Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A developer measured the "exchange rate" between training data and document length for a small transformer versus a zero-parameter count-table cache, finding that 16x more training data (from 500K to 8M tokens) moved the…
A from-scratch autoregressive transformer trained in 1.5 hours on a single Nvidia RTX 5090 GPU reached 44% on the ARC-1 benchmark, matching specialized architectures TRM and HRM, according to a writeup covered by Mango D…
A machine-learning practitioner built a horse-racing ranking model trained on 1.18 million runners and evaluated it with walk-forward validation, finding that the betting market remains a very strong baseline. The projec…
A developer outlined practical techniques for optimizing large language model inference to cut latency and cost, including quantization, KV cache optimization, speculative decoding, prompt optimization, and batch process…
A developer outlined best practices for evaluating AI models, emphasizing a multi-layered framework combining standardized benchmarks like MMLU and HumanEval, adversarial red teaming for prompt injection and jailbreaks, …
Google DeepMind researchers released a public dataset mapping the Drosophila fruit fly brain, documenting roughly 166,000 neurons and more than 125 million connections, which hobbyists have since loaded into simulations …
Researchers trained a publicly released Drosophila connectome of roughly 166,000 neurons and more than 125 million connections to classify emails, reaching about 80% accuracy against a small conventional neural network u…
A study of sequence weighting across in-house and open-weight language models found non-monotonic scaling behavior: as models grow from small to medium scale they shift from learning general patterns independent of data …
A benchmark of seven reinforcement-learning trainer algorithms on Qwen3-14B in a decentralized exchange (DEX) arbitrage gym found that CISPO posted the highest training reward curve in thinking mode (0.8951 last-10 mean)…
A developer added semantic search to the Blazing Story documentation site, a Blazor WebAssembly standalone app hosted on GitHub Pages with no server-side code, API, or database. The implementation generates document embe…
Capsize Games released Spikeforge, an open-source toolchain for spiking neural networks that encodes data into spikes, trains and inspects networks, and deploys models to targets including the lava_loihi2 backend. The pr…
A paper submitted to arXiv on 10 Sep 2026, titled "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement," proposes a development roadmap for recursive self-improvement (RSI) in AI systems. The paper use…
A developer behind NNTC, an open-source texture compression project on GitHub, reports that a bilinear/degree-2 polynomial decoder fitted to each PBR material can match or beat MLP-based neural texturing on raw PSNR afte…
Researchers Jacob Fein-Ashley and colleagues posted a paper to arXiv on 12 May 2026 introducing Attractor Models, a recurrent architecture in which a backbone module proposes output embeddings and an attractor module ref…
Preston Fu proposed progressive point matching (PPM), an asymptotically unbiased framework for assigning partial credit in reinforcement learning of long-horizon language model tasks, published in September 2026. PPM tre…
MoonshotAI researchers submitted a paper on 18 Feb 2025 introducing Mixture of Block Attention (MoBA), an attention mechanism that applies Mixture of Experts principles to let long-context large language models select wh…
OpenGEMM, an open-source project from developer aramesh10, released CUDA GEMM kernels for NVIDIA's B200 GPU (sm_100a), installable via pip and requiring PyTorch 2.8+ and CUDA 12.9+. The library emits standalone CUDA kern…
Princeton researcher Yifan Zhang published a technical report proposing a Recurrent Looped Transformer (RLT) that pairs a causal encoder with a recurrent decoder carrying its final hidden state and layerwise sliding-wind…
Researchers introduced PC-ALM, a local-learning method that trains 1000-layer neural networks using only local dynamics and without backpropagation, generalizing predictive coding by replacing its energy function with an…
NVIDIA Triton Inference Server is being integrated into modern ETL architectures to enable AI-driven data enrichment, dynamic batching, and multi-framework model serving across GPU and CPU environments. The approach, des…