Post · 𑐴𑐶𑐟𑐶

Timnit Gebru (she/her) @timnitGebru@dair-community.social · last month

Content Warning

Friends, for something to be open source, we need to see

1. The data it was trained and evaluated on

2. The code

3. The model architecture

4. The model weights.

DeepSeek only gives 3, 4. And I'll see the day that anyone gives us #1 without being forced to do so, because all of them are stealing data.

gunstick @gunstick@mastodon.opencloud.lu · 4 weeks ago

Content Warning

@timnitGebru they don't keep the training data!
Data is scraped from the internet and fed into the AI.
What could be kept is the list of scraped URLs.
But that's also not kept.
The only info on training data is the source code of the scraper.

+ -

Raptor :gamedev: @raptor85@mastodon.gamedev.place · last month

Content Warning

@timnitGebru not universal but a lot of models, including major ones, give #1 on huggingface fyi, it's in the model card info. (not always easy to find though, often you have to go back to the base model)

Like 99% of models though you'll find it's just common crawl, wikipedia, and webtext2, with some specialized datasets used on top of that for more specialized models (like github for the code models)

Timnit Gebru (she/her) @timnitGebru@dair-community.social · last month

Content Warning

@raptor85 I’m hoping you knew that I know this info when you commented right? Given that you reference the framework that I co-created to do such things, called…..model cards?

https://dl.acm.org/doi/abs/10.1145/3287560.3287596

pettter @pettter@social.accum.se · last month

Content Warning

@timnitGebru OLMO and Bloomz among others seem to provide 1. according to https://opening-up-chatgpt.github.io/

Mr. Teatime @Mr_Teatime@social.tchncs.de · last month

Content Warning

@timnitGebru
Just came across this paper, published in *Nature*, no less, which is probably a good background read on the topic:

https://www.nature.com/articles/s41586-024-08141-1

Full disclosure: I'm not deep enough into the topic to properly get it. The abstract speaks volumes already, though: "open" AI is usually not actually open, and even then it can still be a big problem.

Fábio Costa 🐇🕐🎩☕ @fabiocosta0305@ursal.zone · last month

Content Warning

@timnitGebru TBH, almost everyone "provides" 2, as they are based on TensorFlow/PyTorch

sn 🐦‍⬛ @sn@mastodon.ping.de · last month

Content Warning

@timnitGebru have you seen TEUken? It's not SOTA but improving … and they published their training data

lordjeff @lordjeff@mastodon.social · last month

Content Warning

@timnitGebru i believe china doesn't recognise copyright of non-chinese sources

@BjornW@mastodon.social @BjornW@mastodon.social · last month

Content Warning

@timnitGebru thank you for sharing.

Quick question have you seen this initiative "European Open-Source AI index"?

https://www.osai-index.eu

By @dingemansemark & @andreasliesenfeld from @Radboud_uni

Looks good to me, to help people determine how open an AI model actually is. Are you aware of other initiatives like this?

I'd like to gather these initiatives and share it with @publicspaces so more people learn what to look for to determine if an AI is truly opensource.

#AI #OpenSource

jeremiah @jeremiah@tldr.nettime.org · last month

Content Warning

@timnitGebru In terms of knowing the costs of a model, would we also want to know the hardware used?

Manish @mani@writing.exchange · last month

Content Warning

@timnitGebru Most people in various tech communities on the internet don't even know that they can't view the source code of 'opensurce' Deepseek. Makes me wonder if anyone is even reviewing the code of actual opensource projects nowadays to verify their claims.

Efi (nap pet) 🦊💤 @efi@chitter.xyz · last month

Content Warning

@timnitGebru who called it open source without having the source code? wth?

FluffyBunnikins @FluffyBunnikins@mastodon.social · last month

Content Warning

@timnitGebru The Beaverton was On Point yesterday, though.
https://thebeaverton.com/2025/01/chinas-new-and-cheaper-magic-beans-shock-americas-unprepared-magic-bean-salesmen/

Cai @cai@mastodon.social · last month

Content Warning

@timnitGebru 💯

Lien Rag @lienrag@mastodon.tedomum.net · last month

Content Warning

@timnitGebru

Sky-T1 is supposed to fill all these criteria, from what I read...

netskaven @netskaven@metalhead.club · last month

Content Warning

@timnitGebru I don´t need open source stuff, we wanna FREE stuff ;)

Pixdigit @Pixdigit@layer8.space · last month

Content Warning

@timnitGebru Nextcloud has an "ethical ai rating" which takes the availability of the training data into account. Admittedly there are very few getting the "green" rating but it's not nothing:
https://docs.nextcloud.com/server/latest/admin_manual/ai/overview.html#ethical-ai-rating

bureauxx @bureaux@mastodon.social · last month

Content Warning

@timnitGebru based

@festal

i have never understood what people mean, when they talk about the "stolen data" of AI. every search engine crawls the net, the internet data base backs versions of webpages up, the idea of open source is based on re-using and modifying "data" of others, artists make collages from other images, the whole concept of "knowledge" ist based on the use of other knowledge... so, what exactly does Ai do differently?

Adam Wulf @adamwulf@mstdn.social · last month

Content Warning

@timnitGebru I think the Pleias 1.0 model gets close? unless I've misunderstood what they've open sourced. https://huggingface.co/blog/Pclanglais/common-models

POSUTtRUmp @POSUTtRUmp@mastodon.social · last month

Content Warning

@timnitGebru All Ai enter the octagon. Only one will emerge victorious.

Ingo Wichmann @ingo_wichmann@digitalcourage.social · last month

Content Warning

@timnitGebru the folks at OSI did a good job defining that just in time:
https://opensource.org/ai

Ilkka Tengvall @ikkeT@mementomori.social · last month

Content Warning

@timnitGebru IBM granite models are said to be open source. They even give cover customer in court for accusation data would be stolen. https://www.ibm.com/granite

Bart Janssens 🇧🇪 @bart@sociabl.be · last month

Content Warning

@timnitGebru Eugh, that’s disappointing, but it was to be expected of course. MSM completely misrepresented this again.

David G. Gándara @mr_rookes@mastodon.gal · last month

Content Warning

@timnitGebru What about the euporean ALIA. I couldn't figure out if it was wholly trained with licensed stuff. Some people claim they did.

Michael Slade @michaelslade@mastodon.cloud · last month

Content Warning

@timnitGebru One reason some prefer DeepSeek is that it steals more data without regret and may be therefore more robust.

nonlinear @nonlinear@social.praxis.nyc · last month

Content Warning

@timnitGebru Openwashing is everywhere. If we don’t protect our concepts, they’ll be taken away from us for profit.

PhDog @dogfox@kpop.social · last month

Content Warning

They didn't publish the code? Lame.

And sus.

@timnitGebru

A 𑐴𑐶𑐟𑐶 (hiti) is an ancient Newah urban water feature: a stepped, sunken courtyard with waterspouts and a central pool. These are fed through a remarkable network of pipes, reservoirs, and irrigation systems drawing water from mountain springs dozens of kilometres away. Every day, hundreds of people from the neighbourhood come to draw water, bathe, and exchange news. They are a unique and precious feature of Newa urban architecture through which neighbourhoods are created.

𑐴𑐶𑐟𑐶: About · Code of conduct · Privacy ·

Bonfire · 0.9.10-beta.134 no JS en

Automatic federation enabled