Stella Biderman

13.3K posts

Stella Biderman

Stella Biderman

@BlancheMinerva

Ensuring that tech companies don't have a monopoly on being able to do research on cutting edge AI @AiEleuther. She/her

Katılım Mayıs 2019
841 Takip Edilen17.7K Takipçiler
Sabitlenmiş Tweet
Stella Biderman
Stella Biderman@BlancheMinerva·
The actual American thing to do is not report people to a government that regularly breaks the law, ignores people's rights, and sends them to torture and death camps. You have a moral obligation to hide people from ICE just like you did to hide them from the Gestapo.
Stella Biderman tweet media
English
9
7
106
22.2K
Stella Biderman
Stella Biderman@BlancheMinerva·
How do you identify which problems are interesting and valuable? When people don’t work on problems that matter, why do you think that is?
English
1
2
24
2.8K
Maiv
Maiv@Viam_Invenias_0·
@xlr8harder Oops, does it even matter now? The incentive is just not there anymore for the investors Relying on the goodwill of a few labs is just delusional Opensource llm in the west is nearly dead now
English
2
0
0
422
xlr8harder
xlr8harder@xlr8harder·
A tragedy is unfolding at AllenAI. Fs in the chat
xlr8harder tweet media
English
29
53
659
171.7K
Stella Biderman
Stella Biderman@BlancheMinerva·
@mengyer Extreme over compliance is always the legally safest route. Obviously the more conservative you are the safer you are. That doesn’t mean that they don’t deserve moral condemnation.
English
1
0
1
227
Mengye Ren
Mengye Ren@mengyer·
@BlancheMinerva I am not a legal expert here, but my interaction with AI told me that the 2004 carve out still doesn’t exempt government backed personnels, and the NeurIPS approach is legally the safest approach. NeurIPS has their legal counsel and it is probably not a light decision.
English
2
0
0
758
xlr8harder
xlr8harder@xlr8harder·
@StefanABaumann @MichalBrzozows2 @kirillzubovsky Yeah, great work, but the main Pythia line was released in 2023. There are more recent releases, but the scale is small and it's not complete end-to-end model training. AllenAI really stands alone here.
English
1
0
15
926
Delip Rao e/σ
Delip Rao e/σ@deliprao·
After chatting with a bunch of people I am convinced this @NeurIPSConf "sanctions" policy is BS. The NeurIPS board owe it to the research community that has served them for decades with transparency by reporting why they took this step, whether this push came from the current administration, and if so, from whom? Otherwise, NeurIPS (and other US-based conference organizing bodies) will lose credibility and trust with the broader international community operating in and outside of the US.
Delip Rao e/σ@deliprao

From my understanding, powerhouses like Tsinghua, Peking, SJT, and Zhejiang that produce most of the Neurips/ICML works from China are not impacted by this. Trying to understand how bad this is (IMO this is terrible for open science). Can mainland followers confirm?

English
5
10
107
18.4K
(((ل()(ل() 'yoav))))👾
(((ل()(ل() 'yoav))))👾@yoavgo·
what do you consider to be BS exactly, and what do you think they can do differently? the most serious response I saw was that ACM has a different policy, but I could not verify that this is the case, as the policy page I found just ignored this particular case (but does adhere to other US sanctions).
English
3
0
1
1.9K
Stella Biderman
Stella Biderman@BlancheMinerva·
@mengyer People elsewhere in this thread have linked to the OFAC 2004 thing, but it’s also the case that each company on the list has something specific that’s restricted. For example, Huawei is only restricted in the context of securities trading: ofac.treasury.gov/media/99111/do…
English
0
0
0
116
Stella Biderman
Stella Biderman@BlancheMinerva·
@mengyer NeurIPS is misrepresenting the law. It seems pretty clear to me that this is either preemptive compliance or someone in the admin leaned on them.
English
2
0
7
1.1K
Kirill Zubovsky
Kirill Zubovsky@kirillzubovsky·
@xlr8harder Did not know that. Was under impression it's all moved to China. I wonder if others don't know that either.
English
2
0
1
1.3K
Manuel Mager (Turatemai)
Manuel Mager (Turatemai)@pywirrarika·
@ToadyBabirusa @miniapeur Interesting point. I have not read the list (my fault). But tbh, also a lot of companies/orgs that are allowed to publish could be a accused of war crimes. Not easy to try to be a moral judge ATM.
English
1
0
0
286
Mathieu
Mathieu@miniapeur·
NeurIPS 😬
Mathieu tweet media
English
10
12
136
137.9K
Kia Rahmani
Kia Rahmani@Kiarahmani_·
@pywirrarika @miniapeur If you were alive in the 1940s, would you publish a paper with a Nazi scientist at a Nazi institution?
English
1
0
0
253
Stella Biderman
Stella Biderman@BlancheMinerva·
Well, the good news is that with ICML providing acceptence at the end of April this can't get any earlier...
NeurIPS Conference@NeurIPSConf

