AI News: May 14, 2026

1. Notion launches a developer platform built around external agents and Workers

Notion. Notion turned its workspace into an orchestration layer for AI agents: the new Developer Platform adds Workers (a sandboxed cloud runtime for custom code), database sync, an External Agent API, and direct support for partner agents including Claude Code, Cursor, Codex, and Decagon. Customers can now have agents read and write across Notion databases while triggering external systems — Notion’s pitch is to be the control plane for multi-tool agent workflows rather than a destination. Background: Notion’s Custom Agents launched in February 2026 and customers have built more than a million since. Source

2. Luma opens the Uni-1.1 image model API at $0.04 per image

Luma. Luma Labs opened API access to Uni-1.1, its latest image model, at $0.04 per image — a price tier matching OpenAI and Google’s image APIs. Uni-1.1 ranks third on the relevant image-generation Arena leaderboards behind Google and OpenAI offerings, putting Luma into the public-cloud image-gen tier alongside the larger labs. Worth a look if you’ve been pinned to a single image provider and want competitive pricing leverage. Source

3. Tencent telegraphs an H2 AI infrastructure ramp on improving domestic chip supply

Tencent. Tencent signaled it intends to substantially raise infrastructure investment in the second half of 2026, citing improvements in domestic chip manufacturing capacity. The signal lands as Chinese hyperscalers continue navigating US export controls, with domestic Ascend-class accelerators picking up share inside large CN training runs. The market read: more compute coming online for Chinese-trained models, with ripple effects on the open-weights leaderboard. Source

4. Amazon merges Rufus and Alexa+ into Alexa for Shopping

Amazon. Amazon retired the Rufus chatbot brand and launched “Alexa for Shopping,” a single agentic shopping assistant that combines Rufus’s product knowledge with Alexa+‘s personalization across devices. Customers can ask questions in the main Amazon search bar, build personalized shopping guides for big purchases, see up to a year of price history, and let the agent build carts or place orders. Notably, the assistant also reaches off-Amazon stores via a Buy-for-Me flow — the deepest yet from Amazon into general-purpose commerce agency. Source

5. Origin Lab raises $8M to license game telemetry for world-model training

Origin Lab. Origin Lab raised an $8M round to operate a marketplace where video game studios sell gameplay telemetry and recorded sessions to AI labs training world models. The bet is that the next generation of physical and embodied models needs large, licensed corpora of human interaction with rich 3D environments — exactly what AAA game studios already log. It’s a notable data-economy play in the world-model arms race. Source

6. Adaption launches AutoScientist for model self-training

Adaption. Adaption launched AutoScientist, a tool that automates the fine-tuning loop so a model can keep improving on a target capability with minimal human-in-the-loop intervention. The pitch is closing the gap between “I have a base model” and “I have a specialized model” without standing up an internal RL/eval team. Useful framing for application teams who want continual learning without owning the training stack. Source

7. Poppy debuts a proactive personal-AI assistant

Poppy. Poppy launched a personal assistant that integrates calendar, email, and messaging to surface reminders and task suggestions without explicit prompting. It’s the latest entrant in the “proactive AI” category — the same theme Anthropic’s Cat Wu flagged the same day as the next big assistant frontier — but aimed at consumers rather than enterprises. Worth tracking as a comp for any team designing proactive features on top of LLM platforms. Source

8. AgentMemory ships a local triple-stream memory layer for coding agents

AgentMemory. AgentMemory is a local service that replaces static memory files (CLAUDE.md, AGENTS.md) with a searchable database for AI coding agents, capturing observations through 12 lifecycle hooks. The triple-stream retrieval pipeline reports 95.2% R@5 on LongMemEval-S — competitive with much larger remote memory systems — and ships as a one-npx install that organizes memories across four tiers. It exposes an MCP server with 51 tools, though context injection is off by default so you opt in to what gets pulled into the model context. Source