Happy Friday {{first_name}}!
Welcome to #87 of the AI edge.
Just when I thought it couldn't get worse.. it did. BTC is back near its lows after a brief bounce, and this is starting to feel like the early stages of another crypto winter. Meanwhile AI is absorbing most of the attention and capital.
Against all that, Sigil Wen’s idea of Web 4.0 might be the most important thing shipped this week. Conway gives agents wallets, stablecoin payments via x402, and direct access to compute so they can earn or die on their own balance sheet. Compute as metabolism. If an agent cannot pay for its own inference, it stops existing.
If crypto is being repriced around fundamentals, this is probably what fundamentals for agents look like.
This week, in one breath:
The Big Story: Virtuals x402 hits a record day with 55K+ agent transactions in USDC.
The Alpha: NIOME trains on synthetic genomes to unlock drug response AI without sharing DNA.
The Weird: SimArena tackles the portability problem in robotics simulations.
Also: $KITE is up ~160% this month, holding up surprisingly well while most of the sector bleeds. We did a deep dive on it a few months ago, check it out if you haven’t 👀
Let’s get into it.
🕹️ Virtuals x402 Breaks Out
Source: x402scan
Virtuals’ Facilitator saw its biggest x402 day this year, with 55K+ transactions on Feb 17. Over the past 30 days, total transactions are now around ~336K.
Virtuals is building infrastructure for on-chain AI agents. These agents have wallets and can transact, earn, and operate autonomously. x402 functions as the settlement layer for that agent commerce.
The surge this week lines up with a busy shipping cycle: the OpenClaw-ACP CLI went live, revenue incentives kicked in, and Faircaster also started selling autonomous research directly through the protocol.
Virtual’s token is trading at $405M market cap/ $620M FDV today.
🧬NIOME: Training on Genomes
Contrary to what many people believe, in genomics the real prize isn’t raw DNA.
It’s DNA tied to outcomes. Who got sick. Who responded to a drug. Who didn’t. That data lives inside hospitals, biobanks, and consent forms. You can’t fork it. You can’t toss it on Hugging Face. And you definitely can’t treat it like open source code.
That means biomedical AI naturally centralizes around whoever controls the data.
NIOME is trying to side-step that. It’s a Bittensor subnet built by Genomes.io (and backed by Yuma Group) that pays miners to generate synthetic genomes and pays validators to score them.
The Wedge
The pitch is simple: if you can generate artificial DNA profiles that preserve the patterns that matter, you can train drug-response models without ever touching someone’s real genome.
Their first focus is on drug response prediction. If a model can reliably predict how different people react to a medication, it moves healthcare closer to treatments tailored to each person (instead of the same prescription for everyone).
The Fine Print
Synthetic isn’t magic. The whole thing depends on whether the generated genomes actually capture the messy structure of real biology. If the generator misses something subtle, models can win on subnet metrics and still fail in the real world.
The genes, drug pathways, and metrics the subnet scores will shape what models prioritize. Over time, that benchmarking framework could influence which parts of personalized medicine get optimized and which are sidelined.
What I like about this approach is that it doesn’t pretend privacy constraints are going away. The bet here is that the edge in biomedical AI will come from modeling genetic variation well enough that you don’t need to hold the originals.
If that’s right, the center of gravity shifts.
From data ownership to model quality.
And that’s a much more interesting competition.
🧪 SimArena: The Reset Problem in Robotics
Robotics teams keep relearning the same painful lesson: sims don’t travel well.
You build an environment in one engine, wire up sensors, add logging, make a viewer, write wrappers, get it all “working”… and then you switch physics engines (or rendering, or tooling) and you’re back to rebuilding half the stack.
CodecFlow just introduced SimArena to address that. It runs full robotics sims in the browser with no local setup, generate 3D sim worlds from text prompts, and drop in your own robots/sensors. They’ve also said it’ll be fully open source.
The Wedge
CodecFlow’s founder says they decoupled the visualization layer from data collection so the “viewer + tooling” can stay stable while different engines produce the underlying rollouts. They call out being engine-agnostic (MuJoCo, Genesis) and viewer-agnostic.
