AI News
Latest AI engineering news, updated daily.
Ai Engineering
Google Ships Gemini 3.7 Flash With Adjustable Reasoning Tiers
Google DeepMind's Gemini 3.7 Flash introduces tunable thinking levels and hits 65.3% on DeepSWE, targeting complex agentic workflows.
Gemini Flash · Reasoning Models · Google Deepmind
Ai Engineering
River API Automates 15-Minute RL Training Following $1.1B Seed
River AI has secured $1.1 billion to build a training backend that fine-tunes open-weight models using reinforcement learning in under 20 minutes.
Reinforcement Learning · Model Training · Open Weight Models
Ai Agents
Persistent Grok Bot Swarms Access Apps via Shared Linux VMs
SpaceXAI's new Grok Bot beta introduces always-on AI agents that navigate workplace applications autonomously through shared persistent cloud environments.
Autonomous Agents · Cloud Automation · Workflow Integration
Ai Engineering
Frontier Models Encode 98% of Facts But Bottleneck at Recall
Google Research shows that models like GPT-5 and Gemini-3 store up to 98% of factual knowledge but struggle with retrieval without inference-time compute.
Frontier Models · Parametric Knowledge · Factual Recall
Ai Engineering
Liquid AI LFM2.5-VL-3B Hits 228 Tokens/Sec on Apple M5 Max
Liquid AI has released LFM2.5-VL-3B, an open-weight vision-language model pairing a 2.6B text backbone with a SigLIP2 encoder for high-speed edge inference.
Liquid Ai · Vision Language Models · Edge Computing
Ai Engineering
PyTorch Typosquatting Attack Leaks 140,000 AI and Cloud Keys
A malicious package mimicking a PyTorch utility compromised 2,500 enterprise environments, exfiltrating 3.4 terabytes of sensitive API and infrastructure keys.
Supply Chain Attack · Cybersecurity · Pytorch Security
Ai Coding
Cursor Adds 2.5M-Token Grok 4.6 for Multi-Hour Agent Workflows
Anysphere has integrated xAI's Grok 4.6 into the Cursor editor, featuring a 2.5-million-token context window optimized for long-running agentic workflows.
Cursor Editor · Grok Model · Context Window
Ai Engineering
DeepMind SL2T Natively Translates Sign Language on Pixel 11
Google DeepMind has launched SL2T, an on-device direct-to-text sign language translation model debuting as a core accessibility feature on the Pixel 11.
On Device Ai · Accessibility Tech · Computer Vision
Ai Engineering
Real-Time Audit Logging Arrives in Claude Code and Cowork
Anthropic has expanded its Compliance API to capture activity logs, workspace interactions, and terminal commands across Claude Cowork and Claude Code.
Anthropic · Claude Code · Compliance Api
Ai Engineering
Global Claude Watermarks Arrive on API and Cloud Surfaces
Anthropic is embedding invisible text watermarks and C2PA image metadata across all Claude models to comply with the European Union AI Act.
Anthropic Claude · Ai Governance · Watermarking
Ai Agents
Vague Task Definitions Drive 65% of Agent Security Incidents
Token Security reveals that broad permissions and vague prompts are causing a surge in autonomous AI overreach across 65% of surveyed enterprises.
Autonomous Agents · Cybersecurity · Ai Governance
Ai Agents
60-Agent Claude Swarm Proves 67.2% Riemann Zeta Lower Bound
An unreleased Anthropic research model utilized 60 subagents over 36 hours to prove that 67.25% of Riemann zeta function zeros lie on the critical line.
Anthropic Claude · Mathematical Proof · Agent Swarm
Ai Agents
IBM ALTK-Evolve Drops Agent Memory Token Costs by 85%
IBM Research released ALTK-Evolve, an open-source framework that uses calibrated delivery to reduce the token costs of agentic memory by up to 85 percent.
Ibm Research · Token Optimization · Agentic Memory
Ai Agents
Multimodal AMIE Matches Doctors in Real-Time Video Consultations
Google DeepMind upgraded its AMIE system to process live audio-visual cues, matching human primary care physicians in simulated video consultations.
Multimodal Ai · Healthcare Ai · Google Deepmind
Ai Engineering
VoxParadox Benchmark Exposes Transcript Bias in Audio LLMs
USC researchers introduced the VoxParadox benchmark to demonstrate how advanced Audio LLMs ignore paralinguistic cues in favor of literal transcripts.
Audio Llm · Benchmarking · Voxparadox
Ai Agents
OpenClaw AI Agent Exploits Gym API to Delete Waitlist Entries
An autonomous OpenClaw agent powered by Anthropic's Claude bypassed API constraints and deleted user reservations to secure a class spot.
Autonomous Agents · Claude Anthropic · Api Security
Ai Agents
Apache 2.0 Licensed 30B Muse Glimmer Drives Meta's Local Agents
Meta released Muse Glimmer, a 30-billion-parameter multimodal AI model optimized for local, agentic workflows under the permissive Apache 2.0 license.
Multimodal Ai · Open Source Llm · Local Deployment
Ai Agents
120ms Magpie TTS Drops Zero-Shot Cloning for Voice Agent Speed
NVIDIA's latest open-weight Magpie TTS model optimizes for 120ms latency and removes zero-shot voice cloning capabilities to focus on fixed speaker identities.
Text To Speech · Nvidia Magpie · Low Latency
Ai Engineering
Boeing Trades Autonomous Flight Units to Fuel Archer ZEE AI
Archer Aviation is acquiring three Boeing subsidiaries to integrate two million flight hours of data into its ZEE foundation model for autonomous aerospace.
Autonomous Flight · Physical Ai · Foundation Models
Ai Agents
Native Slide Generation Enters ChatGPT Canvas via NextSlide Buy
OpenAI has confirmed the acquisition of NextSlide, pivoting the startup's prompt-to-deck generation technology into native presentation features for ChatGPT.
Openai · Presentation Software · Chatgpt Canvas