Archive
新模型、開放權重、多模態、語音、影像與模型架構。
AI 研究
The package deletes or truncates older content through an API proxy, bringing rule-based compaction to existing coding agents. The paper’s cost comparison uses an idealized caching model, and the published defaults differ from the experimental settings.
AI 研究
The method retains all recurrent steps, using early exploration to narrow the scope of subsequent attention computation. Performance figures cover only attention workloads; the quality drop at 4K context length and the public reproduction materials still require verification.
AI 基礎設施
BioNeMo combines grouped expert operations, expert parallelism, and MXFP8 to improve throughput for specific training workloads. The fastest configuration changes both precision and parallelism strategy, so its gains still need to be evaluated alongside convergence quality.
GitHub Repo
The new path reuses ggml Metal kernels to run quantized weights within the Transformers workflow. It currently requires installation from the main branch, with limitations in model coverage, batching, and kernel compatibility.
GitHub Repo
When plain-text generation hits the output limit, the new version first tries to free up context and retry. Turns with tool activity are excluded from replay through this mechanism, and the actual recovery rate remains to be measured.
AI 研究
The study uses unlabeled English speech for distillation, partially restoring recognition performance after pruning. The reported 1.75× speedup includes a change in batch size; gains at the same batch size and recognition quality for Chinese require separate evaluation.
AI 研究
The dataset preserves separate tracks for each speaker on a shared timeline for research into full-duplex speech and turn-taking. No human annotation hours are listed for the Chinese subset, and automated transcripts should not be used directly as training ground truth.
推論系統
Tests running DeepSeek-R1 on eight B200 GPUs show that confidential computing increases average per-token latency by less than 5%. The results depend on adjustments to memory transfers, kernel timing, and multi-GPU communication, and still require validation under actual serving workloads.
最新模型
Google introduces Flash and Flash-Lite text-to-speech models with reusable custom voice resources. The update changes prompt structure and the default audio format, while dialogue using custom voices still requires turn-by-turn synthesis.
AI 安全
The vulnerability involves the RPyC service on disaggregated prefill/decode nodes and requires an attacker to reach the control port. A public reproduction shows that health checks can still pass after exploitation; the advisory has yet to list a patched version.
GitHub Repo
The new approach separates portable model layers from hardware-specific optimizations, with the first components integrated into the Transformers backend. Initial H100 tests show throughput close to the existing path, but layer coverage and cross-platform performance still need verification.
推論系統
Physical AI Toolchain adds support for deploying inference on remote GPUs, with two robotic arm examples. Prior research shows that added latency and image bandwidth demands can offset computational gains, so deployments must assess the entire control pipeline.