News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedPremium subscribers can inspect how Spotify categorizes their music, podcast and audiobook tastes, then request changes in plain language. The beta adjusts Home recommendations within a few hours.
Patreon cofounder Sam Yam will lead OpenAI’s Creator Product group alongside former product and engineering chiefs Drew Rowny and Shannon Ma. The hires hint at new tools, but no product is confirmed.
A background assistant will flag newly relevant back-catalog videos and suggest fresh titles or thumbnails. Creators can also test three cuts of a video, but YouTube has not shared evidence of higher views.
The telecom company is applying AI to customer service, tower planning, maintenance and software. It also expects fewer jobs as cloud networking replaces old hardware, though it says new oversight roles will remain.
The open-source tool gives coding assistants structured cross-file context through MCP instead of repeated text searches. Recent parsers improve Rust, Kotlin, C++, Terraform and mixed documentation support.
Google’s profiler now combines compiler views with live hardware counters for Pallas and other custom kernels. One demo cut matrix-multiplication time 30%, though the sampling tools target TPU v7.
The province alleges OpenAI failed to act on violent conversations before the Tumbler Ridge shooting. Its lawsuit asks for public chat logs, safety changes and payment for a replacement school and care.
Alibaba’s five-model audio lineup covers transcription, speech generation and interruptible voice interaction. It also cuts listed prices by roughly 70% for TTS, 85% for real time and as much as 95% for ASR.
Ema’s agent teams carry out multi-step HR, IT and finance processes across existing software. The startup prices by completed tasks rather than seats or tokens and says it has over one million active users.
Bernie Sanders and Greg Casar propose banning artificial superintelligence and requiring permission for advanced AI development. Company leaders who knowingly violate the rules could face 20 years in prison.
A macOS setting let local software redirect Muse transcription and capture the token controlling a user’s agent account. Meta issued a hotfix after disclosure, but the design shows the risk of privileged assistants.
Nvidia’s open-weight 100-million-parameter model marks who spoke when, including overlapping speech. It led an initial English benchmark, but its labels identify channels rather than people.
Radical Numerics argues that genomic language models must advance for defenders as biological design grows more capable. The same long-context systems that generate DNA could also help evaluate emerging threats.
Modal replaced centralized container coordination with parallel schedulers and worker-owned state. Its benchmark created one million sandboxes in under a minute, with median startup below half a second.
Microsoft says EvilTokens helped compromise 12,000 accounts and used AI to map trusted relationships inside stolen inboxes. A coordinated operation seized 200 domains, and UK police arrested two men.
Google’s new text-to-speech models add custom voice design, line-by-line direction and two-speaker scenes. Voice replication requires recorded consent, and generated audio carries SynthID watermarking.
The San Francisco-area facility will let Claude direct physical drug-research workflows with human safety oversight. Anthropic says it will not run clinical trials for now.
The model leads an open-weight ranking while activating 42 billion of its 1.02 trillion parameters. Xiaomi says reinforcement training cost $2.62 million.
Third-party estimates put Muse at 1.8 million iOS downloads in the US and Canada versus ChatGPT’s 1.3 million at the same stage, though launch conditions differ.
The company wants shared measurements, incident reporting and human-oversight requirements for automated AI research, while acknowledging full recursive improvement does not exist today.
A new independent group hosted at the Institute for Advanced Study will review the importance and release of AI-generated mathematics without decision-making power over OpenAI.
Ron Johnson argues agents can narrow choices but will struggle to take over expensive, tactile purchases where customers want to try devices and seek human help.
The AI infrastructure company reports more than $103 billion in contracts, but roughly 85 percent comes from Microsoft and Anthropic agreements.
An 18-month study found residents unite over water, power, pollution and distrust of industrial promises, while construction unions remain a major source of support.
The Horowitz Andreessen Academy will take 50 recent high-school graduates for a free one-year program in San Francisco starting in fall 2027.
The US and China are expected to discuss AI safety alongside semiconductor restrictions and critical minerals, tying technical policy to a broader trade negotiation.
A WIRED experiment found that bots built from public traces could mimic surface habits and help review pitches, yet invented details and failed to match real editors’ thinking.
Viture’s $299 camera-free glasses use bone-conduction microphones and multiple AI models to organize spoken ideas. Preorders open September 28, with shipping planned for November.
oMLX creator Jun Kim has joined Hugging Face to maintain the Apple-silicon inference server full time. The project will remain Apache 2.0 and community-led.
Amazon now serves xAI’s Grok 4.6 through two Bedrock interfaces, with four reasoning levels and support for long-running agents, coding and visual application work.
An AWS architecture processes body-camera video, audio, images and scanned documents, then exposes the extracted results to Salesforce Agentforce through MCP.
OpenAI’s GPT-6 Sol and Luna are generally available on Bedrock, giving teams lower-cost options for recurring coding, document processing and routing below the Astra tier.
Tata Elxsi’s system filters video at the edge and sends selected frames and metadata to AWS for computer-vision analysis, avoiding continuous raw-video uploads.
Nvidia’s open-weight 30-billion-parameter model activates about 3 billion parameters per token, targeting tool calls and other repeated execution steps inside AI agents.
TypeSafe’s Jev returns typed answers, confidence scores and option probabilities for narrow software decisions, using a training method focused on calibrated reliability.
In OpenRouter’s 3,080-example test, Jev reached 81 percent accuracy versus Claude Opus 5’s 84.4 percent, while answering 13 times faster at one twenty-second of the cost.
The asteroid-mining startup built a transformer-based control stack to diagnose spacecraft problems onboard. It will first fly in shadow mode before a planned autonomous mission.
A startup says a software module derived from lab-grown nerve-cell activity can cut video-model costs by 80 percent. The product is not yet available for independent testing.
Seven new laws create a separate utility rate class for data centers, require them to fund local upgrades and disclose expected water use to local governments.
CAIRN fingerprints AI-connected malware and uncovered CLOSEDQUORUM, which polls DeepSeek, Qwen, Mistral and Gemini before choosing its next action.