Articles
Kyutai teaches a speech model to do arithmetic out loud
Kyutai released two open-weight speech-to-speech models that reason through math without transcribing audio first, posting a large jump on spoken GSM8K.
Editor's Picks
Hugging Face ships a $399 duck robot you can retrain
Hugging Face unveiled Microduck, a $399 open source duck robot that learns tricks through reinforcement…
Keenable scores $26M from Accel for an agent-first web index
A Yandex search veteran's startup raised $26M to build a web index tuned for AI…
ChatGPT Work signs into your web accounts all by itself
ChatGPT Work can now store login cookies in its cloud browser and sign into your…
NVIDIA doubles edge AI speed with Jetson Orin Nano 2 module
NVIDIA's Jetson Orin Nano 2 doubles entry-level edge inference and cuts power draw 40 percent.
Latest Articles
Watermarking can loosen a model’s grip on its refusals
New research finds provenance signals change what models do, not just what they say.
Vals wants to rebuild how the industry scores AI models
A 25-year-old founder and $40M from Andreessen Horowitz back an effort to replace benchmarks models have learned to game.
Jina AI fits a document parser into a 3.4B open model
The new jina-ocr-v1 turns PDFs, scans and tables into clean Markdown on a single low-cost GPU.
Linkup ships a sparse retriever that trains on one GPU
SPARSEUP packs a 149M-parameter encoder under Apache 2.0 weights and reports 56.4 average nDCG@10 on BEIR-13.
WSO2 ships a control plane for agents across model stacks
Agent Manager reaches general availability with identity, MCP governance and a sandboxed Kubernetes runtime.
DeepMind opens an institute to host its own AGI disagreements
The new DeepMind Institute will publish essays where its own researchers disagree about AGI.
Anthropic puts a number on how much Claude runs its own research
A prototype index says Claude leads 26 percent of Anthropic's AI research and development work.
Scale AI and Korea build a bilingual safety test that shifts context
ROK-FORTRESS varies language and national grounding to expose what translation-only benchmarks miss.
