleta että kukaan tulee toiselta puolen Suomea tätä varten :)

1

0

Osma Suominen @osma@sigmoid.social

· 5mo ago

This "Traffic Light Protocol for AI generated content" seems like a useful way of expressing how carefully any AI tools have been applied in producing code, text or other content. I've seen other AI usage disclosure frameworks (e.g. CEUR-WS AI taxonomy) but they typically concentrate on *what* the AI did, not how *much human judgement* was involved, which is the focus of TLP:AI. That is crucial if we don't want to fall into comprehension debt.

https://nlkw.de/en/blog/ai-tlp/

Any similar ones?

nlkw.de

1

0

5

0

Osma Suominen @osma@sigmoid.social

· 37mo ago

Replying to

Recording of my #swib23 lightning talk "Extracting metadata from grey literature using large language models" now available: https://www.youtube.com/watch?v=wOzWG60SS_g #metadata #llm #ai #cataloguing #llama2 #gpt

17

0

15

0

Osma Suominen @osma@sigmoid.social

· 8mo ago

Replying to

@poetaster @lari Wow, thanks a lot to both of you! I will seriously consider switching to this client. A noob question: can you select the language for a new toot? I tried an earlier version and couldn't figure it out.

1

0

Osma Suominen @osma@sigmoid.social

· 9mo ago

Replying to

@sarahalang@fedihum.org It seems to me that your LLM usage was quite prompt-heavy since the documents were quite long (IIRC you didn't state how many tokens average/median) and the output was just a classification (relevant/irrelevant) plus a few bits of structured information. Thus your energy use was likely severely underestimated. 40 kWh is not that much.

1

0

1

0

Osma Suominen @osma@sigmoid.social

· 9mo ago

Replying to

@sarahalang@fedihum.org Thanks for confirming! I think that prompt (input) processing is important too, especially for tasks where the input size is large relative to the output (e.g. in document classification, where the input can be long and the output is just a few tokens). PP is typically much faster than generation, but can still be a significant factor in response time and thus energy use.

1

2

0

Osma Suominen @osma@sigmoid.social

· 9mo ago

Replying to

@sarahalang@fedihum.org Thanks for confirming!

1

0

Osma Suominen @osma@sigmoid.social

· 9mo ago

Replying to

@sarahalang@fedihum.org OpenAI style API endpoints like the one you used report both the number of input and output tokens in their responses, so it shouldn't be hard to use those counts to estimate energy use. Or you could use a tokenizer library (e.g. tiktoken or transformers) to count the tokens before sending them to the API. Obviously there are limitations in estimating the energy spent per token (as you described), but I don't think getting the prompt/input token count should be the hard part here.

1

0

1

0

Osma Suominen @osma@sigmoid.social

· 9mo ago

Replying to

@sarahalang@fedihum.org Thank you, there's a lot of insight in this paper!

A couple of questions, if I may:

Did you only consider LLM-generated tokens in your calculations? Or also the tokens in the prompt that need to be processed before generating new ones?
Am I right that the total energy consumed by LLMs was around 40 kWh (or 200 kWh, if you consider the 5x pessimistic estimate)? I was expecting this figure to be shown prominently in the paper, but couldn't find it stated directly.

1

3

0

Osma Suominen @osma@sigmoid.social

· 11mo ago

Replying to

@ink Please do, sounds very interesting! @swib

1

0

Osma Suominen @osma@sigmoid.social

· 44mo ago

Replying to

@pygy@mamot.fr @L_howes@mastodon.social Happened to me as well. Also GitHub Copilot did it, suggesting I use Math.randomseed() for seeding the RNG (which IIRC isn't possible in the JS stdlib)

1

0