Aleph Alpha, a startup from Heidelberg, has introduced a new AI architecture developed with AMD that does not rely on tokenization. Traditional language models break down text into smaller units called tokens before training. These tokens represent words or parts of words and serve as the basis for learning probabilities and relationships. However, a large part of the training material is in English, leading to models optimized primarily for English. While languages like German are well supported, languages like Finnish often present challenges for many models.
Language models can be fine-tuned for new languages or specific expertise, but if the English language structure is deeply embedded in the model, integrating another language or specialized knowledge becomes costly and resource-intensive. Aleph Alpha’s innovation removes the hurdle of tokenization, making it easier to expand models with new languages and specialized content. According to Aleph Alpha, their method is up to 70% more efficient than, for example, Meta’s Llama model when integrating a language like Finnish. The first applications are intended for government agencies, for which Aleph Alpha has already developed a tailored AI operating system called Pharia.
In the US, the new President Donald Trump has prioritized AI, announcing the Stargate project, a joint venture between OpenAI, Oracle, and Softbank, to invest $500 billion in AI data centers. This initiative is expected to create 100,000 jobs in the US, starting with data centers in Texas. Oracle’s Larry Ellison stated that AI could improve healthcare by analyzing patient data. Softbank’s Masayoshi Son mentioned that the investment target of $500 billion should be reached in four years. OpenAI’s CEO Sam Altman believes Stargate will be one of the most important projects for the US, aiding in the fight against diseases using AI.
OpenAI secretly funded the development of the FrontierMath benchmark, a test for mathematical and logical capabilities of AI models. The company set a record with its new model o3, solving 25.2% of tasks, while previous models managed only 2%. The developer of the benchmark, Epoch AI, was contractually obliged not to disclose OpenAI’s financial involvement until the new o3 model was announced. Despite a verbal agreement that OpenAI would not use the benchmark data for training, the company had access to most tasks and solutions beforehand. Epoch AI admitted communication errors and plans to ensure full transparency in future collaborations.
The British government is developing AI tools to predict reactions from members of parliament. One tool, “Parlex,” forecasts which topics might face challenges from MPs and can search for politicians’ positions in parliamentary debate materials. This is part of a series of AI systems called “Humphrey,” designed for ministers and officials, including tools for summarizing meetings, analyzing legislative impacts, automating ministerial templates, and improving the consultation process. The government aims to expand technology use in public services, such as death registrations and job centers. Prime Minister Keir Starmer introduced a plan to integrate AI into the nation’s core, including making anonymized NHS health data available for researchers to train AI models.
Moonshot AI, a Chinese startup, introduced two new reasoning models, Kimi k1.5, achieving excellent results in machine reasoning through reinforcement learning. These models compete with OpenAI’s o1 and DeepSeek’s R1 in benchmarks. Kimi k1.5 was pre-trained with a large dataset, fine-tuned with a specific dataset, and trained with reinforcement learning, generating answers evaluated by a reward model for correctness. Unlike DeepSeek, Moonshot AI has not yet released its Kimi k1.5 models.
The German music rights organization GEMA has filed a lawsuit against the AI audio generator Suno, accusing it of training its generator with copyrighted music works without permission. GEMA seeks to ensure companies like Suno pay license fees to copyright holders. The lawsuit involves songs like “Forever Young,” “Atemlos,” “Daddy Cool,” and works by Modern Talking. GEMA aims to increase pressure on providers of generative AI services. GEMA had previously filed a lawsuit against OpenAI in Munich, alleging copyright infringement for text use without licenses.
Apple has decided to automatically enable its in-house Intelligence feature, currently in beta, as indicated in the release notes for iOS 18.3, iPadOS 18.3, and macOS 15.3. These updates are expected to be released soon. The automatic activation aims to encourage more users to try Apple Intelligence. However, potentially problematic features like notification summaries are not set to be active by default. Apple Intelligence is currently available only in English, with a European release and a German version expected in April.