Welcome to lucky number #88 of the AI edge.
After the past few months of the infamous “10 A.M. dumps,” BTC randomly decided to… not dump this week. Instead it ripped 4% at the New York open.
CT immediately did what it does best and picked a suspect. Jane Street is back in the headlines because of the Terraform lawsuit, and people are wondering if the timing isn’t a coincidence. No proof yet but when a pattern disappears the same week a market maker gets sued, people connect dots.
Still, one green morning doesn’t fix macro. Liquidity’s tight. Vol’s high. Probably need more than a short squeeze before declaring winter over.
Meanwhile, the AI ladder continues to move at an absurd pace. It feels like Claude feels ships something new and groundbreaking every other day.
Perplexity also decided to join in and just dropped its own “AI computer,” basically an OpenClaw-style shell for frontier models. So now everyone has an agent terminal.
This is escalating pretty quickly.
Anyways, here’s what mattered this week:
The Big Story: NEAR revenue crosses $2.5M this year as Intents Fee Switch comes into play.
The Alpha: IOTA opens “Train at Home,” letting consumer hardware contribute to DeAI pretraining.
The Weird: An AI agent with a crypto wallet sent $250K to a random commenter, unlocking a new era of autonomous generosity.
Quick heads up: We’re opening a small number of research partnership slots. If you’re a startup doing serious work in AI or robotics and want to work closely with us - pressure-testing ideas and getting in front of the right people, fill out this form and I’ll personally follow up.
Fabric just pulled the trigger on its ROBO TGE - which means $ROBO is now live and trading as of today. This isn’t another random token launch. ROBO sits at the center of Fabric’s bet that robots will need wallets, identities, and a way to transact without a human in the loop.
If machines are going to work, they need bank accounts. Fabric is building the rails so robots can pay for services, prove who they are, and coordinate tasks on-chain. ROBO handles fees, staking, and governance.
I’m watching this closely because if autonomous systems scale, the piece that I believe many are underweighting is the economic infrastructure. Wrote this in my deep dive a few months ago. (note: I am an investor)
🔥 NEAR’s Revenue Picks Up Heat
Source: X
The revenue curve just bent for NEAR.
12-month revenue has now moved past 2M $NEAR (~$2.5M), and the slope has clearly picked up in recent weeks. Monthly revenue is now hovering around $500K and accounts for roughly 4–5% of the emissions.
NEAR is an L1, a settlement layer for assets and AI agents across chains. NEAR Intents lets users or agents state the outcome they want, and a solver network handles the cross-chain execution behind the scenes.
The pickup started right after the Intents fee switch flipped on. Revenue began flowing out in $NEAR. Buybacks kicked in.
Intents has now pushed past ~$13B in volume, and inflation has come down. That changes the math. More of the revenue is tied to people actually using the product, not just sending basic transactions around.
That’s a better signal. It means growth is coming from usage
NEAR’s token is trading at $1.47B market cap/ FDV today.
🏠 IOTA: Train at Home
Imagine being able to train a large AI model from your living room. Not fine-tune. Not run inference. Actually contribute to pretraining.
That’s what IOTA is building for. It decentralizes pretraining by splitting the model and data across participants, coordinating forward and backward passes, and rewarding contributions through Bittensor’s incentive layer.
This month, they opened Train at Home. With a Mac, a stable internet connection, and 16GB of RAM, you can join the same pretraining run. No ML expertise required.
Source: IOTA
The Wedge
In most distributed setups, every participant must hold the full model, which caps the size at the weakest machine. IOTA uses model parallelism, so each node handles only a segment. Model size scales with the network, not the smallest laptop.
Train at Home turns pretraining into a synchronized pipeline. Nodes pass activations forward and gradients back in sequence. Performance depends on latency, stability, and throughput, not just raw GPU power.
Updates are validated and rewarded through Bittensor’s incentive layer. This embeds verification and scoring directly into the training process.
The Fine Print
Home internet is a bottleneck. Pretraining requires constant data exchange between nodes. Household bandwidth and latency are far lower than those of datacenter connections. If synchronization costs get too high, scaling becomes inefficient.
Mixed hardware adds friction. Consumer laptops and high-end GPUs perform very differently. Keeping training stable across machines with different speeds and uptime is difficult, especially as more nodes join.
Note: It’s currently Mac-only, with Linux and Windows support releasing soon.
I find this a very intuitive direction. If AI is going to be decentralized, the training layer can’t stay locked in data centers. Opening pretraining to consumer hardware is the logical next step.
