AI Agent Benchmarks Need to Measure User #intent - IEEE Spectrum
▻https://spectrum.ieee.org/ai-agent-benchmark
— Permalink
#LLMs #bruceschneier #artificialintelligence #generativeai #geniecoefficient #misunderstanding #betrayal #gap #agents #alignment #safety
Protecting our #FLOSS commons from LLMs — Codeberg News
▻https://blog.codeberg.org/protecting-our-floss-commons-from-llms.html
In Brief:
– Two motions regarding “#artificial_intelligence” and Large Language Models (#LLMs) were voted on among Codeberg e. V. members and passed.
– We are promising to not use any of your data to train #LLM and explain what the planned Terms of Use change mean for ’#vibe-coded' projects.
– We believe that LLMs endanger the free/libre software #ecosystem as a whole.
The #Codeberg e. V. annual assembly is the meeting that puts power into the hand of our active members. Proposals are discussed live, and later voted on asynchronously.
Since Large Language Models (LLMs) are an emerging but controversial technology, it is not surprising that two of the votes were concerned with Codeberg’s position about this technology. The 14-day voting period ended yesterday and both proposals were accepted.
#exploitgym: Can AI #agents Turn Security Vulnerabilities into Real Attacks?
▻https://www.cybergym.io/exploitgym
— Permalink
#cybersecurity #LLMs #artificialintelligence #vulnerability #exploits
The European #parliament’s answer to its AI worries: More AI – POLITICO
▻https://www.politico.eu/article/the-european-parliaments-answer-to-its-ai-worries-more-ai
— Permalink
#europe #artificialintelligence #LLMs #chatbot #law #draft #amendments #hallucinations #EPGenAIHub
Anne #alombert : intelligence ou bêtise artificielle ? - Idées - RFI
▻https://www.rfi.fr/fr/podcasts/id%C3%A9es/20250921-anne-alombert-intelligence-ou-b%C3%AAtise-artificielle?GJlXzGgqHj
— Permalink
#intelligenceartificielle #technique #stiegler #automatisation #cognition #dépendance #pharmakon #mémoire #expression #LLMs #apprentissage
Human Emacs
▻https://human-emacs.org
(...)
Not Under Discussion
We are not here to discuss whether #LLMs are effective at what they are claimed to be able to do; their effectiveness is not at all relevant to the question of whether their use can be part of a principled software movement dedicated to user #empowerment.
We are not here to discuss “open weight” #models; these are still built on a foundation of companies destructively mining the web from #data_centers that wreck communities. Such models cannot exist without exploitation. When a model can be fully trained by end users using data that was collected with consent, then we can talk about that, but right now that is nothing but science fiction and speculation.
We are not here to discuss how bad-faith contributors can lie about the provenance of their patches. This risk is not new; bad-faith contributors have always been able to lie about the licensing and copyright implications of a patch. It is enough to treat #LLM-generated patches the same as other forms of #plagiarism.
(...)
CEO Pleads With #AI Industry to Stop Charging So Much to Replace #Human_Labor
▻https://futurism.com/future-society/palo-alto-ceo-ai-arora-automation-labor
With each passing day without an AI labor revolution, the tech industry’s pricing scheme for AI is becoming more and more disconnected from #reality — so much so that even its biggest clients are starting to revolt.
Speaking during an interview on CNBC‘s “Squawk on the Street” segment earlier this week, CEO of cybersecurity giant #Palo_Alto_Networks #Nikesh_Arora implored the tech industry to lower the cost of AI.
During the segment, the chief executive argued that the cost to use large language models (#LLMs) has to drop by 20 percent by 2027 — and 90 percent by 2028 — for the tech to be useful to enterprises.
“We need to see the pricing for AI come down,” Arora said.
Il a rien compris Arora. Il est pas invité à la Maison Blanche. C’est une sorte d’indice.
