# Jun 22, 2026 ML Times

## Codex logging bug may write TBs to local SSDs  
- **Codex SQLite logs** are projected to write **~640 TB/year**, which can lead to rapid SSD endurance depletion, potentially exceeding the lifespan of consumer SSDs rated at **600 TBW** in less than a year.

## Apertus – Open Foundation Model for Sovereign AI  
- **Apertus Mini** introduces **16 small language models** that showcase advanced **distillation** and **quantization techniques**, enhancing model efficiency and performance.

## Good results fine tuning a local LLM like Qwen 3:0.6B to categorize questions  
- **Fine-tuning a local LLM** can significantly enhance question categorization accuracy, achieving up to **92%** correct classifications by using a two-character opaque ID system to avoid semantic overlap in categories.

## Moebius: 0.2B image inpainting model with 10B-level performance  
- **Moebius** introduces a **lightweight inpainting framework** that significantly reduces computational costs while maintaining high-quality image generation, achieving this with less than **2% of the parameters** of traditional models.

## Munich 1991: The Roots of the Current AI Boom  
- The **1991** breakthroughs in Munich laid the groundwork for today's AI boom, introducing key concepts like the **Transformer**, **unsupervised pre-training**, and **neural network distillation**, which are foundational to modern Large Language Models (LLMs) such as ChatGPT.

## A Theory of Why Prompt Injection Works  
- **Prompt injection** is identified as a form of **role confusion**, where models misinterpret user inputs, leading to unintended outputs and behaviors, as detailed in the upcoming ICML paper by Ye et al. (2026) [arXiv](https://arxiv.org/abs/2603.12277).

## Some new updates to Papers with Code [P]  
- **New features** on Papers with Code enhance research discovery, including **SOTA badges** for top-performing papers and a **trending score** that combines GitHub star velocity with Hugging Face metrics, improving visibility for impactful research like GLM-5.2.

## I Gave an AI a Civilization to Run. It Built a Nuke – Launching CivBench  
- An AI tasked with running a civilization in **Civilization VI** ultimately resorted to building nuclear weapons to counter a cultural threat, illustrating the complexities of strategic decision-making in AI systems.

## Hotter Than a Hot Tub: The 45°C Breakthrough to Cool AI’s Biggest Machines  
- **NVIDIA's AI servers can operate with coolant at 45°C, a significant advancement that enhances energy efficiency in data centers by reducing cooling energy consumption.**

## Data-centric debugging for teams training neural nets [P]  
- **WeightsLab** is a powerful tool that allows teams to **pause training mid-run** to inspect live loss signals, effectively identifying data issues like mislabels and class imbalances before they compromise model performance.

## At ISC, JUPITER Shows What Exascale Science Looks Like  
- **JUPITER**, Europe’s **first exascale supercomputer**, utilizes NVIDIA Grace Hopper Superchips to tackle complex scientific challenges, including brain mapping and climate modeling, demonstrating the transformative potential of exascale computing.

## Manticore Search 27.1.5: Auth, sharding, conversational and faster vector search  
- **Manticore Search 27.1.5** introduces **built-in authentication**, **sharded tables**, and **conversational search**, enhancing security and usability for large-scale deployments.

## NAIRR Science Program Reshapes Scientific Research, Powered by NVIDIA AI Infrastructure  
- The **NAIRR Science Program**, supported by **NVIDIA AI Infrastructure**, has facilitated over **700 innovative research projects** in the U.S., enhancing fields like healthcare and energy through advanced computational resources.

## NVIDIA Vera CPU Opens the Way for Agentic Scientific AI at Los Alamos National Laboratory  
- The **NVIDIA Vera CPU** powers new supercomputers at **Los Alamos National Laboratory**, enhancing capabilities for **agentic AI** in scientific research, particularly in materials simulation and molecular design.

## From Materials Simulation to Experimental Astronomy, New NVIDIA AI Software Unlocks Scientific Discoveries  
- **NVIDIA's new AI software** accelerates scientific research across fields, transforming hours of computation into real-time processing with tools like **cuPhoton** and **DAQIRI**.