G.Vanderstichele (@GStichele) profile picture

Judge Crt of Appeal Ghent (Be) D.Phil in Law @OxfordLawFac - Aff. res. KUL CITIP + COHUBICOL. Legal Tech + Adjudication. RTnotEndorsement

Oxford, England
Joined February 2014
Photo by G.Vanderstichele (@GStichele)Photo by G.Vanderstichele (@GStichele)Photo by G.Vanderstichele (@GStichele)Photo by G.Vanderstichele (@GStichele)Photo by G.Vanderstichele (@GStichele)Photo by G.Vanderstichele (@GStichele)
G.Vanderstichele retweeted
More output of the COHUBICOL project: AI AND THE FUTURE OF LEGAL METHOD, a Dutch article in the Journal of the Judiciary (honoured to be part of the special issue - it will be key for the judiciary to engage with this) rechtspraak.nl/SiteCollectio…
2
4
371
G.Vanderstichele retweeted
How do judges and courts deploy AI? Interesting overview: UNESCO Global Judges’ Initiative. A Survey on the Use of AI Systems by Judicial Operators unesdoc.unesco.org/ark:/4822… 1/
1
8
12
656
Extremely good paper on AI policy. Questions the unexamined assumption that the amount of pretraining compute predicts the safety risks of a model. Five key points: 1. The relationship between compute and risk is rapidly changing. 2. Evidence to-date suggests we are not good at predicting what abilities emerge at different scales. 3. FLOP has to be better specified as a metric to be meaningful and current specifications are subject to Goodhart's law. 4. Governments should be transparent about what risks they are concerned about and where they are allocating limited resources. 5. Applying hard coded thresholds to a quickly changing distribution is likely to fail. Conclusion: "compute thresholds as currently implemented are shortsighted and likely to fail to mitigate risk."
Does more compute equate with greater risk? What is our track record at predicting what risks emerge with scale? I don't have much time anymore to write, but this felt important enough to write down some thoughts on. arxiv.org/pdf/2407.05694v1
18
2
86
14,877
Stanford study finds AI legal research tools prone to hallucinations by @bendee983 in @VentureBeat: “One of the most important arguments we make in the paper is that we have an urgent need for transparency and benchmarking in legal AI,” Dan Ho said. venturebeat.com/ai/stanford-…
2
5
2
15
4,581
'Die houding van media en politici is illustratief voor het gebrek aan een doorleefde democratische cultuur in ons land. Dat is een probleem, want op de lange termijn is geen enkele rechtsstatelijke instantie tegen zo’n apathie opgewassen.'
Quinten Jacobs | Het zijn maar verkiezingen tijd.be/opinie/column/column…
54
Quinten Jacobs | Het zijn maar verkiezingen tijd.be/opinie/column/column…
1
103
G.Vanderstichele retweeted
Which dimension reduction method should you use for data visualization? Check out our blog post on this topic! sites.duke.edu/dimensionredu…
3
27
5
108
21,536
G.Vanderstichele retweeted
Blog post with additional context about our new paper "AI Agents That Matter": - What does the term agent mean? Is it just a buzzword? - Do agents even work? - How does the poor reproducibility of agent evaluations compare with other areas of ML? aisnakeoil.com/p/new-paper-a…
20
4
53
18,521
G.Vanderstichele retweeted
Rule of law fellowship at Stanford, including comparative law, law and development. careersearch.stanford.edu/jo…
21
1
44
6,151
Will Europe succeed in building truly public digital infrastructure for AI, or will its investments entrench Big Tech’s dominance? These and other questions will guide AI Now’s work shaping the EU’s industrial policy for AI; read more here: ainowinstitute.org/general/p…
2
11
5
41
8,804
Atos crisis deepens as biggest shareholder ditches rescue plan ft.com/content/9c3667d7-1c79… via @ft
35
G.Vanderstichele retweeted
Granted - many things are broken in how we think about power in analyzing capitalist dynamics. Even most perceptive critics don't know what to do with the military or intelligence agencies (let alone entities like the State Department) in explaining the rise of Silicon Valley.
2
3
54
6,192
G.Vanderstichele retweeted
Join the Law Tutorial for CS folk this Friday from 12-14 on the implications of the AI Act for the development of human-centric generative AI, given by the inimitable @Gianmarco_Gori and myself. In the context of the @humaneainet
#HumaneAIEvent Online tutorial on "The AI Act’s relevance for the use of Generative AI in Human-Centric AI" - Fri June 28 2024 (12:00 - 14:00 CET) 👩‍🏫 By @mireillemoret and Dr. gianmarco_gori ✒ Registration: email to Bert.Frans.P.De[email protected] (DL: 27 noon CEST)
2
5
478
This is a nice paper. It is an elaborate and somewhat convoluted way to show that AI safety is not a model property. Here's a much simpler argument for why that's the case aisnakeoil.com/p/ai-safety-i… It's great to see that this view of AI safety, already widely recognized outside the AI safety community, is finally being recognized in the AI safety community.
Model developers try to train “safe” models that refuse to help with malicious tasks like hacking ...but in new work with @JacobSteinhardt and @ancadianadragan, we show that such models still enable misuse: adversaries can combine multiple safe models to bypass safeguards 1/n
3
18
1
73
20,243
'We find that the algorithm has significant methodological flaws, suffers from information leakage, relies on inappropriate proxy values for maltreatment assessment, generates inconsistent risk scores, and exhibits age-based discrimination.' dl.acm.org/doi/abs/10.1145/3…
72
G.Vanderstichele retweeted
Paper alert: Happy to share a new preprint on the #AIAct, authored with #KarenYeung. We discuss the #AI Act’s history & structure, and #critically assess its (inadequate) enforcement architecture, incl the problematic reliance on standards. Now on SSRN: papers.ssrn.com/sol3/papers.…
5
40
2
136
12,182
G.Vanderstichele retweeted
With the cover all set, I’m delighted to announce my upcoming #book😊 In #Algorithmic Rule by Law, I argue that the uncritical use of algorithmic regulation in public services can erode the rule of law & hurt democracy. Available this Autumn (with CUP): cambridge.org/core/books/alg…
12
49
5
239
21,292
👇
This quoted post is unavailable.
66