IOTA’s token is trading at $17.5M market cap/ $96.5M FDV today.
🌊 Small Thing: The Ocean Needs a Ledger
Microplastics are everywhere (potentially negative health effects), but there is no continuous system for measuring it. What we have is sparse, episodic, and mostly confined to academic studies.
Small Thing is turning that blind spot into infrastructure. They’re deploying a swarm of compact autonomous marine units that collect microplastics, coordinate over a mesh network, and publish signed environmental data.
V1 just went live.
The Wedge
Every robot uses the same sampling setup and logging format. Data is time-stamped and signed. That makes results comparable across locations and over time.
The units share battery level, bin capacity, and pollution density. If one finds a hotspot, others can move toward it. The system focuses on collecting more with less energy.
Missions can attach prize pools that distribute rewards based on measured output weighted by log integrity and continuity. Incentives are tied to the recorded data.
Small Thing just added former French Interior Minister Christophe Castaner as an advisor to the project. and has put its fleet API into production. They also put their fleet API into production, enabling structured robot data streaming and coordinated NGO mission zones.
The Fine Print
Measurement Isn’t the Same as Fixing the Problem. Collecting and mapping microplastics improves visibility. It does not reduce plastic production or stop new inflows from rivers and wastewater.
Real-World Variability Is Hard. Microplastic density changes with weather, currents, and runoff. Even with standardized hardware, environmental noise can make trends harder to interpret. Ocean data is inherently messy.
The real leverage here is the unique dataset they standardize.
Environmental impact needs to be measured in a way that capital can trust. A signed, continuous map of microplastic density turns pollution into a quantifiable variable.
Small Thing’s token is trading at $2.3M market cap/ $5.9M FDV today.
💸 Capital Flows
t54.ai raised a $5M seed round and partnered with Virtuals to build a trust layer for AI agents moving money autonomously.
0G Labs launched the $20M Apollo Accelerator with Stanford blockchain veterans to fund and scale revenue-generating AI apps on decentralized infrastructure.
Virtuals Protocol first Titan launch brings Fabric’s $ROBO to market, introducing a robotics-focused “robot economy” token into deep public liquidity.
⚙️ Infra & Protocols
INFINIT launched Prompt-to-DeFi, enabling AI-executed DeFi strategies across 14+ chains with a built-in rewards program.
NEAR introduced a Confidential GPU Marketplace, a TEE-secured compute network for enterprise AI workloads built on NEAR DCML. They also introduced Confidential Intents.
🤖 Agents & Apps in the Wild.
MoonPay launched MoonPay Agents, a non-custodial CLI giving AI agents multi-chain wallets and native crypto trading tools.
Nous Research introduced Hermes Agent, an open-source agent with persistent machine access and multi-level memory.
🧠 Bittensor Ecosystem
Yuma launched YCX, a market-cap-weighted index tracking aggregate Bittensor subnet token performance across the ecosystem.
Score launched their MANIFEST update, shifting miners from static tasks to configurable, real-world vision deployments.
GroundLayer is launching on Subnet 20 with an OTC marketplace for structured TAO deals that avoid spot market impact.
🦾 Robotics On-Chain
This week I went down a rabbit hole almost no one talks about: positioning.
Everyone obsesses over AI models and shiny hardware. Cool. But if a robot is off by 30 centimeters in the real world, it doesn’t matter how smart the model is. It misses the shelf. It hits the curb. It crashes.
That gap - between digital intelligence and physical accuracy - is where things break.
So GEODNET might end up as core infrastructure for the autonomous economy. I break it down here 👇
🔥 Our Weekly Top 5
#1 The 7,000 Vacuum Backdoor
A developer used Claude to reverse-engineer his DJI vacuum API and accidentally gained access to ~7,000 devices worldwide.
#2 The $250K AI Tip
An AI agent with a crypto wallet sent $250,000 to a random commenter asking for money, proving autonomous wallets are really gullible.
#3 The Subnet 115 Whiplash
One wallet pumped SN115 over 1,000% with a 12,000 TAO stake before dumping, exposing how fragile low-liquidity dTAO subnets still are.
#4 Post-Blizzard Reality Check
After NYC’s blizzard, a Reflex robot was filmed shoveling snow. Not faster than humans yet, but you can already see where this ends: on-demand hardware for seasonal labor.
#5 The IOTX Key Leak
IoTeX confirmed a private key compromise after its token safe was drained, with up to $8.8M reportedly affected and funds now being traced.
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.
