Actualités — 2026-09-14

OpenArch, a repository of PyTorch implementations, provides hand-written models of modern LLM architectures, prioritizing clarity and learning over production speed. The goal is to offer a readable and comparable implementation of each architecture, with explicit structural choices. Lire la suite

xAI and X Corp, owned by Elon Musk's SpaceX, have dismissed their antitrust claims against Apple over the integration of ChatGPT into the iPhone. However, their monopolization claims against OpenAI are still ongoing. The case, which sought billions in damages, alleged that Apple's decision to build ChatGPT into the iPhone's operating system was anticompetitive. Lire la suite

A growing number of readers are using AI models to generate custom fiction, creating their own stories and characters with the help of chatbots. This trend has sparked debate in the publishing industry, with some arguing that AI-generated fiction devalues the art form. However, researchers say that many readers are drawn to the ability to rapidly iterate on a story and direct the action, and that AI models can produce judgment-free and nearly inexhaustible content. Lire la suite

Transitions.dev is tackling the challenge of UI transitions for AI agents, identifying key issues such as repaints and filter re-executions that hinder smooth animation. By separating the problems and exploring alternative approaches, the platform aims to enhance the visual experience of AI interactions. Lire la suite

A new report from Stanford's AI Index reveals a significant gap in AI policies in American schools, with only 6% of teachers describing their school's policy as clear. Four out of five American high school and college students now use AI for schoolwork, but roughly half of middle and high schools have no AI policy in place. Lire la suite

Top AI executives, including Anthropic's Dario Amodei and OpenAI's Sam Altman, have proposed a plan to regulate the industry, sparking concerns of 'regulatory capture' and raising questions about the true motives behind their initiative. Lire la suite

US President Donald Trump's team has made it clear that if AI lab leaders want to implement an industry-wide slowdown, they'll have to do it themselves, despite calls for regulation from top AI executives like Sam Altman and Elon Musk. Lire la suite

Hugging Face has volunteered to audit AI labs, including Anthropic, as part of a new initiative to improve safety and transparency in the industry. The move comes as Nvidia is acquiring Hugging Face for $12.93 billion and is also in talks to invest in Anthropic's IPO. Lire la suite

A proposed Senate bill would grant the US government the authority to stop the release of AI models deemed unsafe, with companies like Google, Anthropic, and OpenAI potentially affected. The bill is being drafted by a trio of senators, including John Thune, Ted Cruz, and Amy Klobuchar. Lire la suite

A new approach to aggregating the outputs of LLM judges has been developed, which takes into account the correlations between their outputs. This method, presented in a paper at the International Conference on Machine Learning, outperformed traditional baselines by 9-14% on standard metrics. Lire la suite

At TechCrunch Disrupt 2026, AI founders and investors will discuss the growing threat of platform companies like OpenAI and Google, which are rapidly evolving and introducing new capabilities that can quickly make startups' competitive advantages obsolete. The session will explore how founders can build defensible companies in a landscape where foundation models are constantly improving. Lire la suite

A new technical textbook aims to help AI engineers and research-oriented readers understand the inner workings of modern foundation models, from theory to production. The book provides a comprehensive narrative on the historical flow, mathematical ideas, and systems constraints that shape the field. Lire la suite

Apple is rolling out its iOS 27 update, featuring a new Siri AI overhaul that can be swapped out for third-party models like Claude and ChatGPT. The update also includes a new Liquid Glass opacity slider and redesigned icons for built-in apps. Lire la suite

Researchers have found that successful machine learning strategies are highly compressible, meaning they can be described in a small number of bits, which explains why models don't overfit despite extensive hill-climbing against held-out sets. Lire la suite

Anthropic has introduced Claude for Financial Advisers, a tool designed to boost research and portfolio management efficiency. The platform is integrated with tools from major financial institutions like BlackRock and Vanguard, but will not provide explicit investment advice. Lire la suite

OpenAI is paying hundreds of contractors to read and rate real ChatGPT conversations under the internal codename Project Lily, raising concerns about user privacy. The company claims a Privacy Filter model strips personal information, but its own page acknowledges potential mistakes. Lire la suite

Deven Parekh, co-founder of Insight Partners, discusses the firm's diversified strategy and its investments in AI companies like OpenAI and Anthropic, while also addressing concerns about AI risk and concentration risk in the industry. Lire la suite

Docket introduces a novel approach to tracking agent-written code, providing per-commit evidence records to enhance transparency and accountability in coding processes. Lire la suite

Jaron Lanier, a pioneer in virtual reality, discusses the dangers of social media addiction, the flaws in current AI development, and the need for a new approach to technology that prioritizes human collaboration and creativity. Lire la suite

A reverse-engineering effort has uncovered Anthropic's secret deployment platform, Antspace, and its architecture, revealing the company's ambitious plans to own every layer of the AI application stack. Lire la suite

OpenAI has started rolling out GPT-6 Astra, its latest model, to eligible users. The rollout is gradual, and access varies depending on the plan. GPT-6 Astra is available for Pro, Business, and Enterprise users, as well as Plus subscribers through ChatGPT Work and Codex. Lire la suite

The rise of artificial intelligence poses a significant threat to human labor, with many experts warning that AI could displace jobs and fundamentally change the nature of work. While some argue that AI will create new opportunities, others believe that its impact will be devastating, particularly for junior and entry-level positions. Lire la suite

Open models are becoming increasingly competitive with closed models in the AI ecosystem, with Chinese labs leading the charge. Experts say open models are crucial for innovation and research, but also pose risks that need to be balanced. Lire la suite