Good morning. It's Thursday, April 16, and we're covering AI’s uneven intelligence, the rise of a compute-powered economy, Google bringing AI directly into the desktop experience, and more.

If you enjoy this email, share it with a friend. First time reading? Sign up here to keep up with the future of tech.

Cadence, NVIDIA Partner to Train Robots
Instead of relying on scarce real-world data, Cadence Design Systems and NVIDIA are combining physics simulation and AI to generate synthetic training environments for robots. The partnership integrates Cadence’s physics engines with NVIDIA’s AI models to improve how robots learn tasks in virtual settings. Executives said more accurate simulated data can significantly reduce training time and boost real-world performance.

Google’s Gemini Desktop App
Instead of living in a browser tab, Gemini now sits a keyboard shortcut away on macOS, signaling a shift toward ambient, always-available AI. Google has launched a native desktop app that integrates system-wide, allowing users to access AI help without switching windows. The app supports screen sharing for real-time context, works on macOS 15 and newer, and is available globally at no cost starting April 2026.

OpenAI Agents SDK for Long-Running Tasks
Instead of relying on loosely connected tools, OpenAI’s updated Agents SDK introduces a built-in system where AI agents can safely run code, edit files, and operate across extended workflows in controlled environments. The release adds native sandbox execution and a model-aligned “harness” that standardizes how agents interact with files, tools, and memory. The SDK supports integrations with platforms like AWS S3 and Cloudflare.

Meta Commits to Custom AI Chips
Rather than relying solely on GPUs, Meta is doubling down on its own AI silicon, committing to gigawatt-scale deployments of custom chips built with Broadcom through 2029. The expanded partnership centers on Meta’s MTIA accelerators, designed for training and inference across its growing AI infrastructure. Meta plans an initial 1-gigawatt rollout, scaling to multiple gigawatts by 2027, with next-generation chips built on a 2-nanometer process. The move comes as Meta targets up to $135 billion in AI spending in 2026.

After selling its footwear brand and assets to American Exchange Group for $39 million in March 2026, Allbirds is rebranding as NewBird AI and pivoting to AI infrastructure. The company announced a $50 million convertible financing facility from an undisclosed institutional investor and plans to offer GPU-as-a-service and AI-native cloud solutions.

The move allows the Nasdaq-listed shell (ticker: BIRD) to redirect into the high-demand AI compute market. The transition remains subject to shareholder approval, with a vote scheduled for May 18, 2026, and a potential dividend payout expected in Q3 if approved. → Continue reading here.

A growing group of researchers and economists argue that artificial intelligence is best understood as “jagged intelligence,” a term coined by Andrej Karpathy to describe systems that excel in some tasks while failing at others. Recent examples highlight the contrast: AI models solved five of six problems at the 2025 International Math Olympiad but struggled with simple real-world reasoning questions.

The concept is reshaping how experts evaluate AI’s impact on jobs, suggesting automation will target specific tasks rather than entire professions. While capabilities are improving rapidly, uneven performance makes it difficult to predict where systems will succeed or break down. → Read the full article here. (Paywall)

Greg Brockman says the global economy is shifting toward a “compute-powered” model, where access to AI and processing power increasingly determines productivity. He argues that AI is already accelerating software engineering and is poised to transform all computer-based work by reducing the need for step-by-step human instruction.

Tools like ChatGPT and Codex are enabling individuals and small teams to build faster and tackle more complex projects, lowering barriers between ideas and execution. Brockman notes this shift could disrupt job stability and institutions, while also expanding entrepreneurship and creative output. He emphasizes that how these systems are deployed—and who benefits—remains an open question. → Read the full article here.

Snap Cuts 1,000 Jobs Citing AI: Snap Inc will cut 16% of its workforce after pressure from activist investor Irenic Capital and a 30% stock drop, citing AI-driven efficiency gains and halted hiring.

Claude Adds Multi-Session Workspace: Redesigned Claude Code brings side-by-side sessions, integrated terminal, and previews, aiming to streamline developer workflows with customizable layouts and native plugin support.

Emergent Launches Wingman AI Agent: India’s Emergent unveils a messaging-based autonomous agent that executes tasks across apps via chat, targeting workflow automation as competition intensifies in the AI agents market.

Duolingo Drops AI Usage Metrics: Duolingo will no longer evaluate employees on AI usage after backlash and confusion, shifting performance reviews to outcomes as CEO Luis von Ahn admits overemphasis on AI adoption.

Adobe Unveils Firefly AI Assistant: Adobe launches an agent that runs Photoshop, Premiere, and more from one prompt, coordinating 100+ tools to automate workflows as it pushes deeper into AI-driven creative software.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team