The EU AI office proposes external tests for large models. The first draft of the EU code of conduct for universal AI systems requires providers of GPAI models with systemic risk to notify the EU two weeks before training begins and allow external tests by the EU AI office and third parties. The code mandates a comprehensive security framework with risk analysis, protective measures, and evaluation procedures. Each model must have a detailed security report documenting risk analyses, effectiveness evaluations, and test results. Providers must also adhere to strict copyright rules, respect rights holders’ reservations, and support the standard robots.txt. Measures should be taken to exclude piracy websites from crawling activities. The draft is now being discussed and further developed with stakeholders.
The language model Intellect-1 was trained simultaneously on multiple continents. The AI startup Prime Intellect has completed the first decentralized trained language model with 10 billion parameters after eleven days of training. Intellect-1 was trained on computers in the USA, Europe, and Asia simultaneously and is set to be released as open source in a week. The technological basis is “OpenDiLoCo,” Intellect’s open-source implementation of DeepMind’s Distributed Low-Communication method. It allows training on globally distributed devices with reduced communication requirements. Intellect-1 is based on the LLaMA-3 architecture and was trained with high-quality open-source datasets. Prime Intellect sees the project as a milestone for democratizing AI training. There are plans to expand decentralized training to the most powerful open-source models. However, with 10 billion parameters, Intellect-1 is still relatively small compared to commercial models. The scaling of the idea remains to be seen.
Nvidia’s new AI can generate music and give voices accents. Nvidia presents Fugatto, a novel AI technology for audio generation, which aims to surpass existing competing services in versatility and quality. The AI enables various audio transformations such as converting piano playing into singing or modifying voices regarding accent and mood. The system, trained exclusively with open-source material, is controlled via text commands or audio files. In demonstrations, Nvidia showcases various applications: from generating train noises that transform into orchestral music, extracting and regenerating voices, to adding instruments to existing music pieces. The technology is intended for music production, game development, and creative individual users. Due to potential risks, there are no concrete plans for public release yet.
Elon Musk vs. OpenAI is more thrilling than any HBO show. Newly released emails document the tensions that led to the split between Elon Musk and the AI company OpenAI, which he co-founded. Musk was driven by concerns from the start that Google and DeepMind might dominate too much. An initial conflict arose in 2016 when OpenAI negotiated with Microsoft over computing resources. In 2017, OpenAI leadership expressed concerns that Musk, as CEO of a planned subsidiary, could gain “unilateral absolute control” over a potential artificial superintelligence. Musk responded with an ultimatum. By early 2018, Musk’s criticism of OpenAI’s strategy intensified. As a last resort, he suggested OpenAI join Tesla. After OpenAI’s rejection, Musk resigned from the board in February 2018. Tensions persist to this day. Former DeepMind top researcher Nando de Freitas dismissed the allegations against DeepMind as “absolute nonsense.” He accuses Musk and OpenAI leadership of exploiting DeepMind’s openness and sees the emails as an argument for more transparency in AI development.
The pilot project for Robo-Buses in Friedrichshafen has been extended. Self-driving minibuses will continue to roam the streets of Friedrichshafen for a few more months. Baden-Württemberg’s Transport Minister Winfried Hermann announced the extension of the RABus consortium project in the city on Lake Constance at the end of the first funding phase. RABus stands for “Real Laboratory for Automated Bus Operation.” The consortium includes the Research Institute for Automotive Engineering and Vehicle Engines Stuttgart, DB Regio Alb-Bodensee, and Friedrichshafen city transport. The initiative started in September 2020. Four years later, ZF Friedrichshafen received a “Level 4” approval from the Federal Motor Transport Authority for the shuttles from the technology partner, allowing fully automated driving without constant human intervention. After successful completion of the multi-year test operation, the so-called test phase began on October 28 in Friedrichshafen and Mannheim, where selected passengers were allowed on board for the first time. Originally, the project, funded with 14 million euros, was to end by the year’s end. Now, passengers can use the autonomous shuttles until the end of June 2025. This extension aims to gain additional insights for the future use of automated vehicles in public transport. So far, 1450 citizens have signed up for a test ride, according to accompanying researchers from the Karlsruhe Institute of Technology. They mainly wanted to use the service to bridge gaps in public transport offerings and on shorter routes.
AI usage in companies has almost doubled. AI technologies are increasingly being adopted in the German economy. According to a survey by the Federal Statistical Office, one in five companies reported using artificial intelligence. In last year’s survey, it was 12 percent. Larger companies tend to use AI more frequently. According to the survey, AI technologies are most commonly used for text analysis and speech recognition. The main areas of application for these technologies are marketing and sales. Additionally, companies use AI for production and service processes, organizing their administration, and bookkeeping. Those who have not adopted AI mainly cite a lack of knowledge, uncertainty about legal consequences, and concerns about data protection as reasons. About one in five non-AI users do not consider using artificial intelligence in their company to be worthwhile.