Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Traditional cloud reliability models, which focus on individual instances, are inadequate for trillion-parameter AI models, necessitating a shift to cluster-level reliability. Google presents its cluster-level framework …
Traditional intrusion detection systems (IDS) like Snort rely on specific signatures, which creates a critical gap in coverage for novel or modified attacks during the exposure window before a new rule is written and dep…
Researchers at BalCapRL introduced a balanced reinforcement learning framework for multimodal large language model image captioning that jointly optimizes utility-aware correctness, reference coverage, and linguistic qua…
DeepSeek-V4's million-token context capability stems from a hybrid attention architecture that compresses context before KV storage, reducing cache pressure. Together's early bring-up on NVIDIA HGX B200 found that servin…
The article describes the author's successful setup for running local AI models on an M4 Mac with 24GB of memory, specifically highlighting Qwen 3.5-9B (Q4 quantized) as the best performing model at ~40 tokens per second…
AMD has released the MI350P, a PCI-Express version of its Instinct MI350X GPU that is intentionally built with half the components to enable air cooling for on-premise AI inference and small model training. The new card,…
Researchers from NVIDIA have developed a new sparse data format and custom GPU kernels, called TwELL, that reshape unstructured sparsity in transformer language models to align with GPU architecture, achieving over 20% s…
Researchers have introduced Adaptive Parallel Reasoning, a new approach that allows AI reasoning models to autonomously decide when to decompose problems into parallel subtasks and how many concurrent threads to spawn. T…
Apple hosted a two-day Workshop on Privacy-Preserving Machine Learning & AI in early 2026, bringing together its researchers and the broader academic community to discuss advances in private learning, foundation model pr…
Netflix released void-model on Hugging Face, and a developer used the Goose CLI agent with Together's dedicated containers skill to deploy the model for inference on release day with a single prompt. The agent automatica…
Researchers at Duke University introduced Reward-Variance Policy Optimization (RVPO), a risk-sensitive alignment method that penalizes inter-reward variance to prevent language models from neglecting critical constraints…
Researchers have developed Speculative KV coding, a lossless compression method that reduces the size of large language model KV caches by up to approximately 4x using a smaller, faster predictor model. The technique, in…
Zed Industries released Zeta2.1, an updated edit prediction model that emits 3x fewer output tokens than its predecessor, reducing response times by up to 50 milliseconds and requiring 30% fewer servers to handle the sam…
Artificial intelligence systems are now helping to build improved versions of themselves, with OpenAI's GPT-5.3-Codex assisting in its own creation and Anthropic reporting that Claude Code writes the majority of its code…
Despite predictions that longer-horizon training would slow AI progress as models require more time and FLOPs to complete complex tasks, AI capabilities have accelerated rather than stalled. The METR horizon-length graph…
Researchers Chen Huang, Xianhang Li, Vimal Thilak, Etai Littwin, and Josh Susskind have developed Text-Conditional JEPA (TC-JEPA), a visual self-supervised learning model that uses image captions to reduce prediction unc…
Researchers at Apple have developed a new learned image codec that achieves 2.3–3x bitrate savings against traditional codecs like AV1, AV2, VVC, ECM, and JPEG-AI, while also outperforming the best learned alternatives b…
Here is a 2-3 sentence factual summary of the article: The article describes the process of migrating an online reinforcement learning (RL) training system from the vLLM V0 engine to the V1 rewrite, prioritizing backend…
Evolution of "structured output" in AI, from late 2022 to 2024, where developers moved from prompting models to produce valid JSON to using specialized libraries and grammar-constrained decoding. Key libraries like Micro…
Impact of AlphaEvolve, a Gemini-powered coding agent developed by Google DeepMind, which has graduated from pilot testing to become a core infrastructure tool. It has optimized the design of next-generation TPUs, improve…