Google Ships Gemini 3.5 Flash at I/O as Karpathy Joins Anthropic to Lead Pretraining

Google's I/O conference was the main event this week. The company made Gemini 3.5 Flash generally available, and it's now beating its own Gemini 3.1 Pro on basically every major benchmark, Terminal-Bench at 76.2%, GDPval at 1656 Elo, MCP Atlas at 83.6%, and CharXiv at 84.2%, while running at roughly 4x the speed. They also launched Gemini Omni, which collapses text, image, audio, and video into a single any-to-any model with Google's SynthID watermarking built in. The notable miss was Gemini 3.5 Pro, which Sundar Pichai punted on, asking the audience to "give us until next month to get it to you." One observer pointed out that GPT-5.5 medium is already smarter and cheaper than what Google is shipping, and suggested "it might genuinely be over for anyone not named OpenAI or Anthropic." Make of that what you will.

The open source frontier isn't ceding ground though. Prime Intellect released a training environment with 4,504 tasks that tripled small-model performance on BFCL (a benchmark for AI function calling) through self-play. Odyssey released Agora-1, which lets four players share a GoldenEye-trained deathmatch as a learned game engine. And Cursor's Composer 2.5 wraps Moonshot's Kimi K2.5 with 25x more synthetic training tasks, with xAI now co-training a 10x larger successor on its Colossus 2 cluster.

Google is also tying its science work into one stack. The new Gemini for Science package combines Co-Scientist, AlphaEvolve+ERA, and NotebookLM into a single workflow, with Nature papers and 100+ partners onboard. The distribution side is staggering too. The Gemini app just crossed 900M monthly users, and Google is now processing 3.2 quadrillion tokens per month, up from 9.7 trillion two years ago. That's roughly a 330x increase in two years.

The agentic layer is also moving from the cloud into the actual operating system. Android Halo now pulses on the system tray whenever Gemini agents are running in the background, fed by Antigravity 2.0 (which replaces the old Gemini CLI) and AI Studio's new one-prompt Android app builder. Google's Project Genie can now reskin Google Street View as "Stone Age" or "Ocean World" environments, basically generating playable game worlds on top of real geography. Search AI Mode crossed 1B users, which Google is calling the biggest search box upgrade in 25 years. And the new Universal Cart plus AP2 protocol lets Gemini Spark agents transact across merchants from cloud-based virtual machines that keep running even when your device is off.

A few interesting crossovers and security beats showed up too. OpenAI's verifier adopted Google's SynthID watermark, which is one of the only places the two rivals are converging. Apple, choosing accessibility, made Vision Pro able to steer power wheelchairs by eye gaze. On the flip side, IEEE researchers showed imperceptible audio attacks hijacked 13 different audio language models with 79 to 96% success rates. Amazon's Alexa Podcasts now generates on-demand episodes from AP and 200+ outlets. And Anthropic reversed course this week and told Glasswing partners it "fully supports" them sharing their Mythos cyber findings publicly, which is a meaningful shift toward transparency.

The hardware substrate underneath all of this is straining. Intel is pushing PC manufacturers onto 18A while throttling its older Intel 7 node, and OpenAI launched a new "Guaranteed Capacity" tier that sells multi-year compute lock-ins. Sam Altman warned that the world "will be capacity-constrained for some time." Microsoft shipped Azure Linux 4.0, its first server-grade Linux distribution. Armada raised $230M to mass-produce modular data centers. And Ornn listed the first GPU compute futures on ICE, which is a meaningful step toward compute becoming a tradable commodity. The capstone is Google and Blackstone's $25B joint venture to build out 500 megawatts of TPU capacity by 2027.

The physical world is following the digital one into autonomy. Figure's F.03 humanoid just cleared its seventh straight day of fully autonomous package sorting without failure. The White House also unveiled its new ballroom with a "drone-proof" steel roof that apparently doubles as a landing port for friendly drones. Google's Android XR audio glasses ship this fall, putting Gemini one tap away from the temple. Higher up, SpaceX will buy Cursor about 30 days after its June IPO, exercising a $60B option it already holds. And Astrolight opened a new ESA-backed satellite laser-link station in Greece. Above all of that, the next PURSUE UAP file release is imminent after the last drop pulled 1B views, which prompted Elon Musk's deadpan follow-up: "Where are there aliens?"

The economy is rerouting around all of this in real time. Standard Chartered is shedding 7,000 jobs, with CEO Bill Winters describing it as "replacing in some cases lower-value human capital." Demis Hassabis pushed back on that framing, urging companies to use AI gains to do more rather than to fire people. Minnesota became the first state to criminalize hosting prediction markets like Polymarket and Kalshi (the CFTC sued the state the same day) and the first state to ban AI nudification apps at $500k per violation. Meta is cutting 10% of its staff while reassigning 7,000 employees into "AI native" reorgs.

Some big talent and capital moves too. Andrej Karpathy joined Anthropic this week under Nick Joseph to lead pretraining, which is essentially training Claude to accelerate Claude. OpenAI's Noam Brown reframed the move as frontier labs "collectively advancing the most important tech of our era." Polymarket and Nasdaq Private Market are now pricing Anthropic at a 93% chance of crossing $1T this year and a 69% chance of IPOing before OpenAI does. Demis Hassabis himself was also outed as an early Anthropic angel investor, which is a strange detail given his public posture on safety.

Two more notable beats. The FBI is shopping for nationwide license-plate-reader access. And Iran launched something called Hormuz Safe, which is Bitcoin-backed "shipping insurance" for the Strait of Hormuz, projected to bring in $10B. Finally, Pope Leo XIV's first encyclical, called "Magnifica humanitas," releases on May 25, alongside Anthropic co-founder and interpretability lead Christopher Olah. The Vatican is officially engaging with the AI question.

That's today. More tomorrow.