Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Fine-tuning large language models can cause catastrophic forgetting, where new knowledge overwrites old knowledge, making models specialized but less versatile. Techniques like LoRA and multi-task training can mitigate t…
Newton Research launched Unlimited Analytics, a new agentic layer for its AI ad analytics platform that uses causal modeling to predict campaign outcomes, on Tuesday. The tool, which includes agents for planning, activat…
Kakao's Kanana-o model now accepts plain-language voice instructions such as 'read it in a sad voice' or 'read it in a Gyeongsang dialect', reflecting them in speed, volume, pitch, emotion, intonation, and intensity. On …
Soup v0.72.4, an open-source CLI tool, now supports preference alignment (DPO, ORPO, SimPO, KTO) via layer streaming, enabling fine-tuning of 8B models on a 4 GB laptop GPU. The update claims bit-exact results against no…
Bridgewater Associates, in partnership with Thinking Machines Lab, fine-tuned an AI model on its own investment managers' workflows, achieving 85% accuracy and reducing errors by almost 30% compared to frontier models at…
In a blog post on AI Safety, the author argues that causal evidence for latent representations in neural networks is insufficient for deployment in safety-critical systems, because a representation may be causal without …
A developer has open-sourced a Decay-Gated O(N) Causal Linear Attention architecture with fused Triton/CUDA kernels, aiming to bypass quadratic multi-head attention bottlenecks. The project includes a repository and rese…
A comparison of two personal-data detection models by an unnamed company reversed when the evaluation dataset changed, showing that benchmark choice can flip conclusions. On the company's own 1,800-example corpus, its ex…
A developer seeking advice on building a retrieval-augmented generation (RAG) system for government and internal organizational documents asks about optimal chunk sizes for BM25 and semantic search (cosine similarity) re…
AI4Bharat at IIT Madras and Josh Talks launched a broader Voice of India evaluation platform in New Delhi on August 4, 2026, expanding beyond its earlier speech-recognition benchmark to cover voice, language, vision, and…
A developer trained a 4-billion-parameter language model using reinforcement learning to rewrite AI-generated text into more human-sounding prose, using AI detectors as reward functions. The project, described in a blog …
A study submitted to arXiv on 3 Aug 2026 finds that large language models (LLMs) fail at tabular prediction primarily due to input dimensionality, not noise handling, CSV formatting, tokenization, or test-point count. Th…
Homebench, a new open-source command-line tool, benchmarks local large language models for speed, memory, and quality, presenting results in a live terminal leaderboard. The tool, installable via pip, supports Ollama, LM…
A developer built a deterministic solver for a Dota 2 drafting game after finding that an LLM underperformed. The solver, tested with common random numbers and strict train/validation/test splits, achieved a higher title…
In 2026, edge inference has advanced to the point where an 80B Qwen model runs in just 4.3GB of RAM on a MacBook Pro, and a 35B model runs on an iPhone 18 Pro. This is achieved through a combination of hybrid quantizatio…
AirLLM, a new tool from GitHub user lyogavin, enables inference of 70B-parameter models on a single 4GB GPU by aggressively quantizing and offloading layers, potentially reducing hardware costs by 10–20×. This allows dep…
Bespoke Labs is hiring a contract researcher to design and evaluate reinforcement learning environments and benchmarks for long-horizon agentic tasks, which require hours, days, or weeks of coherent multi-step reasoning.…
For travel apps, retrieval-augmented generation (RAG) is recommended over fine-tuning for frequently changing facts, with fine-tuning reserved for stable behavior, according to a technical guide referencing Hugging Face …
Peter Kenny responded to Pat Walters' critique of the OpenADMET initiative, defending the role of machine learning in modern drug discovery. Kenny argued that Walters mischaracterized the utility of ML models, noting tha…
Researchers introduced CaRL (Capability-aligned Reinforcement Learning), a method that trains large language models to refuse futile reasoning on tasks beyond their capability, reducing computationally expensive yet sema…