Профиль

"AI" means shoddy software that hasn't been audited, but doesn't need to deliver on accuracy claims to effect concentration of power. RFPs for physical infrastructure always require compliance with codes and quality/reliability. The "AI" RFPs never do, and contracts disclaim fitness for any purpose

Google's behavior is indistinguishable from a comicbook villain intent on destruction of shared reality. SynthID doesn't stop the spread of disinfo, is not even reliable, and at-best sets up an arms race. A bunch of websites claim to remove SynthID. It also gives permission to disregard real images

RAVEN: Erasing Invisible Watermarks via Novel View SynthesisInvisible watermarking has become a critical mechanism for authenticating AI-generated image content, with major platforms deploying watermarking schemes at scale. However, evaluating the vulnerabilit...arxiv.org

Bargaining gets results! Not only is this a serious labor issue, this agreement is saving the reputation and epistemic integrity of ProPublica from the worst impulses of management. This dynamic applies to so many institutions, and the clock is ticking to organize and win bargaining agreements.

Flailing at manufacturing social permission.

The reputation-laundering operation known as AlphaFold has been shut down. It changed culture in damaging ways, and similar protein-structure packages carry the mantle of black-box short-cuts that break the "chain of custody of why", offloading labor to often-haphazard and de-emphasized validation.

Google shuts down its Nobel-prize winning AlphaFold project as it focuses on Gemini - EngadgetDeepMind's award-winning AlphaFold team is no more.www.engadget.com

While faculty and students experience nausea at betrayals like this, administrators at public and private universities across the country are experiencing an insatiable urge to do the same. It is only through broad-base solidarity and organizing that we can mitigate these destructive impulses.

Fantastic not-so-quick thread on critical theory in information retrieval, including avoiding ethics-washing, algorithmic bothsidesism, and other forms of epistemic violence, as well as the need for not shying away from the "political" and a call for praxis.

Tag yourself. I'm MOAR TORT LAW!!!!

Strict liability is the correct framework because these "mitigations" are inherently a scam. Musk knows they don't work and he has no interest in them working. He wants to advertise "spicy mode" (wink, wink) to customers paying for NCII while pointing the gov to "guardrails" made of damp cardboard.

New investigation from AI Forensics finds HuggingFace Spaces (which are subsidized) for "Image Editing" are primarily used to generate non-consensual intimate imagery of non-celeb real people, including CSAM, and lack even cursory "guardrails" (which are a scam). aiforensics.org/uploads/NCII...

Debian is considering a resolution with two choices. One would forbid use of LLMs within the Debian project (does not apply to upstream packages). The other is permissive responsible-washing. I think it is not incidental that rationale for the former has citations, while the latter does not.

Thakur v Trump continues to deliver the goods. Recall that it was brought by faculty (not the University of California) and is the same case that got UCLA funding restored last year, before UC could sign a betrayal agreement with the Trump administration.