AI news, benchmarks & engineering blog curation
Daily AI industry news — funding, products, policy, and major moves from global AI companies and startups.

Lovable is seeking $300M in funding at a $13.2B valuation. The startup is proving that building software through simple descriptions is a ma

Meta AI glasses will disable cameras if the recording LED is damaged. Prototypes suggest a shift toward continuous audio and photo collectio

xAI released Grok 4.5 with a focus on coding and agentic workflows. The model delivers 110 tokens per second at the lowest cost in its class

Prime Intellect secured $130 million in Series A funding at a $1 billion valuation. The platform provides full-stack infrastructure for doma

cognica contributed Bayesian hybrid search to Apache Lucene 10.5.0. The new Maek app provides a local AI memory workspace for macOS. Maek us

SpaceXAI released Grok 4.5 to optimize autonomous coding agents. The model uses 4.2 times fewer tokens than Opus 4.8 for complex tasks. Trai

Weave Router provides a proxy that automatically routes LLM requests to the most cost-effective model. The tool reduces API expenses by 40%

Developers are replacing traditional search with LLM interfaces. Coding workflows have shifted from writing to design and review. Predictabl

Kokoro is an 82M parameter model providing high-quality local TTS on CPUs. The model is OpenAI API compatible and runs efficiently on legacy

OpenAI released GPT-Live-1 and GPT-Live-1 mini for real-time interaction. The models use full-duplex technology to eliminate conversation la

SpaceX released Grok 4.5 specialized for coding and autonomous agents. The model reduces API costs by 50 percent compared to premium tiers.

Google Photos introduces Video Remix powered by Gemini Omni. The tool offers cinematic relighting and artistic style transfers. Premium AI f

Anthropic released Fable but faced immediate US export controls. The model frequently refuses technical tasks due to over-sensitive filters.

OpenAI launched GPT-Live with a new full-duplex audio architecture. The system separates interaction layers from GPT-5.5 reasoning. New mode

Google updated Android Bench with 8 new models and a more user-friendly framework. It now measures AI performance across 100 dev tasks, incl

GeoSQL introduces a map-in-the-loop system for geospatial LLM analysis. The tool increases geospatial task performance by four times over te

Cognition released SWE-1.7 using a 1 trillion parameter architecture. The model achieves 1,000 tokens per second via Cerebras infrastructure

AI startups are reaching massive revenue milestones in record time. Growth is accelerating as companies scale rather than slowing down. Reve

Mistral released Robostral Navigate for single RGB camera navigation. The 8B model outperforms systems using expensive depth sensors. New tr

Slackbot now integrates Salesforce and Tableau data via MCP. The update shifts AI from personal tools to team orchestration. Over 1,500 engi

Mistral AI released Leanstral 1.5 for formal mathematical verification. The model uses a 119B parameter MoE architecture for efficiency. Dev

SambaNova Systems reached an $11 billion valuation after a $1 billion Series F. The company enables trillion-parameter AI models to run on a

ZML released LLMD to optimize LLM inference across diverse hardware. The tool supports Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc.

Noma Labs discovered the GitLost vulnerability in GitHub Agentic Workflows. Attackers can leak private repository data via indirect prompt i

Red Hat advocates for semantic routing to optimize AI agent token costs. The company warns that AI-driven threats require a 14-day patch cyc

The zkao AI agent identified seven vulnerabilities in Cloudflare's CIRCL library. Technical flaws included float64 precision loss and Rogue

Discord's automated moderation system incorrectly flagged 8,000 user accounts. The AI misidentified benign images like chess boards as prohi

The traditional path of learning to code for a high salary has declined. Coding fosters essential meta-skills like debugging and logical com

Ternlight enables serverless semantic search via a 7MB WASM file. The model uses BitNet b1.58 ternary weights to replace multiplication with

Ilya Sutskever curated a list of 27 essential papers for AI mastery. The curriculum covers everything from Transformers to information theor