Jul 12, 2025

• ETH Zurich and EPFL are set to release a fully open large language model (LLM) in late summer 2025, developed on the “Alps” supercomputer, emphasizing transparency and multilingual capabilities across over 1,000 languages.

• Bad actors exploit LLM vulnerabilities to disseminate propaganda, manipulating outputs through a process known as LLM grooming, which involves flooding models with false narratives to influence their responses.

• Computer scientists have developed a method to exploit the Fiat-Shamir transformation, allowing them to prove false statements within blockchain and cryptographic systems, fundamentally challenging the security of these technologies.

• Kimi K2 is a mixture-of-experts (MoE) language model with 32 billion activated parameters and 1 trillion total parameters, optimized for advanced reasoning and coding tasks using the Muon optimizer for stability during training.

• Differentiable Physics offers a more reliable approach to simulations by combining traditional numerical methods with autodiff, addressing issues of reproducibility and weak baselines prevalent in the Scientific ML community.

• Context engineering redefines prompt engineering, emphasizing the design and optimization of instructions and context for LLMs to enhance task performance, particularly in complex AI workflows.

• Watson, developed through the DeepQA project, achieved human-level performance in real-time on the quiz show Jeopardy!, showcasing its advanced capabilities in precision, confidence, and speed.

• Modeling non-Gaussian labels requires specialized approaches, such as density mixture networks for bimodal distributions and hurdle models for zero-inflated data, to effectively capture the underlying patterns in sparse ratings.