AI News: April 17, 2026

1. Physical Intelligence Ships π0.7 with Unsupervised Task Generalization

Physical Intelligence. The robotics startup published research on π0.7, a robot brain that handles tasks it was never explicitly trained on by recombining skills learned in other settings — compositional generalization in practice. In the demo, the model made a passable attempt at using an appliance to cook a sweet potato with zero coaching, and performed successfully with plain-language step-by-step guidance. The “verbal coaching” angle is what makes this interesting: it suggests robots could be deployed in new environments and improved in real time without additional data collection or model retraining. Source

2. Factory Raises $150M at $1.5B Valuation for Enterprise AI Coding Droids

Factory. The three-year-old startup closed $150 million led by Khosla Ventures, with Sequoia, Insight Partners, and Blackstone participating. Factory’s pitch targets the full software lifecycle — code generation, testing, review, documentation, deployment — through “Droids” already in use at Nvidia, Adobe, EY, Palo Alto Networks, Adyen, MongoDB, Bayer, Zapier, and Clari. Keith Rabois joins the board. Notable in a week that also saw Cursor adjacent and Cognition Devin-adjacent entries shipping new capabilities. Source

3. Upscale AI Reportedly Raising at $2B Valuation Seven Months After Launch

Upscale AI. The AI networking startup is in talks for a $180-200M third round at roughly $2 billion valuation, Bloomberg reported. Upscale has raised over $300M since launch (a $100M seed in September, $200M Series A in January) and is building SkyHammer, a scale-up networking architecture for AI compute clusters. It has yet to ship a product. Investors include Tiger Global, Xora Innovation, and Premji Invest. Source

4. Adobe: AI Traffic to US Retailers Up 393% in Q1, Converting 42% Better Than Humans

Adobe Analytics. AI-referred traffic to US retail sites rose 393% year-over-year in Q1 2026, and for the first time converts better than non-AI traffic — 42% better in March 2026, flipping from 38% worse a year earlier. Engagement rate is 12% higher, session duration 48% longer, pages per visit 13% more, and revenue per visit 37% higher. The numbers end a long-running debate about whether agent-driven shoppers would ever monetize at parity with humans. Source

5. DeepL Launches Voice-to-Voice Real-Time Translation with Zoom and Teams Add-ons

DeepL. The translation company unveiled a Voice-to-Voice product suite that translates live speech in real time. Early access opens for Zoom and Microsoft Teams add-ons where listeners hear real-time translation while others speak their native languages, or follow on-screen text. Benchmark scores reported 96.4/100 for Zoom and 96.3/100 for Teams on DeepL’s internal quality metric. 40+ languages are supported, including all 24 official EU languages plus Vietnamese, Thai, Arabic, Norwegian, Hebrew, Bengali, and Tagalog. Full rollout starts in June. Source

6. Canva AI 2.0 Turns Assistant into Tool-Calling Agent with Slack, Gmail, and Drive Integrations

Canva. Canva launched AI 2.0 in research preview (initial cap: one million users), adding a model orchestration layer where the assistant autonomously selects and executes Canva’s design tools via function calling. It produces fully editable layered designs without requiring users to pick fonts or arrange elements. Integrations with Slack, Gmail, Google Drive, Calendar, and Zoom let the assistant read email, conversations, and meeting data for context, and a web research skill adds general browsing. Scheduling support runs repeatable tasks in the background. Source

7. InsightFinder Raises $15M for Agent Failure Diagnosis

InsightFinder. The AI observability startup closed a $15M round led by Yu Galaxy, bringing total funding to $35M. Its Autonomous Reliability Insights platform handles the full incident lifecycle — detection, diagnosis, remediation, prevention — for AI and IT systems. Customer list includes UBS, NBCUniversal, Lenovo, Dell, Google Cloud, and Comcast. Revenue grew threefold in the past year. A concrete sign that enterprise agent deployments are hitting the “why did my agent do that?” scaling wall. Source

8. Roblox Assistant Gains Agentic Planning, Procedural 3D, and Self-Correcting Test Loops

Roblox. The platform’s AI assistant picked up a Planning Mode that analyzes game code and data models, asks clarifying questions, and translates prompts into editable action plans. New capabilities include procedural 3D model generation, mesh generation, and self-correcting loops that test outputs and feed results back into planning. Roblox is working on multi-agent parallelism and cloud-based long workflows, plus first-party support for Claude, Cursor, Codex, and other third-party tools in Studio. Source

9. ByteDance Launches Seedance 2.0 Video Model in 100+ Countries, Excludes US

ByteDance. Byteplus rolled out Seedance 2.0 to enterprise customers in 100+ countries but pointedly left out the US, likely due to ongoing copyright disputes with Disney, Warner Bros. Discovery, Paramount Skydance, and Netflix. The model first launched in China in February. Guardrails added for international rollout: realistic human faces can’t be used as source material, filters block copyrighted content generation, and approved customers get access to a 10,000+ virtual-person library. Source

10. Runway CEO Argues for 50 AI Films Instead of One $100M Blockbuster

Runway. Speaking at Semafor World Economy, CEO Cristóbal Valenzuela argued studios should redirect the $100M typically spent on a single film toward 50 smaller AI-assisted productions to diversify hit odds. He pointed to “Bitcoin: Killing Satoshi” — a forthcoming studio-quality AI feature reportedly produced for $70M versus an estimated $300M for a traditional equivalent. Runway’s $5B+ valuation rests on convincing Hollywood this math works. Source

11. Luma and Wonder Project Launch AI Production Studio “Innovative Dreams,” Backed by AWS

Luma Labs. Luma and faith-focused streamer Wonder Project launched Innovative Dreams, an AWS-backed production services company where filmmakers work alongside Luma’s creative technologists and Luma Agents. The first project, “The Old Stories: Moses” starring Ben Kingsley, was shot entirely on a virtual stage and ships this spring on Prime Video. Creative teams direct Luma Agents in real time to adjust sets, props, lighting, and integrate human-actor footage. Source

12. Antioch Raises $8.5M Seed to Close the Sim-to-Real Gap for Robotics

Antioch. The robotics simulation startup closed an $8.5M seed at a $60M valuation led by A* and Category Ventures. Antioch lets robot builders spin up multiple digital instances of their hardware connected to simulated sensors that mirror real-world data. Initial focus is sensor and perception systems for autonomous vehicles, farm and construction machinery, and drones. CEO Harry Mellsop previously worked on Tesla Autopilot; his prior startup Transpose was acquired by Chainalysis in 2023. Source

Alpha Signal. The newsletter featured Architecture Diagram Generator, a Claude Code skill that produces system architecture diagrams from plain-English descriptions as standalone HTML or SVG. The piece walks through setup and demonstrates the tool on custom projects and public GitHub repos — a practical showcase of how the Skills abstraction is being used for genuinely useful developer documentation workflows, not just novelty. Source