Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Developer jlt-commons released Lev, an open-source decision engine that exposes the same wire API as TypeSafe's Jev classifier model and adds an escalation path to a local 2.5B LLM when the fast encoder tier returns low …
A feasibility experiment called qwen3-ebpf ran a full Qwen3-0.6B forward pass inside Linux eBPF, with BPF programs computing all 28 decoder layers, the output logits, and the next-token argmax while a C driver handled to…
A new paper establishes information-theoretic bounds on span-level latent reasoning, framing the Span-level Pause-of-Thought (SPOT) framework as an entropic optimal transport problem between distributions of hidden state…
TypeSafe shipped a model called Jev on 15 September 2026 that returns typed answers with probabilities instead of generating text, a category TypeSafe calls a System One model. Jev accepts a state and one or more questio…
Skild AI announced a post-training result in which its S1 robotics foundation model taught itself to play soccer through physical self-play, competing against previous versions of itself in NVIDIA's Isaac Sim with no hum…
The second part of an AI fundamentals explainer on transformer attention mechanisms details four widely adopted techniques — Rotary Position Embedding (RoPE), the KV cache, Flash attention, and cross-modal attention — th…
A paper submitted to arXiv on 22 Sep 2026 resolves why diffusion models overfit catastrophically rather than benignly, showing through U-Net experiments on CelebA and a random-features model with closed-form learning cur…
Developer Shrijith Venkatramana explains Muon, an optimizer that treats neural network weight matrices as matrices rather than collections of scalar coordinates, orthogonalizing momentum updates by preserving singular di…
A developer building speech recognition for Crimean Tatar found that a large Whisper model fine-tuned on the language scored 34.6% word error rate on clean audio, improving to 20.1% WER after roughly 1.5 hours of fine-tu…
A developer published a guide for deploying Kubeflow on Kubernetes as a self-hosted alternative to Azure Machine Learning, covering installation via Kustomize manifests, notebook environments, pipeline orchestration, dis…
Nvidia released Nemotron 3 Diarization on September 23, an open-weight speaker-diarization model with roughly 100 million parameters that identifies up to eight speakers simultaneously at latency as low as approximately …
NVIDIA released Nemotron 3 Diarization, a 100M-parameter open-weight speaker diarization model on Hugging Face under the OpenMDW License 1.1, which permits commercial use and tracks up to 8 speakers, including overlappin…
Skild AI revealed on September 22 that its S1 model learned to play football through self-play training entirely inside NVIDIA Isaac Sim, compressing roughly 140 years of simulated experience into a few weeks and transfe…
A developer ported Andrej Karpathy's microgpt Python code into Rust and into an array programming language called sw-MLPL, presenting the work as a Show HN post. The sw-MLPL version is more compact than both the Python a…
A user is asking the local AI community what VRAM is realistically required to fine-tune 1B to 7B parameter models on a laptop GPU, noting that most laptop GPUs offer 8GB and some newer ones 12GB, and questioning how far…
Google Cloud AI Research has published RRSI, a framework that applies classical regularization techniques to recursive self-improvement of LLM agent harnesses. The method constrains the proposal and selection of harness …
Fine-tuning, including techniques such as LoRA and QLoRA, offers a way to customize language models when prompting and retrieval-augmented generation (RAG) leave recurring gaps, according to an article on customizing mod…
An AI system detected hidden bladder cancer warning signs up to five years before diagnosis, according to research on a disease that causes about 220,000 deaths worldwide each year. The tool analyzes faint signals accumu…
A pure-Java implementation of NVIDIA's Parakeet automatic speech recognition model, called Parakeet.java, matches the reference parakeet.cpp engine word-for-word at F16 quantization with 0.00% WER difference on LibriSpee…
TypeSafe's Jev, a "System One Model" that returns typed decisions with probabilities instead of generated text, cannot deliver the calibrated probabilities it advertises, according to a critique citing tweets in which Je…