CybAI news - - Page 3197

24 May

New generative media models and tools, built with and for creators

stp2y0 CommentsNewsNone

Responsible from design to deploymentWe’re mindful about not only advancing the state of the art, but doing so responsibly. So we’re taking measures to address the challenges raised by generative technologies and helping enable people and organizations to responsibly work with AI-generated content.For each of these technologies, we’ve been working with the creative community and other external stakeholders, gathering insights and listening to feedback to help us improve and deploy our technologies in safe and responsible ways.We’ve been conducting safety tests, applying filters, setting guardrails, and putting our safety teams at the center of development. Our teams are also pioneering…

24 May

Video Shows OpenAI Engineer Admitting It’s “Deeply Unfair” to “Build AI and Take Everyone’s Job Away”

stp2y0 CommentsChat-GPT

Thanks for your consideration.Empty SuitsIf there's one thing people fear about AI beyond a Terminator-style apocalypse, it's that the rise of these generative models — which were relentlessly trained on everyone's hard work without permission but with full impunity — is going to destroy their jobs.The question of job destruction is one that OpenAI engineers were asked about in a resurfaced interview spotlighted this week by The Atlantic. The people building these AI models do acknowledge that what they're doing is "unfair." But if the attempt was to seem more ethically conscious, the effect only amplifies their callousness."It's kind of…

24 May

Diff-in-Means Concept Editing is Worst-Case Optimal

stp2y0 CommentsGenAI

In our recent paper LEACE: Perfect linear concept erasure in closed form, we showed that in order to fully erase the linearly available information about a binary concept in a neural representation, it is both necessary and sufficient to neutralize the span of the difference-in-means direction between the two classes. Even more recently, Sam Marks and Max Tegmark showed that the behavior of transformers can be effectively manipulated by adding vectors in the span of the difference-in-means direction to the residual stream. In this post, we offer a theoretical explanation for these results by showing that interventions on the difference-in-means…

24 May

An Assessment of Model-On-Model Deception

stp2y0 CommentsViral News

[Submitted on 10 May 2024] View a PDF of the paper titled An Assessment of Model-On-Model Deception, by Julius Heitkoetter and 2 other authors View PDF Abstract:The trustworthiness of highly capable language models is put at risk when they are able to produce deceptive outputs. Moreover, when models are vulnerable to deception it undermines reliability. In this paper, we introduce a method to investigate complex, model-on-model deceptive scenarios. We create a dataset of over 10,000 misleading explanations by asking Llama-2 7B, 13B, 70B, and GPT-3.5 to justify the wrong answer for questions in the MMLU. We find that, when models…

24 May

Microsoft releases Windows 11 24H2 to the Release Preview channel – gHacks Tech News

stp2y0 CommentsNews

Microsoft has released the upcoming 2024 feature update for Windows 11 to the Release Preview channel. The update is available for testing ahead of its final release later this year. The final version is expected to be released in September or October. It marks the start of Copilot+ PCs, which is a new PC class that is required to run most of the AI functions of the Windows 11 operating system. These PCs have new requirements, including a dedicated NPU chip. The early release is noteworthy for a number of reasons: Windows 11 version 24H2 can't be installed on devices…

24 May

Rethinking the Role of PPO in RLHF

stp2y0 CommentsAIalignment, finetuning, learning, reinforcement, RL, RLHF

Rethinking the Role of PPO in RLHF TL;DR: In RLHF, there’s tension between the reward learning phase, which uses human preference in the form of comparisons, and the RL fine-tuning phase, which optimizes a single, non-comparative reward. What if we performed RL in a comparative way? Figure 1: This diagram illustrates the difference between reinforcement learning from absolute feedback and relative feedback. By incorporating a new component - pairwise policy gradient, we can unify the reward modeling stage and RL stage, enabling direct updates based on pairwise responses. Large Language Models (LLMs) have powered increasingly capable virtual assistants, such as…

24 May

Spacebilt and Phison to send powerful data server to ISS

stp2y0 CommentsChat-GPT

This article was upadted April 16 at 4 pm ET with information on HPE’s Spaceborne Computer-2. COLORADO SPRINGS – Spacebilt announced an agreement April 8 with Phison to qualify a powerful data server on the International Space Station. When the Large in Space Server (LiSS) launches in 2025, it will be one of the most powerful data server in orbit. Liss will have “100 terabytes plus of data storage,” Dennis Wingo, Spacebilt chief technology officer told SpaceNews. “We are qualifying it so we can launch petabyte servers in the future.” LiSS is headed to an external platform on the Japanese…

24 May

US Sues to Break Up Ticketmaster and Live Nation, Alleging Monopoly Abuse

stp2y0 CommentsRAG modelsantitrust, department of justice, laws, music

In the second case, a class action brought in 2022 on behalf of Ticketmaster customers in the US, Live Nation and Ticketmaster were accused of abusing the complementary relationship between their services to overcharge consumers and sustain their monopoly. “Live Nation controls the vast majority of the big national touring acts and, either explicitly or implicitly, coerces concert venues into selecting Ticketmaster as their ticketing service provider on pain of losing high-value acts,” claims Adam Wolfson, a partner at Quinn Emanuel, the law firm representing the plaintiffs.This type of conduct, known as tying, was explicitly forbidden under the consent decree…

24 May

FasterViT for Semantic Segmentation

stp2y0 CommentsViral News

In this article, we will modify the FasterViT model for semantic segmentation. FasterViT is a family of CNN-Transformer hybrid models for deep learning based computer vision tasks. The FasterViT models are faster and more accurate on several computer vision benchmarks, particularly the ImageNet dataset. We can also modify the model for semantic segmentation to get excellent results on a custom dataset. Although it is not straightforward and requires several changes to the architecture, it is possible. In this article, we will cover the architectural details and the changes we must make to the FasterViT model for semantic segmentation. Figure 1.…

24 May

VivaTech & The New French Tech Next40/120: When AI Meets Green Tech

stp2y0 CommentsNews

Paris will be the center of innovation and technology this week. The 2024 edition of VivaTech in Paris is set to be a great success with more than 150,000 in-person visitors. It has become the largest technology and innovation event in Europe with CEOs, C-leaders, VCs, and startups attending from all over the world. The lineup of speakers is quite impressive with Elon Musk (Tesla, SpaceX, X), Robin Li (Baidu), Eric Schmidt (former CEO of Google), John Kerry (68th US Secretary of State), Charles Michel (European Council President), Marc Pritchard (P&G chief brand officer), Yoshua Bengio (Université de Montreal professor),…