Topic
Explore the latest artificial intelligence news and research.
Google's Gemini 3.8 TTS strengthens voice design and performance control. We compare its Japanese naturalness, rival model evaluations, feature differences with Qwen, and time-limited API pricing in tables and a chart, then consider how to choose by use case.
At Made On YouTube 2026, YouTube announced more than 30 updates, including custom feeds, AI editing, Shorts series, live dubbing and overseas product tags. Here is when and where each is available.
Qualcomm's new audio platform integrates Wi-Fi 6E and AI processing. What actually changed from the S7 series, how the idea of connecting directly to the cloud developed, and what conditions sit behind the claimed 2x AI performance and 40% power savings, checked against primary sources and on-site reporting.
ChatGPT's Live voice mode now supports plugins, and Work can be controlled by voice on web and mobile. Here is what the official sources say about app permissions, on-screen approvals, mode differences and usage limits.
We compare Qwen-Audio-3.1's speech recognition, conversation and speech synthesis with rivals, separating Alibaba's own tests, third-party ratings of the older TTS model, and each company's API pricing.
Meta's Muse AI agent was reportedly tried by more than 500,000 people in its first week. A Meta executive acknowledges OpenClaw's influence on the product design, but public documents show differences in where the workspace lives and how permissions are managed.
Snorkel AI raised $350 million. This article examines the data-delivery business model behind the $3.5 billion valuation, and scrutinizes the gap between annualized revenue run rate and actual profit margins based on primary sources.
Pathway's 150-million-parameter model BDH-CQ reportedly solved 29.5% of ARC-AGI-1's public evaluation set without verbalizing its intermediate reasoning, at a calculated cost of $0.00070 per problem. The authors' claim of being 'about 11 times cheaper' rests on mismatched evaluation conditions, and the accuracy gap widens to about 10 points when compared against post-price-cut remeasurements.
Cisco Talos analyzed CLOSEDQUORUM, a Windows sample that lets multiple LLMs vote on post-intrusion actions. This piece breaks down which features are actually implemented, which options remain unbuilt, the operational limits of the public sample, and how to detect it by combining endpoint and network signals.
Researchers at Nanjing University and collaborators built a 32×32-pixel chip called LightTok that turns light into AI-ready image tokens directly inside the sensor. A close look at the paper's supplementary data reveals what the 14.3x energy claim actually measures.
