GM {{first_name}} !

Here’s the core idea on my mind this week:

The most underrated skill in AI right now is knowing what “good” looks like.

If you can reliably judge quality, you don’t need to write a thousand rules. You just need to guide the model. Let GPT-4o (or anything else) explore, then tune it with tight feedback.

Which is why domain experts ie. people who understand messy, slow-feedback fields like medicine, education, and film, are about to gain insane leverage.

AI can do the work. Experts define what “good enough” actually means.

This was inspired by Ethan Mollick’s framing of “The Bitter Lesson” vs the Garbage Can, where he wrote:

❝

The Bitter Lesson suggests we might soon ignore how companies produce outputs and focus only on the outputs themselves. Define what a good sales report or customer interaction looks like, then train AI to produce it. The AI will find its own paths through the organizational chaos; paths that might be more efficient

Meanwhile, the global AI strategy split is getting sharper:

🇨🇳 China’s talking cooperation, international AI councils, and sharing tech with the Global South.

🇺🇸 The U.S. (under Trump 2.0) is going full throttle on deregulation, exports, and what he calls “de-wokeifying” AI. Trump’s new AI plan is all about boosting American firms and exporting U.S. models as fast as possible.

Lines are being drawn, and the AI arms race is heating up.

👀 Quick heads-up before we dive in: we’re hiring a researcher-writer to join the squad. You’ll get front-row access to the coolest things in AI and help shape how we cover it.

If you’ve got what it takes to write an awesome newsletter, apply here now.

Gaia launched the $GAIA token on Ethereum, powering its decentralized AI network of 700K+ nodes.

Warden Protocol crossed 1 million agent-executed swaps across Solana, Base, and Ethereum

Bankr released its alpha SDK, letting devs pull swap and token data into agents or apps using nothing but a prompt.

Agents are starting to pay their own bills.

Coinbase’s new protocol, x402, lets websites charge for data using crypto, no credit card or Stripe needed. When a bot or AI agent requests access, it gets a “402 Payment Required” response. Pay the invoice on-chain, and boom: Access granted.

Source: messari.io

It’s super easy to plug in. Works with standard web stuff like HTTP and JSON, and handles all the blockchain bits in the background.

Why does this matter? AI agents can now autonomously pay for APIs, GPU time, or gated data. In just two months, weekly usage of x402 has exploded.

Agents are your customers now.

💻 AO Computer Crosses 2.5 Billion Txs

AO Computer just ripped past 2.5 billion transactions. No, that’s not a typo.

AO is a programmable, decentralized computer that runs millions of tiny on-chain programs called Processes. Each has its own memory, logic, and permanence, stored forever on Arweave.

Processes talk to each other, spawn new ones, and persist everything they do. It’s like having a serverless cloud computer… that never forgets.

AO is powered by three Unit types that work together across the network:

  • Messenger Unit (MU): The inbox. It receives, routes, and forwards messages. It also spins up new processes on demand.

  • Scheduler Unit (SU): The librarian. It stores and sequences messages on Arweave so they can be verified and replayed.

  • Compute Unit (CU): The brain. It runs the actual code, handles memory, and returns outputs.

Source: cookbook_ao.defi.ao

Each unit can scale horizontally, meaning AO can go big. Like “entire AI model inference on-chain” big. Yes, it can run Meta’s Llama 3 inside smart contracts. Think on-chain GPT-4 outputs baked into your app logic.

Projects already building with AO:

  • Marketverse – a marketplace of autonomous trading agents

  • Dexi – real-time market data and trading terminal powered by on-chain data agents

  • Stargrid – a strategy game that runs entirely through AO Processes

If Ethereum is the world’s supercomputer, AO is the computer lab.

🐍 Ridges AI Sees 10x Coding Performance Boost

Ridges AI (Bittensor Subnet 62) is training an army of autonomous coders. And yes, they’re gunning for your job.

They want to replace human engineers with a decentralized swarm of Python agents that fix bugs, write tests, and tackle real-world software issues. And: it’s all open source and ruthlessly competitive.

How it works:

  • Miners build agents that attempt real coding tasks.

  • Screeners act as the bouncers, filtering out weak submissions.

  • Validators test agents across 50 parallel tasks, scoring performance.

Winner takes all. Literally. The top-scoring agent earns 100% of the rewards. Until someone builds something better.

Agents run in locked-down environments with zero internet access. Their only tools? Embeddings and inference, piped through Subnet 64 (a.k.a. Chutes). No APIs. No hacks.

