Parliament monitoring. AI Act co-rapporteur Brando Benifei (S&D, Italy) briefed IMCO on Monday, and newcomer Michael McNamara (Renew, Ireland) briefed LIBE on Thursday on the first meeting (held on 24 October) of the Parliament’s monitoring group of AI Act implementation. In LIBE, Birgit Sippel (S&D, Germany) said she was sad the second meeting was cancelled after it was announced, and Markéta Gregorová (Greens/EFA, Czechia) complained that she learned first about the planned 4 and 11 December meetings in the press. AI Safety network. The AI Office attended the first meeting of the network of AI Safety Institutes (AISIs) in San Francisco on Wednesday and Thursday. The conference focused on AI-generated content and evaluations of general-purpose models. The safety institutes published joint mission and risk assessment statements. The work is "setting the stage for discussions at the Paris AI Action Summit in February 2025,” the Commission’s press release reads. The US will serve as the first chair of the network. EU-Singapore. The Commission signed an administrative arrangement between the AI Office and Singapore’s AI Safety Institute on Wednesday as part of the EU-Singapore Digital Partnership. US taskforce. The US AISI launched a task force that “brings together experts from Commerce, Defence, Energy, Homeland Security, NSA, and NIH to address national security concerns and strengthen American leadership in AI innovation” on Wednesday. The institute is seen as being in jeopardy as Trump has vowed to repeal Biden’s executive order that established it.
Making the rounds. French Minister in charge of AI and digital policies Clara Chappaz met with “20 top VCs [venture capital firms]” and representatives from Anthropic, OpenAI, Meta, Microsoft, Google, and Nvidia in San Francisco in a visit to the US spanning Tuesday to Saturday. The objectives: incentivise funding in French Tech start-ups and prepare the Paris AI Summit in February. She also met with the AI Safety Institute. Inference scaling laws. Much of the progress in general-purpose AI came from surprisingly consistent gains from scaling up compute resources used in training, often called ‘scaling laws.’ After recent reports of diminishing returns, some talk of inference scaling laws that describe gains from scaling up compute use when using an already trained model. NVIDIA was grilled about the new trend in their earnings call this week, as inference-specialised chips like Google’s TPUs that outperform NVIDIA chips for inference could chip off NVIDIA’s dominant position, TechCruchreported on Wednesday. Chinese reasoning. Chinese AI research lab Deepseek released a model on Wednesday rivalling OpenAI’s o1 model in reasoning capabilities, they announced on X. Both models are using extra inference computing to fact-check themselves. Le Chat. French Mistral AI released Pixtral Large and upgraded its ChatGPT-competitor Le Chat with image generation and web search, the company announced on Monday. Uber eyeing Pony. Uber seeks to invest more than $10 million (€9.6 million)worth of shares in Chinese autonomous driving company Pony AI’s US initial public offering, Bloombergreported on Thursday. Autonomous driving rules. Members of Trump’s team have told advisers they plan to make a federal framework for fully self-driving vehicles a priority, sources told Bloomberg. AI readiness. Only 6% of EU companies are “ready to capture AI’s potential, compared with 19% in the US and 20% in China, Cisco said in a study published on Thursday. History of OpenAI distrust. Emails revealed in Elon Musk’s lawsuit against OpenAI showed that Ilysa Sutskever and Greg Brockman worried about Sam Altman’s motivations as early as 2017, Transformerreported on Tuesday. | 