Large Language Models
Climbers rescued after relying on Google Gemini for Mount Shasta trek
Climbers on Mount Shasta had to be rescued after Google Gemini said the hike would take eight hours; it took 16, leaving them without supplies in the dark.
Latest news, developments and reporting on Large Language Models.
Climbers on Mount Shasta had to be rescued after Google Gemini said the hike would take eight hours; it took 16, leaving them without supplies in the dark.
An Android Authority writer describes how switching to paid Kagi search, Proton email and Firefox helped reduce Google dependence far more easily than anticipated.
Perplexity has open sourced Lily, a Rust and Metal inference engine for Apple silicon that runs Qwen3.6-35B-A3B and delivers faster prefill and decode than MLX-LM, according to MarkTechPost.
据雷峰网报道，长安汽车已在研产供销服等场景接入千问办公，一线员工用AI开发工具，效率显著提升。
Google launches Gemini 3.8 Flash and a cybersecurity model, wins a court ruling against forced ad-exchange sale, and gains Berkshire backing after its longest monthly losing streak in over a decade.
South Korea selects a Naver Cloud-led consortium to build a sovereign cybersecurity AI foundation model, beating SK Telecom, with 256 Nvidia B200 GPUs and tests across seven critical sectors.
Capsule Security released a detection system built on fine-tuned Nvidia Nemotron models to block rogue AI agents in real time, claiming 98% accuracy on a benchmark.
CrowdStrike opened a cyber superintelligence lab for AI-driven defense; internal tests showed 29% higher detection and sixfold faster remediation.
Google’s new Gemini for Home update lets users combine media and smart-home voice commands, and fixes alarm and Keep issues. Separately, one user reports Gemini gave them an unwanted nickname, “Fish Lemon.”
Microsoft will begin reporting Azure quarterly revenue and consolidate into two business segments, giving investors clearer insight into AI-driven cloud growth.
Meta released Muse Spark 1.3 on Sept. 2, claiming it matches or beats Anthropic and OpenAI. Independent benchmarks put it behind only two Claude models.
Google unveiled Gemini 3.8 Flash, its third Flash release in six weeks, alongside a gated cybersecurity edition, Gemini 3.8 Flash Cyber.
OpenAI's Astra faces safety alarms over opaque reasoning; open-source AURA offers human-gated oversight.
At Fal.Con, CrowdStrike unveiled SafeMind, a security model family built with Nvidia, while CEO George Kurtz argued that security must move from a stack layer to an AI control plane governing enterprise agents.
A Juniper Research report says Chinese AI models cost up to 90% less to run, eroding US market share on OpenRouter as price-driven and quality-driven AI markets split.
ByteDance reorganized its Seed AI unit in August, creating four departments for pre-training data, reinforcement learning and product post-training for workplace and chat. The shift signals a task-oriented approach as AI competition moves to agentic and omni-modal development.
ZeroDrift today introduced Guard for Agents, a service that converts written company policies into enforceable rules to check AI agents' outgoing messages before they reach customers, with API and connectors for document platforms.
World Labs’ Atlas generates bullet-time visuals from just a few cameras, moving spatial intelligence closer to content creation and robotics.
Google is reportedly close to releasing Gemini 3.8 Flash, a new AI model with significantly improved coding capabilities, potentially rivaling Anthropic and OpenAI offerings.
World Labs, founded by Fei-Fei Li, released Atlas, the first multimodal world model, enabling controllable video generation and 3D reconstruction from just a few photos.
Anthropic has unveiled Enterprise Frontier Safeguards, an architecture that stores monitoring data in customer-controlled cloud infrastructure, aiming to deliver zero-data-retention privacy while detecting cross-session misuse.
Alibaba's Qwen3.8-Max update tops the global front-end coding leaderboard CodeArena with 1,691 points and a price of $5 per million tokens.
Gemini now offers agentic video understanding, with up to 88% lower token usage and 7% better accuracy, per Android Authority.
Meta's new Muse Voice Transcribe unifies streaming ASR, diarization for 20+ speakers, and endpointing in one API model, posting a 3.1% WER at $3 per 1,000 minutes.