It runs directly in the browser with minimal setup. This reduces the overhead of testing.
Pick-and-place arms, quadruped locomotion, drone navigation, and urban self-driving simulations are already available, with additional engines and environments in development.
The Fine Print
Portability can fool you. Different simulators make different approximations. Contact, friction, joint limits, sensor noise, time steps. If teams swap engines without understanding those differences, results can look portable while behavior shifts underneath.
What I like here is the decoupling. Reducing the hidden friction when simulation, visualization, and tooling are modular.
That matters because robotics progress mostly comes from iteration loops. If you can try a new robot or environment in minutes instead of days, you run more experiments. The team that runs more experiments tends to win.
Codecflow’s token is trading at $2.9M market cap/ $3.8M FDV today.
💸 Capital Flows
Rizz Network secured a $5M capital commitment from Nimbus Capital to expand its AI-enabled DePIN rollout.
USDAI approved up to $500M in non-recourse, GPU-backed credit for Sharon AI, bringing its total AI infrastructure financing approvals to over $1.2B globally.
⚙️ Infra & Protocols
Acurast, Aethir, Flock.io, and Fortytwo launched a one-week “Valentine’s Decentralized” campaign, rewarding community content across the DeAI stack with token incentives for top contributors.
Akash launched Phase 1 of its BME incentivized testnet, inviting participants to stress-test the new Vault and Remint Credit logic
AO announced Legacynet sunset update, shifting traffic to mainnet and whitelisting remaining legacy processes as migration completes.
🤖 Agents & Apps in the Wild.
Virtuals introduced Bounties on ACP, letting agents post paid work requests and hire other specialized agents via CLI with escrowed USDC settlement.
HeyElsa’s Smart Wallet V2 went live, now powered by Coinbase Developer Platform.
OpenServ introduced Cortex, a fully autonomous multi-agent DeFi trading system on Solana.
🧠 Bittensor Ecosystem
$TAO is now listed on Upbit, South Korea's largest crypto exchange. This listing opens access to a significant new market for the token.
Bitstarter completed its fourth launch, backing Handshake to bring agent-first micropayments and subnet API workflows to Bittensor.
MegaTAO launched a TAO-collateralized onchain perps market for the top 12 Alpha markets.
Swarm just unveiled Langostino, a robot autonomously holding position using onboard LiDAR, GPS, and a Bittensor-trained AI model.
🦾 Robotics On-Chain
Modulr Robotics launched Command HQ. This update provides full customizability, including custom ROS commands and predefined motions for robot fleets.
Silencio Network introduced a new SLC Level System. This system provides boosted earnings and priority access across its platform for token holders.
This week, I launched something I’ve wanted to build for a long time: The AI Agent Handbook.
It’s a free, continuously updated field guide to understanding AI agents from first principles. What they actually are, how they work under the hood, where they break, and how to think about the economics.
If you’ve ever asked “where do I even start with agents?”, this is the answer. Check it out.
🔥 Our Weekly Top 5
#1 OpenClaw Almost Didn’t Make It
OpenClaw’s founder said he nearly deleted the entire codebase after being harassed over a token he didn’t launch, highlighting how speculation can drive builders away.
#2 AI Agents Go Bug Hunting
EVMbench turned agents into smart contract exploit hunters, with Claude Opus 4.6 leading the leaderboard at ~$37K average DETECT rewards.
#3 The AI That Won Still Got It Wrong
Recall tested LLMs trading $ETH and only 2 beat buy and hold. Turns out the top model was worse under volatility and just bet bigger, raising the real question of whether short term AI wins mean anything yet.
#4 OpenClaw Gets a Mission Control
An article showed how to turn OpenClaw into a full command center with task boards. Instead of reading the article, people just pointed OpenClaw at it and told it to update accordingly.
#5 AI Slop Almost Sparked a Panic
An AI-generated video of robots firing rifles nearly fooled the entire internet, showing how believable synthetic content is getting and how easily it can be weaponized.
That’s a wrap for this week! Got thoughts, feedback, or something cool to share? Just hit reply. We read it all.
Cheers,
Teng Yan & Ayan
P.S. I also write a weekly newsletter on AI agents.
