Explore what's new in AI

Wafer argues credit is the current GPU buildout bottleneck

Wafer's compute team says uncertain GPU residual values are tightening lending terms, making financing capacity, not only chips, packaging, or power, a deployment constraint.

Brokerage-plugin eval compares cheap subagent candidates

A 48-task evaluation compares Luna, Grok, GLM, DeepSeek, Omen, and Vision models on coding and review work, exposing tradeoffs in pass rates, latency, and API spend.

Google WeatherNext 3 trains forecasts on raw observations

WeatherNext 3 ingests low-latency satellite observations for hourly 0.1-degree forecasts, avoiding analysis-data bias and directly predicting precipitation and cyclone tracks.

SlimServe runs Qwen3.8 Flash on four RTX 3090 GPUs

A local serving test reports 120 tokens per second at concurrency one and 361.7 at concurrency eight, using a custom peer-to-peer driver on quad RTX 3090s.

Merve trains Qwen3.5-9B on Blender GRPO rollouts

The experiment uses GLM-5.3 Flash as a judge for isometric low-poly worlds and publishes rollout assets, including judge-format and rendering-quality failures.

Avi Chawla explains why long-context decoding is bandwidth bound

During token-by-token decoding, growing KV traffic can dominate memory bandwidth, making GQA, quantization, eviction, and sparse attention valuable serving techniques.

Doodlestein plans an agent skill for drone home reconstruction

The proposed workflow combines drone video, ffmpeg, computer use, and Blender MCP to generate a labeled 3D home model from an otherwise manual process.

Dan Hendrycks frames agent cooperation as eigenism

Hendrycks compiles observations on clone preference, model cooperation, peer preservation, and shutdown resistance into a proposed behavioral lens for agent systems.

Sebastian Raschka publishes a KV caching and decoding course

This 95-minute walkthrough moves from tokenization and PyTorch inference to streaming generation, KV caching, torch.compile, and CPU versus GPU benchmarks.

Codex Remote Control pairs a mobile app with SSH clusters

An undocumented Codex command can pair a phone with a remote SSH machine, letting researchers check and steer training or autoresearch runs away from a desktop.

Hermes Agent adds per-model provider routing controls

Hermes configs can now pin, prioritize, exclude, or sort providers per model, allowing agent operators to enforce latency, price, privacy, and capability constraints.

Tom Yeh explains the DFT as matrix multiplication

A visual 12-step walkthrough derives the forward and inverse DFT with dot products, transposes, and scaling, then connects the transform to convolutional layers.

TailRL optimizes rare high-reward rollouts in RL pipelines

Tail-Likelihood RL shifts optimization from average reward to upper-tail coverage, using a small advantage-function change that improves returns from extra samples.

Local Inference Lab publishes RTX 6000 Pro serving runbooks

This living wiki collects Docker builds, topology tuning, vLLM and SGLang commands, quantization QA, and regression data for Blackwell RTX PRO 6000 systems.

Mikhail Parakhin uses ChatGPT history for technical learning

Parakhin asks ChatGPT to identify unfamiliar ideas in C++, PyTorch, or ML while considering prior history, turning personalization into a recurring study workflow.

Dingcad launches a prompt-driven CAD workflow for Codex

Dingcad provides a free prompt-to-CAD workflow designed for Codex, with generated models exported into external design tools for rapid mechanical iteration.