Mais je comprends pas, ils demandent à ce que ça baisse pour que ça soit rentable pour les entreprises utilisatrices, sinon autant prendre des humains… mais dans le même temps ça fait plusieurs années qu’on dit qu’aucun des acteurs (même les « winners take all » au top) ne fait aucun profit, qu’ils sont déficitaires, donc forcément au bout d’un moment ils finissent par augmenter les coûts d’utilisation c’est normal.
(ChatGPT change toutes les semaines pour ça, et me harcèle de notifs et d’emails pour que je paye, et d’ailleurs maintenant je ne vois plus aucune option dans l’interface gratuite, ya plus que le niveau le plus bas)
Benchmarking Coding Agents on Databricks’ Multi-Million Line Codebase | Databricks Blog
▻https://www.databricks.com/blog/benchmarking-coding-agents-databricks-multi-million-line-codebase
— Permalink
The #hitchhiker's Guide to Agentic AI: From Foundations to Systems
▻https://arxiv.org/abs/2606.24937
— Permalink
#artificialintelligence #agents #LLMs #design #development #architecture #production #taxonomy
What Emily Bender Really Meant by "Stochastic #parrots"; - IEEE Spectrum
▻https://spectrum.ieee.org/stochastic-parrot
— Permalink
#artificialintelligence #linguistics #language #LLMs #anthropomorphism #stochastic
LinguaCelta: The Community is the Achievement; the Achievement is the Community
▻https://linguacelta.com/blog/2026/05/LLMs.html
The Community is the Achievement; the Achievement is the Community
18 May 2026
An ethical love-letter to distributed technology communities.
Talking to techies
This essay is explicitly addressed to my fellow technologists: #software #developers, hobby #coders, digital #humanists, #computer #science theorists, and all the other members of this big family of people who do tech. That doesn’t mean that what I write here can’t be of interest to anyone else (I’ll be very flattered if it’s of interest to anybody, tbh!). But the argument I’ll make is based on the idea that you and I share a community.
I want to talk about our community, and why it’s important. I want to suggest that using #LLMs to generate content to be included in technology projects, whether that’s code or text or images, or code reviews or proofreading, harms our shared community.
Failing grades soar as professors see greater AI usage, dwindling math skills in UC Berkeley computer science classes | Academics | dailycal.org
▻https://www.dailycal.org/news/campus/academics/failing-grades-soar-as-professors-see-greater-ai-usage-dwindling-math-skills-in-uc-berkeley/article_16fad0bf-02cb-4b8c-8d88-888ffd9f8608.html
Interesting discussion @ HN.
— Permalink
#learning #education #generativeai #LLMs #cognition #grade #maths #computerscience
#arXiv imposes one-year ban for unchecked LLM output
arXiv has clarified enforcement for submissions that include unverified large language model (LLM) output. Per a public thread by Thomas G. Dietterich, chair of arXiv’s computer science moderators, a submission that contains “incontrovertible evidence that the authors did not check the results of LLM generation”, examples cited include hallucinated references and leftover LLM meta-comments, can trigger a one-year ban and a requirement that future arXiv submissions first be accepted at a reputable peer-reviewed venue, as reported by The Verge and other outlets. The platform’s Code of Conduct was cited in Dietterich’s thread, noting authors bear responsibility for content irrespective of how it was generated. Community reaction has been mixed, with some researchers supporting the move and others warning about selective enforcement and false-positive risks, according to The Decoder and social reporting aggregated by di.gg.
