Cursor, the AI coding company, open-sourced Mixture-of-Kittens (MoK), a deterministic MoE training megakernel for NVL72s that fuses communication and computation into a single kernel, achieving a 1.41…
The EU AI Act's transparency rules, effective August 2, impose fines up to €15 million for violations, forcing East Asian tech giants like Sony and Fujitsu to redesign products or risk losing European…
AirLLM, an open-source inference library, claims to run a 70B-parameter large language model on a 4GB GPU without quantization, distillation, or pruning. The claim is technically true, achieved by loa…
AirLLM, an Apache-2.0 Python library, enables running large language models like Llama 3.1 405B on 8GB of VRAM and DeepSeek-V3 671B on ~12GB without quantization, by streaming one layer or MoE expert …
AirLLM, an open-source inference library, now supports running the 2.8T-parameter Kimi K3 model on a single 4GB GPU, using only 3.72GB of VRAM on an RTX 6000 Ada, by streaming one expert at a time for…
Huawei released openPangu 2.0 Pro, a Mixture-of-Experts language model with approximately 505 billion total parameters and 18 billion activated per token, trained entirely on its Ascend AI processors.…
A leak at DeepSeek exposed proprietary engineering pipelines, including data curation, RL alignment recipes, and system configurations, rather than model weights. The incident highlights that the true…
Claude 3.5 Sonnet is the current gold standard for agentic coding benchmarks, with superior ability to parse stack traces and fix bugs, while GPT-4o often gets stuck in looping behavior and DeepSeek-V…
Token dropping in mixture-of-experts (MoE) layers occurs when a capacity factor limits each expert's buffer, causing excess tokens to bypass the expert MLP and degrade model quality under production l…
A practical guide recommends hardening AI-generated code for legacy systems by using a multi-model verification loop, cross-model validation, and isolated test generation. The guide suggests using Cla…
Claude 3.5 Sonnet is generally superior for complex architectural planning and nuanced refactoring, while DeepSeek-V3 offers industry-leading performance in raw logic and competitive benchmarks at a s…
A guide for beginners on prompt engineering outlines techniques such as role prompting, few-shot prompting, and output format control to improve AI responses, noting that GPT-4o excels at few-shot pro…
A new benchmark suite called JAXBench, comprising 50 JAX workloads for AI-generated kernel optimization on Google Cloud TPUs, shows that target-specific documentation context matters more than model s…
Mixture of Experts (MoE) is a machine learning architecture that divides a model into specialized subnetworks called experts, with a gating network routing each input token to only the most relevant s…
Anthropic chief national security officer Tarun Chhabra identified Chinese AI developer Zhipu among companies that have allegedly distilled capabilities from frontier American models, but hours later …
NVIDIA GB300 NVL72 set a world record for pre-training DeepSeek-V3 671B at 1,648 TFLOPs per GPU, demonstrating how advances across the AI platform—from silicon to networking to software—push training …
An engineer reverse-engineered DeepSeek's AI assistant by interviewing it and checking its self-reported answers against public papers, finding the model carefully separates observation, inference, an…
China's open-weights AI strategy is winning, as labs like DeepSeek and Z.ai release competitive models under permissive licenses, building developer ecosystems and bypassing the proprietary API moat o…
A new multi-agent framework called GraphDx improves diagnostic success rates from 50-68% to 79-93% while reducing test costs by 20-54%, according to a preprint on arXiv (2607.15280v1). The knowledge-e…
Chinese AI startup DeepSeek has begun developing its own AI chips focused on inference workloads to reduce reliance on NVIDIA and cut costs, sources told Reuters. The project is in early stages and ha…