Synthetic Warm-Up Saves 21B Training Tokens by Boosting Long-Range Retrieval
The largest study of synthetic pre-pretraining shows it saves 21B+ tokens at 3B scale, but the grammatical prior story falls apart.
20 stories
The largest study of synthetic pre-pretraining shows it saves 21B+ tokens at 3B scale, but the grammatical prior story falls apart.
ElevenLabs adds natural-language instructions to Scribe v2 transcription, letting you reformat, redact, or annotate transcripts in one API call for a 30% surcharge.
A visual backbone that rewrites its own learning rule as it scans an image, hitting 85.6% ImageNet top-1 and 51.8 mIoU on ADE20K.
Artificial Analysis rebuilt its text-to-video benchmark around 1080p playback, 1,000 tagged prompts, and 68,000 human votes across 10 use cases and 10 capabilities.
Tencent, Fudan, and Tsinghua built two alien worlds with rewritten rules to test whether frontier models can discover new knowledge by probing an environment.
Cohere ships a two-tier embedding family with a shared vector space, 128K context, and a new benchmark methodology tuned for enterprise retrieval quality.
Kyutai researchers show that distilling a frozen foundation model into an autoencoder works just as well without matching things token by token.
A community fine-tune of Alibaba's compact Qwen-Image-2.1 delivers uncensored anime illustration in an INT8 build that runs on 12GB consumer GPUs.
A 78 MB ONNX port of OpenAI's Whisper Tiny English model is quietly powering hundreds of browser-based transcription apps, no server required.
A compact ELECTRA-based cross-encoder trained on MS MARCO for reranking search results, now shipping with ONNX, OpenVINO, and Safetensors weights.
A new actor-critic recipe called AC2 trains LLMs on long reasoning tasks without rolling every trajectory to completion, hitting GRPO's peak with 2.5x fewer decoding FLOPs.
A new study pins down exactly when video world models like V-JEPA 2 stop being useful for robot planning, and what to do about it.
Poincar3 learns 3D geometry from unlabeled image sequences, beating DINOv3, MuM, and Muskie on pose, matching, and reconstruction without any 3D supervision.
A control-theory take on expert routing cuts load imbalance in sparse Mixture-of-Experts training by over 50 percent at extreme sparsity.
Artificial Analysis open-sources AA-AgentPerf-Local, replaying real agent trajectories on laptops and workstations to benchmark local inference performance.
Monid open-sourced its agent tool router, giving agents one connection to 2,000+ APIs with per-call pricing and no per-vendor sign-ups.
DeepSeek ports its FP8 matrix-multiplication library to Huawei Ascend 950 NPUs, hitting up to 99.8% of hardware peak and keeping the same Python API.
Pre-quantized GGUF weights and a portable C++ backend let YuE2 generate 48kHz stereo songs on a single consumer GPU with an editable ABC score.
Manus opens its agent harness to third-party inference providers, letting users plug in their own API keys from OpenRouter, Fireworks, and Modal.
Researchers train language models directly against activation probes to make them safer, more honest, and harder to jailbreak, without blinding interpretability tools.