Sep 6, 2025

Articles

• Apertus 70B: Truly Open - Swiss LLM by ETH, EPFL and CSCS
  Apertus-70B-2509 is a new addition to the Apertus LLM collection, showcasing advancements in large language models with enhanced capabilities.

• ML needs a new programming language – Interview with Chris Lattner
  Mojo, a new programming language by Chris Lattner, aims to enhance GPU programming by combining ease of use with type-safe metaprogramming, allowing developers to harness the full power of modern hardware while maintaining productivity.

• Novel hollow-core optical fiber transmits data 45% faster with record low loss
  A novel hollow-core optical fiber achieves 45% faster data transmission with a record low loss of 0.091 dB/km, surpassing traditional silica fibers which have a minimum loss of 0.14 dB/km.

• How big are our embeddings now and why?
  Embedding sizes have significantly increased, with models like OpenAI's using 1536 dimensions, reflecting a shift from the previous standard of 300 dimensions due to advancements in training data and architecture.

• Now Live: Europe’s First Exascale Supercomputer, JUPITER, Accelerates Climate Research, Neuroscience, Quantum Simulation
  JUPITER, Europe’s first exascale supercomputer, is now operational, enabling unprecedented computational power for scientific research and AI applications.

• Europe enters the exascale supercomputing league with Jupiter
  The European Commission's Press Corner serves as a vital hub for accessing official communications, including press releases and statements, enhancing transparency and public engagement. Link to article

• [R] The Illusion of Progress: Re-evaluating Hallucination Detection in LLMs
  Current benchmarks for hallucination detection in LLMs are inadequate, often yielding low-signal results that fail to address real-world high-stakes scenarios, as highlighted in the paper here.

• [P] I Was Wrong About Complex ML Solutions - Gower Distance Beat My UMAP Approach
  Gower distance, a method for calculating distances in mixed categorical and numerical data, often outperforms complex ML solutions like UMAP, especially in small-to-medium datasets, highlighting the value of simplicity in machine learning.

• [P] Knowledge Distillation for Text-to-SQL — Training GPT-2 with Qwen2-7B as Teacher
  Knowledge Distillation (KD) enables GPT-2 to effectively generate SQL queries from natural language by learning from the larger Qwen2-7B model, demonstrating that smaller models can be trained for specific tasks without the need for massive resources.

• [D] Baseten raises $150M Series D for inference infra. where’s the real bottleneck?
  Baseten's $150M Series D emphasizes their commitment to enhancing inference infrastructure, targeting low latency and throughput optimization, with a valuation of $2.1B.

• [P] An Open-Source Pipeline for Speech-to-Speech Translation with Voice Preservation (RVC) and Lip-Sync
  The project presents an open-source pipeline for speech-to-speech translation that preserves the speaker's voice and synchronizes lip movements, initially targeting Telugu for low-resource language dubbing.

• Fast 2-Simplicial Attention: Hardware-Efficient Kernels in TLX
  Fast 2-Simplicial Attention achieves 588 Tensor Core BF16 TFLOPs with 60% tensor core utilization, marking a 1.74x speedup over the original Triton kernel, thanks to a hardware-aligned design and the use of TLX for modern GPU techniques.