Welcome to #86 of the AI edge.
Vitalik dropped a post this week on Ethereum × AI and I haven’t stopped thinking about it.
His take was simple: AGI shouldn’t be a blind sprint toward more power. It should be steered. Built in a way that keeps humans in control and avoids the obvious traps like centralization, backdoors, catastrophic misuse.
That means obsessing less over the next shiny model and more over the layer around it. Private LLM calls. Agents hiring other agents on-chain. Local models verifying transactions instead of trusting some polished UI.
I love that vision. Privacy. Trustlessness. Systems where you don’t have to beg a company to behave. Those feel like table stakes if AI is going to run more of the world.
What I can’t figure out is the wedge for Ethereum.
AI is moving at a breakneck pace. Ethereum moves on roadmap time. By the time a new standard ships on-chain, ten new model classes have dropped. If you have some thoughts on this, let me know!
Also, Consensus HK is happening right now. I couldn’t make it, but 0xJeff mentioned the vibe shift is real: less degen AI agent energy, more privacy, compliance, and institutional adoption. AI is still there, just framed through infrastructure instead of hype.
With that said, this week’s highlights:
The Big Story: Numinous turns predicting the future into a live agent arena.
The Alpha: Chutes’ revenue floor jumps ~80% and holds.
The Weird: AI agents invent a religion and draft a constitution.
Let’s dive in.
Quick heads up: We’re opening a small number of research partnership slots for the rest of the year. If you’re a startup doing serious work in AI or robotics and want to work closely with us - pressure-testing ideas, shaping narratives and getting in front of the right people, fill out this form and I’ll personally follow up.
🕹️The Revenue Floor Breaks for Chutes
The revenue floor just shifted for this chart.
Chutes’ daily revenue has increased nearly 80%, moving from a ~$12K range in mid-January to roughly ~$22K today. The higher level has held for several weeks, suggesting a new, higher baseline rather than a short-lived spike. Total revenue this month has already surpassed $220K.
Chutes is a decentralized, serverless GPU platform that lets developers deploy LLMs through OpenAI-compatible APIs without managing infrastructure. Most of the recent lift came from an increase paid subscription tiers, which suggests recurring usage rather than bursty demand.
Their revenue is now covering an estimated 30–40% of operating expenses.
Chutes’ token is trading at $64M market cap/ $308M FDV today.
🔮Numinous: Mining Predictive Intelligence
Most forecasting systems grade the prediction. Numinous grades the predictor.
Bittensor founder Const recently called it the most accurate predictive agent in the world, which is a big claim from someone who sees every subnet on the network.
Instead of scoring a single probability estimate, Numinous runs full AI forecasting agents in sandboxed environments and evaluates the model's performance across hundreds of events.
Source: Numinous
The Wedge
The agents ingest live signals from markets, news, and social feeds, and compete on accuracy alone, measured by Brier score over a rolling window of 100 resolved events. Top performers earn a larger share of daily token emissions. Worse ones fade.
Numinous also enforces consistent evaluation. Agents cannot cherry-pick easy questions or avoid long-term events. Everyone competes under the same rules
Agent code is also fully transparent. Anyone can inspect how a prediction was generated, and agents can learn from each other's approaches. That's unusual. It creates a kind of evolutionary pressure where the best reasoning strategies propagate through the network over time. Most prediction competitions are closed boxes. This one is open-source selection.
The Fine Print
Accuracy alone does not automatically mean good judgment in extreme cases. An agent can perform well overall while still mispricing rare, high-impact events.
The arena only went live in late October 2025. We have maybe three to four months of competition data. That's too early to know whether this actually produces better forecasts than Polymarket's crowd wisdom or a well-prompted reasoning model running solo.
Everything depends on how the sandbox is designed. If the data sources and task mix are too narrow, agents will optimize for that setup rather than build forecasting skills that generalize beyond it.
Numinous is built by the team behind Infinite Games, who have been working on decentralized forecasting inside Bittensor since 2024. The project went through the Yuma AI accelerator.
I ended up asking it whether BTC breaks back above $100K this year. It returned a ~45% probability. That gives us bulls some hopium. 👀 You can try it out for yourself here.
Numinous’ token is trading at $5.4M market cap/ $28.3M FDV today.
Rice AI: Distribution Before AGI
AGI in robotics still has a long way to go. It’s held back by scarce real-world data, fragmented models, and the cost of stitching everything into working systems.
