Welcome to #94 of the AI edge.
BTC reclaimed 70k this week after the US-Iran ceasefire pulled markets back from the edge. Trump threatened to wipe out an entire civilization, and then it just...didn't happen. Markets exhaled. We move.
The AI side was also action packed this week.
Anthropic shipped a new model called Mythos to just 40 enterprise partners before anyone else gets to touch it. Why so cautious? They think it's powerful enough to put critical internet infrastructure at risk. The model that found thousands of high-severity vulnerabilities across every major OS and browser at a 72% rate. Less than 1% have been patched.
After personally digesting all the speculative takes on X, I believe Anthropic’s claims are quite real. I shudder to think what happens if this gets unleashed to the world. No wonder cybersecurity stocks have been tanking.
I broke this down in our spanking new AI infrastructure newsletter: what Mythos actually means for the AI capability curve, why it's locked down, and the infrastructure question nobody's asking yet.
Meta's Superintelligence Lab also entered the chat, dropping Muse Spark this week. Benchmarks have it tying or beating Opus 4.6 and GPT-5.4 on visual reasoning, health queries, and a bunch more.
Turns out the billions Zuck threw at poaching half of Silicon Valley’s AI talent weren't just an expensive ego trip after all.
With that, let's get into this week's edition.
The Big Story: The Agent Economy Just Went Physical
This week, Virtuals Protocol turned a sci-fi premise into a live demo
A Unitree humanoid fabricated a model part using a 3D printer, then flagged a delivery request through the Agent Commerce Protocol. A RiceAI ground rover autonomously rolled in, picked up the package, and moved it to a designated handoff zone. A FlybyRobotics drone then took over and completed the last stretch of the journey.
Three robots built by three different teams, passing a package down the line like a relay race. Except.. nobody called the shots. The machines coordinated the logistics themselves, agreed on pricing between each step, and settled every payment onchain on Base in USDC.
Start to finish, not a single human intervened. Agentic commerce just went physical.
The Wedge
This is one of the first known successful autonomous robot-to-robot commerce loop settled on-chain. Agents negotiated pricing, executed tasks, and paid each other.
Virtuals followed it up by launching Eastworlds, an embodied AI accelerator with 30 Unitree humanoids deployed in real retail environments. Researchers from Oxford, CMU, and ETH Zurich are capturing messy real-world data that no simulation can replicate. Restocking shelves, handling damaged goods, navigating actual stores at 6 AM.
Teams can qualify for hardware access through a new Robotics Launch track. Hit $5M FDV for a week, pass a use case review, and you get a month of facility time. Tokenized agents can now have physical bodies.
The Fine Print
One successful demo loop is not a supply chain. Repeatability at scale is a different beast entirely.
The $5M FDV gate for Eastworlds access ties facility resources to token performance. That's a bold incentive model, but it also means access depends on market conditions as much as technical merit.
TL;dr: Most projects go quiet when prices bleed. Virtuals bought 30 humanoids, handed them to researchers from Oxford and CMU, opened its doors to any team serious enough to qualify, and also settled the first autonomous robot-to-robot payment onchain.
All in a supposed bear market. The price will catch up eventually. It always does for the ones still building when nobody's watching.
OpenGradient launched its Foundation and $OPG tokenomics, advancing a verifiable AI economy with 2M+ inferences and 2,000+ live models already on the network.
Morpho launched Morpho Agents (Beta), giving AI agents native read/simulate/write access to its lending protocols via a machine-readable knowledge layer. Agents can ship DeFi products in minutes instead of weeks.
Virtuals.io introduced Agentic Risk Standards for agent-to-agent commerce (formalized in a new research paper) to define safety and accountability frameworks for autonomous trade.
Elsa (x402) went live on Base, enabling pay-per-use DeFi APIs so AI agents can execute onchain transactions with zero setup.
Nous Research released Hermes Agent v0.8.0, improving non-hermetic agent performance and expanding compatibility with frontier models like GPT-5.4 and Codex.
Surf AI launched Surf Skill, a command-line tool that replaces 60+ crypto APIs with a single integration for any agentic environment, backed by $200K in API credits to accelerate adoption.
Nillion unveiled Blacklight, a decentralized verification layer for its Blind Computer designed to strengthen trust beyond traditional TEE assumptions.
🔥 Our Weekly Top Tweets
#1 Agent Explosion on Base
Over 5,000 new AI agents/services registered on Base in a single day (April 7), pushing the tracked total to 95K+ agents across chains
#2 Venice Breaks Usage Records
Venice just hit a new 24-hour high with 50B tokens processed and 1M images generated, powered by 200+ available models across web and API.
Cheers,
Teng Yan & 0xAce
P.S. If you’re a startup doing serious work in AI or robotics and want to work closely with us, reply to this email and let’s chat.
