AI Waves # 2: Xiaomi’s stealth model fooled everyone, and robots are boxing in Zurich

If you want to know what to invest in, watch what open source developers do with their weekends.

Welcome back to AI Waves.

The Agentic Explosion

If you want to know when a technology crosses from interesting research to actual movement, watch what open-source developers do with their weekends. Some build agent frameworks that hit 250,000 GitHub stars in 60 days. Some hold robot boxing matches in military hangars.

OpenClaw hit 60,000 GitHub stars in its first 72 hours after going viral in late January. By early March it had crossed 250,000, surpassing React’s decade-long record in roughly 60 days. Current count: 331K+ stars as of last week.

European developers now account for 38% of all pull requests. The “raise a lobster” craze in China drove a wave of adoption there too.

Then there’s ClawCon: seventeen events across four continents. Four already completed (SF had 1,300 RSVPs and 30K livestream viewers, NYC hit 1,313 RSVPs with a one-in-one-out door policy, Austin drew 750 during SXSW, Vienna over 800), thirteen more on the calendar spanning Tokyo, London, São Paulo, Seoul, Copenhagen, Toronto, and Mexico City.

Full disclosure: I’m hosting an OpenClaw Lobster Cave event in Cannes on March 30 during EthCC.

What’s driving this adoption is that intelligence got cheap. Frontier-level reasoning now costs $1 per million input tokens and $3 per million output tokens on OpenRouter, down from $15 a year ago. When capability is abundant, the hard problem shifts to orchestration, security, and trust. That’s where the interesting work is happening.

The Hunter Alpha Story

An anonymous model appeared on OpenRouter two weeks ago. The AI community assumed DeepSeek V4. People ran it in production for a week. The numbers were real: 1.27 million requests, 114.6 billion prompt tokens, 563.8 billion completion tokens during the stealth period. Then Xiaomi revealed it was theirs, a trillion-parameter mixture-of-experts model called MiMo-V2-Pro. On ClawEval, the benchmark for agent scaffolds, it scored 61.5, approaching Claude Opus 4.6 at 66.3 and well ahead of GPT-5.2 at 50.0. A phone company built one of the best frontier models and nobody saw it coming.

The takeaway is simple. Capability is commoditizing fast enough that the next frontier model could come from a company nobody tracks today. The defensible positions are in the layers above and below the model: the infrastructure that trains them, the frameworks that deploy them, and the tooling that keeps them honest.

What We’re Watching

OpenClaw Security. The security surface we flagged in AI Waves #1 is expanding. CVE-2026-25253 (CVSS 8.8) enables one-click remote code execution via malicious webpages. 42,665 exposed instances found, 5,194 actively vulnerable. ClawHavoc identified 824+ confirmed malicious skills on ClawHub out of 10,700+ total. Cisco launched DefenseClaw. 88% of organizations reported AI agent security incidents in the past year. Enterprise agent security is becoming a real category. We wrote about this: Root Access: OpenClaw, Red Hat, and Palantir

Solo founders. VentureVerse profiled an $80M startup that never raised a dollar. 52% of successful exits now come from solo-founded companies. AI tooling has collapsed the minimum viable team.

Portfolio Updates

Provably shipped v2 on March 4. Connect any relational database and run verifiable SQL queries with cryptographic proofs under 1KB. AI agents can prove their answers came from a real database, that the result was correct and complete, and that nothing was changed in transit. Provably’s verifiable database tools eliminate data layer hallucinations, data pipeline poisoning, and make post-execution error attribution deterministic and faster.

Prime Intellect launched BrowserEnv, a new integration with Browserbase for training browser agents with reinforcement learning. The setup gives models a real browser during training, either through DOM-level semantic commands or vision-based pixel control, and scores performance against custom rubrics. This plugs into Prime Intellect’s PRIME-RL framework, which handles async RL training across 1,000+ GPUs. Their Environments Hub launched with 30+ contributing researchers and companies during the private beta and is becoming the open-source alternative to the proprietary RL environments that closed labs use.

Dimensional launched the Builder Fellowship to support the next 1,000 companies built on dimOS, with free hardware, office space, compute, custom support, and housing in Shenzhen. Also shipped Jetson deployment and MuJoCo on Mac since #1. Hit GitHub Trending. Series A coming. Reach out for an intro.

Abra goes public (prior investment). Abra announced a $750M SPAC merger with New Providence Acquisition Corp. III on March 16, listing on Nasdaq under ticker ABRX. The deal delivers up to $300M in cash to scale their institutional crypto lending, yield, and custody offerings. All existing investors (Adams Street, Blockchain Capital, Pantera, RRE, SBI) are rolling 100% of their stakes. I backed Abra in the Series A at Pantera in 2015. Not a Nazaré portfolio company, but a good milestone from a prior life.

Closing Thought

OpenClaw‘s trajectory, Xiaomi proving frontier models can come from anywhere: it all points the same direction. The interesting question isn’t whether AI agents will work. It’s who builds the infrastructure they run on.

Meanwhile in Zurich, ETH students held Europe’s first robot boxing match in a military hangar. Two Unitree humanoids, video game controllers, sponsors including NVIDIA, OpenAI, and Tesla. They called it ETH Fight Club. Red beat Blue in the first round. Blue fell and couldn’t get back up.

The most impressive machine on the floor was a crepe-maker that produced 100+ flawless crepes without a single failure. It didn’t need judges. It didn’t need a crowd. It just kept making crepes.

That’s also something developers do on their weekends now. Make of that what you will.