November 19, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Research Google’s AI agent improved Harness task success eightfold after changing only one line of code, according to an IT之家 report. IT之家 ↗
- Open Source Together AI launched a storage accelerator that caches data from S3-compatible storage and Hugging Face on GPU clusters. Together AI ↗
- Safety Cognition launched Devin Security Threats, a managed index tracking newly emerging vulnerabilities such as Java4Shell and NGINX2Root. Cognition ↗
- Industry Amazon SageMaker AI added detection for latency drift, idle GPUs, and missed optimization opportunities on inference endpoints. AWS Machine Learning ↗
- Funding AI drug-discovery startup Boltz raised $110 million to accelerate small-molecule research. SiliconANGLE ↗
- Agents Amazon Bedrock AgentCore added Strands defaults, built-in tools, prompt caching, context management, and lifecycle hooks for controlling agent tool calls. AWS Machine Learning ↗
- Agents OpenAI said GPT-Live-1.1, already powering voice in ChatGPT, is now available to developers building reliable voice agents. OpenAI Developers ↗
- Models Meta’s Muse Wonder earned the highest overall quality score in Legora’s November BAR benchmark for legal work, according to Meta FAIR. Meta AI ↗
- Agents Sourcegraph reported that 54% of AI-using engineering teams say more than 60% of generated code needs significant human rework before shipping. Sourcegraph ↗
- Research Goodfire released a searchable atlas of language-model neurons, grouping neuron poles by labels inferred from their maximally activating examples. Goodfire AI ↗
- Agents Google previewed Tool Retrieval in the Gemini API, loading tool schemas only when needed to reduce prompt tokens and latency for large agent tool libraries. Google AI ↗
- Industry Pinecone introduced a Cost and Usage Export API that lets enterprise finance teams export project and index spending in the FOCUS format. Pinecone ↗
- Open Source Unsloth collaborated with AMD to bring local model inference to AMD NPUs, claiming two-to-four-times lower power use than traditional GPUs. Unsloth AI ↗
- AIGC Kling 4.0 added native 4K resolution and 10-bit HDR output in both 4K and 1080p modes. Kling AI ↗
- Agents Superhuman unveiled a personal AI assistant and expanded Go across its entire productivity suite. SiliconANGLE ↗
- Industry Sakana AI signed a memorandum with Salesforce to deploy its models in Japanese public-sector software and develop government-focused AI applications. Sakana AI Labs ↗