Jul 13, 2025

News Highlights

• ETH Zurich and EPFL to release a LLM developed on public infrastructure
  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.

• Amazon CEO says AI agents will soon reduce company's corporate workforce
  Amazon's CEO Andy Jassy predicts a significant reduction in the corporate workforce as the company increasingly adopts generative AI tools, aiming for efficiency gains in the coming years.

• Kimi k2 largest open source SOTA model?
  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.

• The upcoming GPT-3 moment for RL
  Reinforcement learning (RL) is poised for a transformative shift akin to GPT-3, moving from narrow task fine-tuning to massive-scale training across diverse environments, which will enhance few-shot, task-agnostic capabilities.

• Context Engineering Guide
  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.

• Hill Space: Neural nets that do perfect arithmetic (to 10⁻¹⁶ precision)
  Hill Space transforms neural network arithmetic by enabling calculation of optimal weights for discrete operations, leading to rapid convergence and high reliability in mathematical tasks.

• Show HN: ArchGW – an intelligent edge and service proxy for agents
  ArchGW is an intelligent proxy server designed to streamline the development of AI agents by managing prompt handling, routing, and safety checks, thus reducing repetitive low-level coding tasks.

• How to scale RL to 10^26 FLOPs
  Scaling reinforcement learning (RL) to 10^26 FLOPs requires a shift from traditional methods to leveraging next-token prediction on web data, enhancing reasoning capabilities beyond mere model size.

• [R] I am building a framework for AI which will allow it to self learn and self evolve
  The Neuro-Schema framework aims to enable AI to learn, adapt, remember, and evolve autonomously, moving beyond traditional response mechanisms to a more dynamic interaction model.

• [P] Hill Space: Neural networks that actually do perfect arithmetic (10⁻¹⁶ precision)