Good morning. It's Wednesday, May 20, and we're covering Google’s Gemini expansion, the growing battle over custom AI chips, and Andrej Karpathy’s return to frontier AI research.

Know someone who'd love this? Forward it their way. First time reading? Sign up here to keep up with the future of tech.

Rather than build an agent that needs onboarding, Google unveiled Gemini Spark at I/O 2026 — a personal AI agent that runs 24/7 on dedicated Google Cloud VMs, taps directly into Gmail, Docs, and Slides, and is built on top of an agentic harness from Google's Antigravity team. Spark is positioned to do the work rather than just answer questions: example tasks include parsing credit card statements for hidden subscriptions, and monitoring kids' school emails for deadlines then submitting a daily report to both parents.

Google is redesigning its search box for the first time since 2001 as longer AI-driven queries replace short keyword searches. Powered by Gemini 3.5 Flash, the overhaul adds follow-up chat, photo and video uploads, autonomous search agents, and integrated shopping tools across Search and YouTube. Google says Gemini now serves 900 million users, while analyst Richard Kramer notes the company’s ad business grew with 6% more clicks and 7% higher pricing last year.

Google is turning Street View into an interactive simulation platform that can model weather, lighting, and disaster scenarios in real-world environments. The company unveiled the integration of Street View with Google DeepMind’s Project Genie at Google I/O 2026, extending its world model beyond synthetic game-like spaces. DeepMind researcher Jack Parker-Holder says the system could help train robots for rare conditions, such as sudden sunlight glare on London streets, before real-world deployment.

Andrej Karpathy says he has joined Anthropic, marking a return to hands-on AI research after focusing heavily on education projects. The former OpenAI co-founder and ex-Tesla AI director calls the next few years in large language models “especially formative” for the industry. Karpathy says he plans to resume his education work later while reentering frontier LLM R&D.

Don’t let tech debt kill your startup. Readdy delivers high-fidelity business logic without the programmer’s price tag.

One $15 sub. Massive margins. Ship today.

Use code "forward20" for 20% off — Start building!

An X user named "cprkrn" successfully used Anthropic's Claude AI to recover about $400,000 worth of Bitcoin (5 BTC) that had been locked away for 11 years.

Google’s Gemini chatbot has rebounded sharply from its troubled 2024 launch, reaching 900 million regular users by May 2026 — roughly matching OpenAI’s reported ChatGPT audience, according to Google. The company is weaving Gemini across its ecosystem, from Gmail and Google Docs to Maps and Android, while also preparing to power future versions of Apple’s Siri.

Unlike many AI rivals still grappling with the cost of running large AI systems, Google is already tying AI directly to its advertising business, which generated $77 billion last quarter, up 16%. The shift underscores how Google’s biggest advantage may not just be the model itself, but its ability to place AI inside products billions already use daily. → Read the full article here. (Paywall)

Blackstone plans to invest $5 billion in a new artificial intelligence infrastructure venture with Google, a move CNBC says could strengthen Broadcom’s position in the AI hardware market. Unlike many cloud providers that rely heavily on NVIDIA chips, the new company will use Google’s Tensor Processing Units (TPUs), which Broadcom helps design and develop.

The announcement highlights Broadcom’s growing role behind the scenes of large-scale AI systems as hyperscalers increasingly pursue custom silicon strategies. Broadcom shares fell nearly 2% after the news, though CNBC’s Jim Cramer argued the market may be underestimating how deeply Broadcom is embedded in the expanding AI infrastructure stack. → Read the full article here.

Cerebras says it has begun enterprise trials for Kimi K2.6, a one-trillion-parameter open-weight AI model focused on coding and agentic workflows. According to benchmarking firm Artificial Analysis, Cerebras achieved 981 output tokens per second on the model — roughly 6.7 times faster than the next-fastest GPU-based cloud provider and 29 times faster than Kimi’s official endpoint on a large inference request.

The company credits its wafer-scale hardware architecture, custom kernels, and speculative decoding techniques for the performance gains. Cerebras argues the speed improvements could make agentic coding feel nearly real time, allowing developers to iterate on debugging, refactoring, and full-stack application generation much faster than with current frontier models. → Read the full article here.

Google Launches Gemini Omni: Gemini Omni Flash creates and edits videos from text, images, audio, and clips using conversational prompts.

Google Unveils AI Audio Glasses: Google partnered with Warby Parker, Gentle Monster, and Samsung on Gemini-powered smart glasses for Android and iOS.

NVIDIA CEO Eyes China Reopening: Jensen Huang said he expects China to eventually reopen to U.S. chip suppliers despite H200 approval delays.

NVIDIA Says AI Demand Is Surging: Jensen Huang said AI demand is “utterly parabolic” as Dell and NVIDIA expanded their enterprise AI infrastructure.

SandboxAQ Brings AI Models to Claude: SandboxAQ integrated its drug discovery models into Claude, enabling natural-language simulations.

🟨🟨🟨⬜️⬜️⬜️ Anthropic self-governs (24%)
🟨⬜️⬜️⬜️⬜️⬜️ US government (9%)
🟨🟨🟨🟨⬜️⬜️ International treaty body (25%)
🟩🟩🟩🟩🟩🟩 No one — open-source it (37%)
🟨⬜️⬜️⬜️⬜️⬜️ Nobody should build it at all (6%)

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team