Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A recent write-up from researchers at HumeAI found that automatic speech recognition (ASR) models exhibiting benchmark-optimized behavior reproduce erroneous reference transcript text, indicating they are pattern-matchin…
The NanoGPT speedrun repository reports that a collaborative effort has trained a language model to 3.28 cross-entropy loss on the FineWeb validation set in under 75 seconds on 8 NVIDIA H100 GPUs, a dramatic improvement …
A developer has built Psycho-Acoustic, a mental health monitoring tool that uses the HuBERT model, HuggingFace Transformers, and Librosa to quantify emotional states from non-verbal acoustic features. The tool extracts p…
In 2018, a multi-institution research team reported in Nature Communications that their learning algorithm, S#, could sustain cooperation with people and other machines in repeated two-player games, achieving human-level…
Ornith-1.5, a family of open-weight foundation models built on Qwen3.5 and Gemma4, trains through a self-improvement loop that jointly generates tasks, constructs scaffolds, and runs reinforcement learning rollouts witho…
Translucent Energy, a Charleston-based startup, opened a 1.3-gigawatt solar module factory in Summerville, South Carolina, on August 17, creating 167 jobs by year-end, with more phases planned. The company aims to serve …
A developer is documenting their journey toward Google's Associate Cloud Engineer Certification, highlighting the certification path provided by Google Skills and MentorMe Collective. The path includes courses, hands-on …
Lightricks released LTX-2.5, a video generation model that uses a 32×32×8 spatiotemporal compression ratio (1:192) and a diffusion-based decoder to avoid blur, achieving an artifact score of 0.28 versus Flux 3's 0.45 and…
A developer experimenting with a Tokyo public transit MCP server found that adding a SKILL.md guidance layer does not automatically improve agent performance. Across six evaluations, the skill layer had little impact on …
Rust 1.98, released on August 20, introduces algebraic floating-point methods that address a performance gap where Rust dot products ran 8x slower than C++ on x86_64 CPUs, allowing per-operation reordering for vectorizat…
A new survey of positional encoding methods in Transformers argues that the field is best understood not as a chronological progression but as answers to a single question: where position information is injected into the…
Microsoft researchers introduced LoRA (Low-Rank Adaptation) in 2021, a parameter-efficient fine-tuning technique that reduces trainable parameters by approximately 10,000 times and GPU memory requirements by 3 times comp…
Speech recognition optimization requires moving beyond a single Word Error Rate (WER) metric, according to a practical guide that recommends integrating Character Error Rate (CER), Keyword Error Rate (KER), and Semantic …
Oh-My-Pi, a fork of the Pi coding agent by Mario Zechner, introduces hash-anchored edits to prevent coding agents from overwriting entire files during small edits. The terminal-native agent, built with ~80k lines of Rust…
A developer created a fork of llama.cpp optimized for AMD GFX906 GPUs (Mi50, Mi60, Radeon VII, GCN HIP), and another developer noted they had previously finetuned llama.cpp for one card, mentioning the Docker image knguy…
A developer's guide details nine techniques to improve retrieval quality in Retrieval-Augmented Generation (RAG) systems, including reranking, hybrid search, and context compression. The techniques address common product…
Propio, a financial tracking platform for short-term rentals, has implemented an AI-powered OCR system called Smart Scan that extracts financial data from receipts and invoices. The system uses asynchronous processing, s…
Clause AI, a platform for analyzing rental and lease agreements, uses a multi-agent AI pipeline built with Mastra and TypeScript. The system coordinates four specialized agents—for parsing, summarizing, risk analysis, an…
The Neural Image Format (NIF), a deep-learning-based image codec developed by Gabriel Philippus, achieves up to 26.4% lower LPIPS distortion than JPEG at 0.80 bpp on the Kodak24 dataset, with a decode latency of ~202 ms …
Inherent, a London AI lab founded by Google DeepMind alumni, says its AI agent Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at replicating scientific research, despite running on a 27-billion-par…