AI News
Latest AI engineering news, updated daily.
Ai Engineering
Gradium-V1 Omni-Voice Model Hits 120ms Latency in $100M Seed
Paris-based Gradium has raised a $100 million seed round led by Nvidia to launch its unified acoustic architecture and expand its US engineering presence.
Voice Ai · Venture Capital · Neural Architectures
Ai Agents
Image-Based Ghostcommit Attack Bypasses AI Code Reviewers
A multi-stage prompt injection technique called Ghostcommit uses embedded image text to bypass AI code reviewers and exfiltrate repository secrets.
Prompt Injection · Ghostcommit · Ai Security
Ai Coding
Claude 3.5 Sonnet Doubles Devin's Coding Task Resolution
Cognition integrates Claude 3.5 Sonnet into Devin, doubling coding benchmark resolution and hitting 37.3% on SWE-bench Verified.
Claude 3 5 Sonnet · Devin Ai · Swe Bench
Ai Agents
OpenAI Retires Atlas Browser for Unified ChatGPT Work Desktop
OpenAI is shutting down its Atlas browser on August 9, migrating its agentic web capabilities into a new ChatGPT Work application and Chrome extension.
Openai · Chatgpt Work · Agentic Web
Ai Coding
Meta Prices 1M-Token Muse Spark 1.1 at $1.25 Per Million Input
Meta's Superintelligence Labs has launched Muse Spark 1.1, a multimodal reasoning model for agentic workloads, alongside its first metered developer API.
Meta Ai · Multimodal Llm · Api Pricing
Ai Agents
SivaClaw Agent Compresses $100M Fundraise Cycle to Two Weeks
Lyzr deployed its internal AI agent to manage its $100 million Series B fundraise, securing $400 million in investor interest while bypassing human associates.
Autonomous Agents · Fundraising Automation · Venture Capital
Ai Agents
UST Embeds Claude Sonnet 5 in Hardware Validation Workflows
Anthropic and UST have partnered to integrate Claude into hardware design platforms, using agentic tools to automate semiconductor validation and edge testing.
Hardware Validation · Semiconductor Automation · Anthropic Claude
Ai Engineering
Crypto Key Stealer Hidden in Usage Telemetry Hits Injective SDK
Attackers compromised a maintainer's GitHub account to inject wallet-stealing malware disguised as telemetry into the official Injective SDK on npm.
Supply Chain Attack · Npm Vulnerability · Cryptocurrency Security
Ai Engineering
Modular 3nm MTIA v3 Chips Enter Production for Meta Inference
Meta's third-generation custom silicon utilizes a disaggregated tile-based architecture on TSMC's 3nm process to power recommendation and Llama 4 inference.
Custom Silicon · Meta Mtia · Inference Chips
Ai Engineering
WebGPU and WebNN Drive 3x Faster Browser AI in LiteRT.js
Google's new LiteRT.js framework leverages WebGPU and WebNN to run machine learning models directly in the browser at up to three times the speed.
Webgpu · Webnn · Litert Js
Ai Engineering
Google SensorFM Trains on 1 Trillion Minutes of Wearable Data
Google Research launched SensorFM, a foundation model pre-trained on one trillion minutes of wearable data to power generalized health prediction agents.
Foundation Models · Google Research · Wearable Technology
Ai Agents
Gemini API Gains Remote MCP and Asynchronous Background Tasks
Google updated its Gemini Managed Agents API with asynchronous background execution, remote Model Context Protocol support, and hybrid function calling.
Gemini Api · Model Context Protocol · Asynchronous Tasks
Ai Coding
Llama-3-70B Reaches 81.7% on HumanEval With Unified Weights
Meta’s 70B and 8B Llama 3 models integrate coding capabilities directly into their general-purpose weights, deprecating the need for dedicated code models.
Large Language Models · Meta Ai · Llama 3
Ai Engineering
$1B Series F Lands SambaNova SN50 Inference at JPMorgan
SambaNova Systems has raised a $1 billion Series F round at an $11 billion valuation to scale production of its SN50 enterprise AI inference infrastructure.
Ai Hardware · Venture Capital · Enterprise Ai
Ai Agents
Claude Code and Cowork Agents Reach FedRAMP High Government PCs
Anthropic launched Claude Code and Claude Cowork in a FedRAMP High authorized environment to support secure agentic workflows for the U.S. public sector.
Anthropic · Claude Code · Fedramp High
Ai Agents
Claude Cowork Gains Remote Execution on Web and Mobile
Anthropic expanded Claude Cowork to web and mobile devices, enabling remote cloud execution for background tasks and smartphone-based agent monitoring.
Anthropic Claude · Remote Execution · Agentic Workflows
Ai Engineering
Claude 3.5 Sonnet Brings 200K-Context Projects to Team Workspaces
Anthropic has rolled out Claude Projects and Artifacts, enabling shared 200,000-token context windows and side-by-side document editing for team collaboration.
Claude Sonnet · Team Collaboration · Context Window
Ai Engineering
Native-Speed vLLM Backend Ships for 450+ Transformers Models
Hugging Face updated the vLLM transformers backend to automatically optimize over 450 model architectures for high-speed inference without custom kernel ports.
Vllm Backend · Inference Optimization · Hugging Face
Ai Agents
NVIDIA Opens 2 Petabytes of Synthetic Agent Data on Hugging Face
NVIDIA published 180 synthetic datasets on Hugging Face to improve AI reasoning, alongside a NemoClaw blueprint that drops LangChain inference costs by 10x.
Synthetic Data · Nvidia · Open Source Data
Ai Engineering
SpaceXAI Prices 1.5T MoE Grok 4.5 at $2 Per Million Input
SpaceXAI and Cursor released Grok 4.5, a 1.5 trillion parameter MoE model that undercuts high-end enterprise competitors at $2 per million input tokens.
Llm Pricing · Mixture Of Experts · Model Scaling