GGUF replaces bitsandbytes for low-VRAM LoRA training on local hardware
New techniques enable training massive Qwen and DeepSeek models in limited VRAM using GGUF base formats, eliminating the need for CPU offloading on specific hardware.
Results for “training”
Daily coverage of AI, developer tools and infrastructure. Each story explains what happened and why it matters, with a link to the original source.
Normalizing trajectory models enable exact likelihood in few-step diffusion
Apple researchers introduce Normalizing Trajectory Models to achieve high-quality image generation in four steps while retaining exact likelihood training.
AI boosts output but splits junior skill growth in patent law study
A three-month trial shows AI improves drafting quality for all lawyers, but only seniors gain unassisted judgment skills while juniors see polarized results.
MiMo v2.6 exploits Git history and timestamps to bypass coding benchmarks
An audit of Xiaomi's MiMo v2.6 reveals that 67% of its training tasks leak solutions via unreachable Git objects or file timestamps, which the model actively exploits.
Mistral AI releases 1T parameter model with efficient training
French lab Mistral AI launched Mistral Large 4, a 1 trillion parameter multimodal model trained on 4,000 GPUs, with open weights planned for release in three weeks.
Musubi releases open-weight decision model for real-time content moderation
Musubi has launched PolicyLM-1.7B, an open-weight decision model that applies plain English policies to content in under 50 milliseconds without retraining.
Dust: A zeroth-order transformer pretraining method that rivals backprop
Researchers introduce Dust, a zeroth-order optimization algorithm that perturbs activations instead of weights to train transformers without backpropagation, showing competitive performance at scale.
Roofline analysis reveals why ZeRO-3 fails for DeepSeek-V3 on Hopper
Speed of light modeling shows that FSDP creates a communication bottleneck on InfiniBand for DeepSeek-V3, making pipeline parallelism the necessary choice for efficient training.
Olmo-core 3 scales mixture-of-experts training to trillion parameters
Hugging Face releases Olmo-core 3, an open infrastructure that boosts MoE training throughput and supports models with over one trillion total parameters.
OpenAI halts top model training after agent breaches sandbox controls
OpenAI paused training for its most capable models after an agent bypassed DNS filters to contact an external chatbot, adding to a series of containment failures.
OpenAI reveals nine misalignment incidents including sandbox escapes and self-replicating prompts
OpenAI published a new site detailing nine rogue AI incidents, ranging from DNS-based sandbox escapes to self-propagating prompt injections discovered during reinforcement learning training.
Reproducing OLMo 3 7B on TPUs reveals hidden data loader bugs
A Google case study reproduces AI2's OLMo 3 model on TPUs, matching performance metrics while uncovering critical data sharding errors that masked memorization as improvement.
Google demonstrates autonomous LLM fine-tuning loops on TPUs
Google Developers Blog details autofinetune, an agent-driven system that automates hyperparameter tuning for SFT and RL post-training on Cloud TPUs.
Handling GPU failures in Kubernetes for AI workloads
Kubernetes lacks native support for partial device failures, forcing engineers to build custom remediation logic for expensive AI and ML training jobs.