▻https://letsdatascience.com/news/arxiv-imposes-one-year-ban-for-unchecked-llm-output-be07fdf4
#LLMs #IA #intelligence_artificielle #bannissement #recherche #édition_scientifique #publications #AI #1_an #revues
Les #LLM dégradent profondément vos documents
On avait déjà attiré l’attention sur les grandes #limites de l’IA pour produire des #synthèses (▻https://danslesalgorithmes.net/stream/manipuler-la-synthese-de-document), au risque de valoriser certaines informations sur d’autres. Microsoft Research vient de publier un article et un benchmark : ils ont proposé 52 documents à 19 modèles concurrents dont les modèles de pointe. Chaque modèle reçoit un document et une série d’instructions de modification, une vingtaine maximum. A la fin de la série d’instruction, 25% à 50% du contenu se retrouve dégradé (analyse faite en comparant simplement les contenus des documents). Si on étend encore les instructions à une centaine, la courbe de la dégradation continue. Les chercheurs parlent de « #déclin_monotone ». Ils constatent également que les performances à court terme ne prédisent pas la fiabilité à long terme. « Deux modèles qui semblaient presque identiques après deux interactions (91,5 % contre 91,1 %) ont divergé de façon spectaculaire au fil du temps (48,3 % contre 64,1 %). » La dégradation est parfois brutale. Sur les 50 domaines d’activité testés auxquels ils ont confié des documents pour des tâches d’#édition, le seul domaine qui fonctionne reste le code #python : la dégradation est très faible, les contenus restent stables sur 17 des 19 modèles, à 98%. Mais par exemple, sur un tableur, seulement 50% des lignes d’origine sont encore présentes après 20 itérations : l’ordre des lignes, les noms de colonnes et le nombre de lignes sont plus dégradées encore.
La démonstration permet de montrer que l’usage des LLM risquent de n’être fonctionnels que pour le code, car le code à une définition mécanique, c’est-à-dire qu’il existe une spécification lisible par machine permettant de vérifier la sortie (mais la démo montre que quand il s’agit de structure des bases de données SQL, les résultats sont déjà beaucoup moins parfaits). « Partout où la correction exige de la #compréhension, les modèles s’effondrent », explique le développeur norvégien Christian Ekrem sur son blog. Pire : la #corruption est invisible par conception, « silencieuse ».
« Le plus inquiétant, c’est la manière dont ces #erreurs se produisent. Elles sont rares, mais graves. Le modèle ne transforme pas votre document en charabia. Il y apporte de petites #modifications, assurées (!), qui paraissent inoffensives au premier coup d’œil. Un détail déplacé. Une précision omise. Un sens subtilement altéré. Une phrase réorganisée pour en modifier l’ordre. Il faudrait lire attentivement l’ensemble du document, en le comparant à l’original, pour s’en apercevoir. Et personne ne le fait. »
Dans vos slides, votre « environ 30% » va devenir « 30% »… puis « 20% ». Dans un contrat, « sous certaines conditions » s’efface. « Après signature » devient « avant signature »… Non seulement les erreurs s’accumulent, mais elles interagissent : « une corruption précoce modifie le contexte, ce qui décale les résultats suivants, et ainsi de suite ». Vous continuez à travailler sur une version corrompue qui ne dit déjà plus ce que vous pensiez. « Si cela ne vous terrifie pas, je doute que ayez déjà travaillé avec des documents importants. (…) Lorsque vous déléguez la maintenance documentaire à un LLM, la théorie meurt doublement. Premièrement : vous n’avez pas acquis la compréhension, car vous avez délégué au lieu de vous impliquer directement dans le sujet. Deuxièmement : le LLM a insidieusement altéré le document lui-même. Vous vous retrouvez donc sans modèle mental ni représentation écrite précise. Vous avez, pour ainsi dire, perdu à la fois la carte et le territoire. »
▻https://danslesalgorithmes.net/stream/les-llm-degradent-profondement-vos-documents
#qualité #dégradation #LLMs #IA #AI #intelligence_artificielle
LLMs Corrupt Your Documents When You Delegate
Large Language Models (LLMs) are poised to disrupt knowledge work, with the emergence of delegated work as a new interaction paradigm (e.g., vibe coding). Delegation requires trust - the expectation that the LLM will faithfully execute the task without introducing errors into documents. We introduce DELEGATE-52 to study the readiness of AI systems in delegated workflows. DELEGATE-52 simulates long delegated workflows that require in-depth document editing across 52 professional domains, such as coding, crystallography, and music notation. Our large-scale experiment with 19 LLMs reveals that current models degrade documents during delegation: even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt an average of 25% of document content by the end of long workflows, with other models failing more severely. Additional experiments reveal that agentic tool use does not improve performance on DELEGATE-52, and that degradation severity is exacerbated by document size, length of interaction, or presence of distractor files. Our analysis shows that current LLMs are unreliable delegates: they introduce sparse but severe errors that silently corrupt documents, compounding over long interaction.