The NeurIPS 2026 Call for Papers is now live: neurips.cc/Conferences/20… Abstracts are due May 4, 2026 (AOE), with full papers due May 6, 2026 AOE. Please review the key changes to submissions this year neurips.cc/Conferences/20…, as well as our new initiative for Strengthening Area Chair Engagement and Calibration at NeurIPS 2026 blog.neurips.cc/2026/03/23/ref…

English
0
0
25
8.8K
Remco
Remco@recmo·
@littmath @monofrogue Weren’t a lot of them found to be solved already through AI literature search? That qualifies as “being marked solved”.
English
2
0
1
868
Daniel Litt
Daniel Litt@littmath·
I’m not sure how the number 50 was arrived at here. I count 3 Erdos problems solved fully autonomously with no known prior solution, and 6 solved by humans in “collaboration” with AI tools (unclear to me what exact role was played by AI in these cases). Am I missing something?
Dwarkesh Patel@dwarkesh_sp

AI has solved 50 Erdős problems in the last year. But on a wider sweep of problems, the models’ success rate is only about 1-2%: labs have just been publishing the wins. This isn’t because AI isn’t useful for mathematicians. Terence Tao thinks the models are currently at the level of a trustworthy coworker. But while they’ve got a strong ability to apply standard math techniques to problems, often more reliably than humans, Terence thinks they currently aren’t great at iterating on partial successes - their understanding of the mathematical object does not advance from session to session. I swear I wasn’t trying to get him to talk about continual learning.

English
18
15
276
43.1K
Igor Shilov
Igor Shilov@_igorshilov·
@BlancheMinerva There is also a tricky bit of what does it even mean to "have never seen the data in the first place". For privacy unlearning it's easy, but for capability not really. For WMDP we can't just remove the forget set, it's clearly not enough. And all biology is probably too much?
English
1
0
0
17
Stella Biderman
Stella Biderman@BlancheMinerva·
If I was going to claim that a finetuning methodology for machine unlearning “really worked,” what evidence would you like to see?
English
14
1
31
8.9K
Stella Biderman
Stella Biderman@BlancheMinerva·
@_igorshilov I really like your paper by the way! Looking forward to citing it regularly over the next year or so.
English
1
0
0
31
Stella Biderman
Stella Biderman@BlancheMinerva·
@_igorshilov But while that question is interesting, its somewhat beside the point for a paper focused on machine unlearning. You're absolutely right about the trajectories. Although there's a variant of the question that still matters: how much work should I do to try to falsify the results?
English
1
0
0
24
Stella Biderman
Stella Biderman@BlancheMinerva·
@megamor2 As there is always more labor one could do to try to elicit the undesirable capabilities from the unlearned model. Also, there's another great recent paper that released models for this work with a PII bent: arxiv.org/abs/2510.19811
English
1
0
1
78
Stella Biderman
Stella Biderman@BlancheMinerva·
@megamor2 We have trained and released models that enable you to do this analysis for real (in one context)! I was thinking more along the lines of "what analysis of the trio {unlearned, filtered, unfiltered} models should one do?" arxiv.org/abs/2508.06601
English
1
0
1
120
Stella Biderman
Stella Biderman@BlancheMinerva·
@anirudhg9119 This doesn’t seem particularly relevant to any real-world unlearning work? Why do you think it is?
English
0
0
0
128

Keşfet

@Viam_Invenias_0 @xlr8harder @mengyer @StefanABaumann @MichalBrzozows2 @kirillzubovsky @yoavgo @deliprao