Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Lightricks released LTX-2.5, an open-weights video generation model that uses a diffusion-based video decoder to recover fine details from a highly compressed latent space, eliminating the need for a separate upscaling s…
Nari Labs' Qwen3-TTS 1.7B CustomVoice implementation achieves 10 requests per second and sub-50 ms p95 time-to-first-audio on a single NVIDIA H100 SXM, maintaining real-time playback and costing about $2 per 1M character…
Researchers at Harvard Medical School, Dana-Farber Cancer Institute, and Mass General Brigham have developed an AI tool called Aladynoulli that estimates a person's risk of developing 348 diseases by analyzing medical hi…
AGY Memory Engine, a zero-dependency SQLite FTS5 fact store and MCP server, delivers sub-2ms local retrieval with 0 MB extra RAM, serving as a persistent semantic backbone for Google Antigravity (agy) and autonomous agen…
A production AI tutor's context engineering experiments, detailed in the newsletter LAI #139, found that reducing tokens via summarization increased costs by roughly 2x despite sending 41% fewer tokens, while keeping ful…
An engineer's blog post explains the Key-Value (KV) cache optimization in large language model inference, walking through the math of attention calculation to show how caching keys and values avoids redundant computation…
NVIDIA's Agentic Variation Operators (AVO) architecture achieved a 100% score on the ARC-AGI-3 benchmark, demonstrating a frontier-level general-purpose architecture for long-horizon autonomous agents. In GPU-kernel opti…
Isopolis, a continuous, zoomable, clickable isometric pixel-art map of San Francisco covering all ~121 km² as a 22-gigapixel image, was built by an unnamed developer using a fine-tuned Qwen Image Edit 2511 model on 3D Go…
Delta Air Lines' new AI-powered pricing engine, built on a custom TensorFlow Serving cluster with Kafka event streaming, targets a 50% profit jump by scoring each shopping request against a contextual bandit model traine…
A developer's analysis of their production AI agent memory store found that feedback signals, which are crucial for learning, are almost never sent in live traffic. The audit of 359,388 concept-graph edges across 65 tena…
Brunswick CEO David Foulkes is betting on AI-powered auto-docking, an AI customer assistant, and a defense partnership with Textron to drive growth, including an unmanned surface vessel of which about 20 have been sold t…
Google DeepMind announced a research partnership with Fenris Creations, the studio behind EVE Online, to prototype new gameplay experiences, building on 15 years of AI research in games. The collaboration extends DeepMin…
Arizona State University-led research found that AI analysis of antibody patterns in blood samples from 4,089 people can predict vaccine response strength before vaccination, distinguishing strong responders from weak on…
Google Research introduced Mobility-Embedded POIs (ME-POIs), a framework that blends text descriptions with aggregated, anonymized mobility patterns to improve language models' understanding of places, achieving up to an…
GenBio AI, a Palo Alto-based startup co-founded by Nobel laureate David Baker, launched AIDO Cell, a virtual cell 'world model' that simulates a human cell in its natural state and in response to drugs and interventions …
A test by an unnamed team found that a quantized 4B multimodal model running on a single AWS Graviton4 CPU server with 16 vCPUs and 32 GB of RAM completed 360 of 380 enterprise AI tasks (94.7%) across 13 industries, with…
In a blog post, an unnamed author argues that AI performance will not plateau, countering the common belief that progress is driven by brute-force scale with diminishing returns. The author contends that combining genera…
SGLang, the open-source inference project associated with Ying Sheng and Banghua Zhu, published deployment recipes for Alibaba's Qwen3.8-27B model, enabling single-GPU serving with NVFP4 quantization and DFlash2 speculat…
A new study from arXiv researchers finds that language models leak sensitive in-context secrets through benign outputs, with 2-digit secrets reconstructed with near-perfect accuracy and 4-digit secrets at 82% exact match…
A researcher at Alignment of Complex Systems showed that a 32B open-source model, Qwen2.5-Coder-32B, can subtly introspect when external concepts are injected into its activations, despite appearing unable to do so. By a…