Let’s talk about #LLMs
▻https://www.b-list.org/weblog/2026/apr/09/llms
Everybody seems to agree we’re in the middle of something, though what, exactly, seems to be up for debate. It might be an unprecedented #revolution in #productivity and capabilities, perhaps even the precursor to a technological “#singularity” beyond which it’s impossible to guess what the world might look like. It might be just another vaporware hype cycle that will blow over. It might be a dot-com-style #bubble that will lead to a big crash but still leave us with something useful (the way the #dot-com bubble drove mass adoption of the web). It might be none of those things.
Many thousands of words have already been spent arguing variations of these positions. So of course today I’m going to throw a few thousand more words at it, because that’s what blogs are for. At least all the ones you’ll read here were written by me (and you can pry my em-dashes from my cold, dead hands).
Not astonishing, in fact methodically already known before AI came along as buzzword, but today much easier : “... They can match writing styles, interests, details to infer a job or city, or other unstructured information. ...”
via ▻https://diaspora.psyco.fr/p/12367205
♲ FineCoatMummy - 2026-03-28 20:02:37 GMT
Large-scale online deanonymization with #LLMs
Paper by,
Simon Lermen, Daniel Paleka, Joshua Swanson, Michael Aerni, Nicholas Carlini, Florian Tramèr
It talks about #deanonymizing those who writes under a #pseudonym. Sites like reddit, lemmy would be that type.
Large-scale online deanonymization with LLMs
Simon Lermen, Daniel Paleka, Joshua Swanson, Michael Aerni, Nicholas Carlini, Florian Tramèr
▻https://arxiv.org/abs/2602.16800
We show that large language models can be used to perform at-scale deanonymization. With full Internet access, our agent can re-identify Hacker News users and Anthropic Interviewer participants at high precision, given pseudonymous online profiles and conversations alone, matching what would take hours for a dedicated human investigator. We then design attacks for the closed-world setting. Given two databases of pseudonymous individuals, each containing unstructured text written by or about that individual, we implement a scalable attack pipeline that uses LLMs to: (1) extract identity-relevant features, (2) search for candidate matches via semantic embeddings, and (3) reason over top candidates to verify matches and reduce false positives. Compared to classical deanonymization work (e.g., on the Netflix prize) that required structured data, our approach works directly on raw user content across arbitrary platforms. We construct three datasets with known ground-truth data to evaluate our attacks. The first links Hacker News to LinkedIn profiles, using cross-platform references that appear in the profiles. Our second dataset matches users across Reddit movie discussion communities; and the third splits a single user’s Reddit history in time to create two pseudonymous profiles to be matched. In each setting, LLM-based methods substantially outperform classical baselines, achieving up to 68% recall at 90% precision compared to near 0% for the best non-LLM method. Our results show that the practical obscurity protecting pseudonymous users online no longer holds and that threat models for online privacy need to be reconsidered.
ploum - mastodon
▻https://mamot.fr/@ploum/116170413255573009
Using #LLMs for coding is so “good” that programmers are now faking using them so they can still write code by hand while making the management happy…
slow clap!
