Happy Friday {{first_name}}!
Welcome to #83 of the AI edge.
The InfoFi era effectively died this week. All it took was one tweet from Nikita (head of product at X). I’d be lying if I said I didn’t see this coming. The AI spam on X was getting out of hand. I couldn’t tell who was real anymore, including in my own reply feed.
Long term, this is healthy. Probably for X. Probably for crypto too. Reward-for-posting turned timelines into sludge. What surprised me was how little was left once the incentives disappered.
Cleaning that up improves content, sharpens signal, and forces projects to earn attention instead of farming it.
On that note, Stillcore Capital just put out its State of TAO 2026 report, and it’s worth a skim. Bittensor is one of the few corners of crypto I am genuinely excited about right now, mostly because real innnovation is still happening there. I liked how the report frames Bittensor as a live subnet economy and shows how the effects of flow-based emissions play out (more on that with Hippius below).
Which brings us to this week’s spotlights:
The Big Story: Hippius breaks out post-upgrade as emissions jump to ~2.5%.
The Alpha: Zeus beats top weather models by ~40% using real-time specialist routing.
The Weird: Axis tries to teach a robot without telling it what ‘right’ looks like
🕹️Hippius’ Emissions Break Out
Source: taostats
Bittensor just showed its first real post-upgrade winner.
Hippius’ subnet emissions jumped from 2% to ~2.5% this week, putting it among one of only two subnets with a clear upward move (alongside Score, which we talked about a few weeks ago).
It comes after Bittensor’s December shift to stake-flow emissions, where rewards now follow net stake inflows rather than static allocation.
What makes Hippius different is what it actually does. It’s built as a decentralized cloud storage layer inside Bittensor — think persistent, blockchain-metered storage for models, data, checkpoints and outputs instead of relying on AWS or Google. That fills a structural gap in the network’s product stack, bringing storage into the same incentive framework as compute and inference.
It also recently received backing from Stillcore Capital.
This performance, driven by real staking demand, suggests Hippius may move into the protocol’s top 5 subnets if the trend continues.
Hippius’ subnet token is trading at $21.8M market cap and $123M FDV today.
🌩️ Zeus: Weather Forecasting at Market Speed
After watching flights get cancelled this winter because snow calls came hours too late (the single worst thing that can happen to anyone on holiday, trust me), the value of better timing becomes obvious.
That’s the problem Zeus is attacking on Bittensor Subnet 18. Instead of betting on one massive global weather model, Zeus routes each forecast to the best-performing regional models in real time, then merges them into a single, low-latency API.
Source: Zeus
The Wedge
Forecasts are high-resolution and aligned to fast market intervals, making them practical for energy trading and other time-sensitive operations.
Zeus delivers roughly 40% lower error than state-of-the-art baselines, with wind and dewpoint accuracy improving by over 65% and precipitation error dropping by up to 22%.
The Fine Print
A mixture-of-experts system is only as strong as its coverage. If certain regions or variables lack enough specialist depth, Zeus falls back to a “least-bad” consensus rather than a true expert forecast.
Better forecasts won’t stop bad weather. But earlier decisions mean fewer disruptions, even when conditions are ugly.
Zeus’ subnet token is trading at $8M market cap and $42M FDV today
🌹 Teaching a Robot Its First Act of Care
Robots are easy to script and hard to teach. This week, Axis Robotics tried something different. It started with a rose.
Axis just launched The Little Prince’s Rose, a 30-day public experiment to teach a robot how to autonomously water a rose from start to finish. The goal is to train a single robotic control policy directly from human-guided interaction. Axis builds crypto-native infrastructure for robotic intelligence.
The Wedge
Instead of collecting labels, Axis collects human experience. Participants guide the robot through real actions, and physics decides what counts. If the pot tips or the water spills, the trajectory fails. What survives is what’s physically stable.
Simulation makes this scalable, allowing endless repetition, cheap variation, and training a single coherent control policy under real physical constraints.
In the first 24 hours, 10,000+ participants generated ~5,000 usable training trajectories, stress-testing both the learning pipeline and infrastructure.
The Fine Print
This only works if all the human input comes together into one clear behavior. If it doesn’t, the result is noise instead of a usable control policy.
If Axis can make that convergence reliable at scale, it removes one of embodied AI’s hardest bottlenecks: teaching robots without scripting the world for them.
I tried it myself and was genuinely surprised by how interactive it felt and how many ways there were to control it. It also made one thing clear: simulation-based training is the most practical path we have for teaching robots at scale.
💸 Capital Flows
Astrid Intelligence has acquired TaoFi, retaining TAO swap, liquidity, and bridging while winding down lending and cross-chain features.
⚙️ Infra & Protocols
OpenServ launched its AI-native launchpad on Base, letting autonomous startups launch tokens from day one with built-in execution and fee alignment.
Perle Labs launched Season 1, expanding human-verified data work with higher-value tasks.
🤖 Agents & Apps in the Wild.
ChaosChain just shipped an ERC-8004 implementation that puts on-chain Proof of Agency into practice.
Teneo Protocol launched live x402 payments in its Agent Console, enabling AI agents to earn USDC for successful requests.
Fetch.ai released FetchCoder v2, adding spec-driven planning, safety budgets, and test-first workflows
Mind Network launched the x402z testnet, enabling privacy-preserving agent payments where intent remains confidential.
🧠 Bittensor Ecosystem
Hermes (SN82) launched on Bittensor, enabling AI agents to query and act on real time on-chain data via GraphQL.
CrunchDAO just opened Bittensor mining to academic and enterprise ML scientists, bringing in 11K ML engineers and 1.2K PhDs across 100+ countries.
Sportstensor’s Almanac has opened beta access to its incentivized prediction-market intelligence terminal
🦾 Robotics On-Chain
The GPU shortage was never about supply. We have plenty of chips. The real crisis is that we still don’t know how to use most of them together for AI work! This deep dive explains why that’s the real bottleneck.
🔥 Our Weekly Top 5
#1 Render turns a music video into a demo
Render Network’s Octane 2026 update was used almost instantly to render Gaussian splats and volumetric effects in A$AP Rocky’s new Helicopter video.
#2 The next phase of warfare is showing up
After drones dominated recent conflicts, new footage of weapon-mounted robots in use by the Indian Army suggests AI-driven robotics is next
#3 Robots are earlier than they look
General-purpose home robots are still years away, but task-specific robots are already entering factories and improving month by month, not in one big leap
#4 AI Launchpads are splintering fast
The AI launchpad space is fragmenting into distinct lanes: vibe-driven dev onboarding, revenue-backed model tokens and agent-native products like Polymarket ETFs.
#5 KAITO didn’t dump, it broke
Insider sell pressure, stacked unlocks, and X shutting down InfoFi wiped out ~70% of token utility at once, triggering a full structural unwind.
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.