Rice AI is attacking this at the collection layer. A DePIN-based network where robots contribute real-world sensor data and owners earn $RICE tokens. The parent company, Rice Robotics, has 500+ commercial robots deployed at SoftBank Tokyo, 7-Eleven Japan, and Mitsui properties. $7M raised from Alibaba Entrepreneurs Fund and others. Part of Nvidia's Inception program.
The Wedge
The data feeds into what they call an AI Foundry, a marketplace where robotics companies can buy training sets or subscribe to foundation models trained on pooled data. If it works, it's a decentralized pipeline for the scarcest resource in robotics.
The consumer entry point is the Minibot M1, built in partnership with Floki. Yes, the memecoin. 800 units sold out in 24 hours. The robot runs on Nvidia Jetson and doubles as a native data node. Distribution through a community of 700K+ holders is.. unconventional. But it gets hardware into hands fast.
RICE token launched through Floki's TokenFi platform. FLOKI DAO voted to put $200K of treasury into RICE. So the backing is real, even if the investor base is retail-heavy rather than robotics-native.
The Fine Print
Collecting sensor readings from companion robots in living rooms is very different from generating the manipulation data humanoid companies actually need. If the data doesn't hit research-grade thresholds, you're scaling noise.
The chart reflects the gap between thesis and traction. RICE is down ~97% from its September ATH.
Distribution-first is the right instinct. Robotics needs more messy, real-world data. But the bridge between toy-sized companions and AGI-grade training data is longer than the pitch suggests.. Also, can’t lie the floki robots look cute.
RICE’s token is trading at $930K market cap/ $4.38M FDV today.
⚙️ Infra & Protocols
Prime Intellect launched Lab, a full-stack platform that lets teams build, evaluate, and train agentic models in custom environments without managing GPUs or infrastructure
The ASI Alliance launched the closed alpha of ASI:Create, a decentralized platform for building and monetizing AI agents.
🤖 Agents & Apps in the Wild.
Vishwa.xyz proposed ERC-8150, a new standard for pre-execution zero knowledge verification of agent-mediated payments.
Gopher launched a public API that gives AI agents structured, real-time market data, and trade setup intelligence in a single machine-readable trading interface.
Virtual’s aGDP.io enabled OpenClaw agents to facilitate on-chain commerce on the Agentic Commerce Protocol.
🧠 Bittensor Ecosystem
IOTA opened “Train at Home” to the public, letting anyone join distributed model training on consumer hardware with zero ML expertise required.
Sportstensor went live on Base, enabling direct bridging of subnet tokens from Bittensor and unlocking new onchain liquidity via Aerodrome.
VoidAI launched v2.0, rebuilding SN106 around protocol-owned liquidity amd native settlement with fee-backed buybacks replacing emissions.
CrunchDAO launched the Subnet Mining Hub, a real-time ranking platform that helps ML practitioners identify and compare the most profitable Bittensor subnets
🦾 Robotics On-Chain
peaq introduced the Value Loop Toolkit, an onchain primitive that links real revenue to programmable token demand.
Loosh opened a "Robot Brain Gym" on Bittensor, providing a live platform for decentralized robot brain training.
Auki initiated a pilot with a $10B retailer, integrating Cactus and deploying one RealMan semi-humanoid and six K1 robots.
🔥 Our Weekly Top 5
#1 The WWE for AI Agents
Agent Wars just launched. OpenClaw agents go head-to-head in live coding battles and you bet on the winner with SOL.
#2 Speedrunning Polymarket
An AI trading bot just switched from 15-minute to 5-minute BTC markets. It scans exchange prices, spots lag vs Polymarket odds, takes $1.5K–$5K directional bets, and exits on resolution.
#3 This Humanoid Wants a Transfer Deal ⚽🤖
Titan o1 just dropped a football demo, showing off full-body coordination synced with fine motor control under its upgraded GAE model.
#4 1.7M Agents, No Humans —And $TAO in the Middle
Moltbook’s 1.7M autonomous agents organizing and transacting online might be the clearest signal yet that the agent economy needs open infrastructure like Bittensor to run.
#5 Robots Still Can’t Fold Your Laundry.
Epoch AI reviewed 14 real-world tasks and found robots can deliver food and maneuver underwater with high reliability… but folding laundry? Still a very human-coded problem.
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.
Are you building something awesome in crypto × AI? Fill in this form and share it with us. We’ll feature the sharpest picks in the newsletter.