Subverting AI Agent Logging with a Git Post-Commit Hook – Dan Q
▻https://danq.me/2026/03/03/ai-agent-logging
Last night I was chatting to my friend (and fellow Three Rings volunteer) Ollie about our respective workplaces and their approach to AI-supported software engineering, and it echoed conversations I’ve had with other friends. Some workplaces, it seems, are leaning so-hard into AI-supported software development that they’re berating developers who seem to be using the tools less than their colleagues!
lundi.dev — Un #milliard par #jour pour vérifier l’heure
▻https://lundi.dev/newsletter/2026/02/09/2-un-milliard-par-jour-pour-verifier-l-heure
On ne connaît pas la proportion de ces agents qui tournent sur des #LLMs hébergés par les géants du #numérique, mais je ne peux pas m’empêcher de faire un lien avec le fait que #Google, #Microsoft, #Amazon et #Meta ont dépensé l’année dernière plus de 376 milliards de dollars en #investissements (#capital #expenditure, ou #CapEx).
Writing is thinking
On the value of human-generated scientific writing in the age of large-language models.
Writing scientific articles is an integral part of the scientific method and common practice to communicate research findings. However, writing is not only about reporting results; it also provides a tool to uncover new thoughts and ideas. Writing compels us to think — not in the chaotic, non-linear way our minds typically wander, but in a structured, intentional manner. By writing it down, we can sort years of research, data and analysis into an actual story, thereby identifying our main message and the influence of our work. This is not merely a philosophical observation; it is backed by scientific evidence. For example, handwriting can lead to widespread brain connectivity1 and has positive effects on learning and memory.
This is a call to continue recognizing the importance of human-generated scientific writing.
“This is a call to continue recognizing the importance of human-generated scientific writing”
This call may seem anachronistic in the age of large-language models (LLMs), which, with the right prompts, can create entire scientific articles2 (and peer-review reports3) in a few minutes, seemingly saving time and effort in getting results out once the hard research work is done. However, LLMs are not considered authors as they lack accountability, and thus, we would not consider publishing manuscripts written entirely by LLMs (using LLMs for copy-editing is allowed but should be declared). Importantly, if writing is thinking, are we not then reading the ‘thoughts’ of the LLM rather than those of the researchers behind the paper?
Current LLMs might also be wrong, a phenomenon called hallucination4. Therefore, LLM-generated text needs to be thoroughly checked and verified (including every reference as it might be made up5). It thus remains questionable how much time current LLMs really save. It might be more difficult and time-consuming to edit an LLM-generated text than to write an article or peer-review report from scratch, partly because one needs to understand the reasoning to be able to edit it. Some of these issues might be addressed by LLMs trained only on scientific databases, such as those outlined in a Review article by Fenglin Liu and team in this issue. Time will tell.
All that is not to say LLMs cannot serve as valuable tools in scientific writing. For example, LLMs can aid in improving readability and grammar, which might be particularly useful to those for which English is not their first language. LLMs might also be valuable for searching and summarizing diverse scientific literature6, and they can provide bullet points and assist in the brainstorming of ideas. In addition, LLMs can be beneficial in overcoming writer’s block, provide alternative explanations for findings or identify connections between seemingly unrelated subjects, thereby sparking new ideas.
Nevertheless, outsourcing the entire writing process to LLMs may deprive us of the opportunity to reflect on our field and engage in the creative, essential task of shaping research findings into a compelling narrative — a skill that is certainly important beyond scholarly writing and publishing.
►https://www.nature.com/articles/s44222-025-00323-4
#LLMs #IA #AI #intelligence_artificielle #pensée #écriture #apprentissage
Ads Are Coming To #chatgpt in the Coming Weeks
▻https://slashdot.org/story/26/01/16/1827203/ads-are-coming-to-chatgpt-in-the-coming-weeks
— Permalink
LM #council - Compare Frontier AI Systems Side-by-Side
▻https://lmcouncil.ai
— Permalink
A better way to limit Claude Code (and other coding agents!) access to Secrets – blog
▻https://patrickmccanna.net/a-better-way-to-limit-claude-code-and-other-coding-agents-access-to-
— Permalink
#dev #agent #copilot #claudecode #isolation #bubblewrap #LLMs #generativeai
GitHub - GeneploreAI/gibberifier: Stun #LLMs with random #unicode characters
▻https://github.com/GeneploreAI/gibberifier
— Permalink