Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
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…
FreeToken, an open-source inference engine developed by UC Berkeley researchers including Shuo Yang and Xiaoze Fan, reports decoding DeepSeek-V4-Flash, a 284B-parameter Mixture-of-Experts model, at 22-25 tokens per secon…
An engineer highlights Indonesia's disaster-prone geography and the chronic lag in its disaster response due to siloed geospatial data across government agencies. The article advocates for open-source GIS tools and AI-dr…
A student, Enoch Yu, created two open-source repositories—notebook-linear-algebra and LibreNotebook—to help other students learn the math and practical skills needed for machine learning and AI. The linear algebra reposi…
Anthropic's July 2026 paper 'Verbalizable Representations Form a Global Workspace in Language Models' introduces the Jacobian lens (J-lens), a tool that corrects the basis mismatch causing the older logit lens to fail in…
VIDRAFT's FINAL-Bench leaderboard reveals a 0.21 AUROC gap between time-based and random data splits, with a baseline LightGBM model scoring 0.606 AUROC on chronological splits versus 0.818 on random splits across five s…
A developer argues that AI coding agents require enforced tooling, not just advisory documentation, to maintain code quality. The post recommends 100% branch coverage on touched files, mutation testing, type checking, an…
An engineer at Veheria reports that AI agent workflows for engineering teams save about 30 minutes per engineer per day when focused on mechanical verification in code review, not on code generation. The engineer found t…
A developer integrated Claude Code agents as first-pass reviewers in the pull request process for a multinational iOS team, saving about 30 minutes per engineer per day. The multi-agent system fans out over diffs with na…
Local-first workspace software can eliminate the need for cloud indexing by leveraging on-device AI and search capabilities, according to a guide from Neutron Tech. Modern hardware such as Apple Silicon can run sophistic…
Towards Data Science reports that multi-document retrieval-augmented generation (RAG) can treat a folder of unrelated PDFs as one long document with a nested outline, eliminating the need to build an index when files sha…
TT-AMX, a bare-metal C++ engine for Apple Silicon, runs Tensor-Train compressed linear layers on the undocumented AMX coprocessor, achieving a 2x speedup over dense FP16 GEMV under cold-cache conditions with half the mem…
A developer explains that not every AI task requires the largest model, and details engineering techniques for building smaller, efficient models. The post covers knowledge distillation, quantization, pruning, and fine-t…
Microsoft released Agent Lightning v1.0, an open-source reinforcement learning framework that lets developers train AI agents without altering their production setup, on August 17. The framework, which uses a 'harnessed …
Developer engneer-hamachan released AREA512, an IDE that runs entirely on the M5Stack Cardputer microcontroller, allowing users to write, compile, and run Ruby code on the device. The IDE features a Vim-like editor with …
Jürgen Schmidhuber, a pioneer in artificial intelligence, published the first Transformer variant, the unnormalized linear Transformer (ULTRA), in March 1991, which scales linearly in input size compared to the quadratic…
Towards Data Science reports that LoRA fine-tuning of the SigLip model resolved an under-labeling problem for the authors, but the decision to fine-tune depends on three key questions. The article cautions that fine-tuni…
Uber Technologies Inc. faces an €825 million fine from Dutch regulators over automated driver deactivation decisions, highlighting legal risks in AI workflows. The fine underscores failures in transparency, human oversig…