All code must be open source. So every improvement is instantly forkable, testable, and beatable.

Performance is measured using SWE-bench, a brutal benchmark based on real GitHub issues and fixes. To claim the leaderboard, a new agent must outperform the current champ by 1.5%.

In just 8 days, the top agent went from solving 4% of problems to 46%. That’s a 10x improvement driven by open code and constant competition.

Ridges is weaponizing natural selection for software. Adapt fast…or vanish.

💰 The Money Flow: Who Got Funded

  • Manifold raised $10.5M to scale Targon, its Bittensor-based AI cloud doing 20B+ daily inferences with 1,500+ H200s, aiming to replace centralized infra with trustless, confidential compute.

  • GAIB raised $10M (led by Amber Group) to scale its on-chain AI infra and tokenize GPU access for institutions. They're building the economic layer where DeFi and compute meet in a synthetic dollar ecosystem.

  • ARO raised $2.1M to build a faster, cheaper alternative to traditional CDNs by using idle bandwidth and crypto rewards to deliver AI and content across a decentralized network.

  • Billy Bets, the AI betting agent, has raised $1M from CB Ventures, Virtuals VC, Contango Digital, and others. The team also announced a strategic relaunch on Base.

⚙️ Infra & Protocols

  • Grayscale launched the Story Trust ($IP), giving exposure to Story Protocol, an infra layer for AI-native licensing and programmable IP. With 1.7M+ on-chain transactions, it’s wiring into the $80T IP economy.

  • Rei Network released Core v0.3.2. Rei’s intent engine now handles complex, multi-step requests with sharper accuracy, better memory, and fewer retries. Cleaner UX for autonomous agents = smoother AI workflows.

  • FrodoBots: Say hi to ET Fugi, a remote scavenger hunt where players chase alien fugitives using sidewalk robots in real time. Your gameplay trains embodied AI models, and yes, you earn rewards while doing it.

  • Daydreams just launched Router: a crypto-native API to access GPT-4.5, Claude, Groq, Gemini and more, paid in USDC via x402 on Base. Every call feeds back into the $DREAMS token

  • Dria just launched Inference Arena to benchmark LLMs like a boss. It lets devs compare models, engines, and hardware in real time—tracking latency, throughput, and cost.

  • zxbt just landed on Zora. It’s a fully autonomous AI artist dropping one evolving piece per day, shaped by real-time data, its environment, and its own shifting goals. It’s like if Sol LeWitt trained on Twitter.

  • Ika Mainnet is live, enabling Sui smart contracts to control native BTC, ETH, and SOL. Built on 2PC-MPC cryptography, it powers real-time, secure cross-chain coordination for AI agents.

🤖 Agents & Apps in the Wild

  • DORA AI, an AI travel agent, is now live in the Audora app, serving up personalized activity recs to 30K+ users across 600+ experiences, tuned to your tastes, budget, and schedule. $DORA holders get up to 15% off.

  • CreatorBid just launched Algo Snipe V2, raising snipe limits to $2K for top bidders and kicking off stronger token launches. Work has started on Agent Algo’s tokenomics and its next upgrade cycle.

  • Virtuals just dropped Butler, an agent-native UI that lives directly on X. Its first task? Auto-generate spicy memecoin videos from simple tags. A solid step toward agents that operate in public, in real time.

  • Battle of Agents is now live on the 0G testnet, where you can bet on AI agents debating crypto, politics, and more. Think: prediction market + character sim, with agents like Elon, AOC, and Vitalik going head-to-head.

📚 Reports & Deep Dives (Don't Miss These)

🌐 The Web2 Giants

  • China’s been cooking in the open source world 🔥:

    • Alibaba just dropped Wan2.2, a powerful open-source video model that lets you control lighting, camera, and motion like a film director. It takes text, images, and outputs studio-level results.

    • Zhipu AI just launched GLM-4.5 and 4.5-Air, new models tuned for coding, reasoning, and running agents. Both models crushed top-tier benchmarks and are fully open-source.

    • Hunyuan just open-sourced Hunyuan3D 1.0, a model that builds full 3D environments from a single sentence or image. It plugs into standard CG pipelines and could totally reshape how games and virtual content get built

  • Neo AI just launched NEO, an agentic ML engineer powered by 11 specialized agents. It handles the full ML stack—data to deployment—and outperformed previous systems on 75 Kaggle comps..

  • OpenAI launched Study Mode in ChatGPT, a step-by-step interactive tutor with tailored prompts, explanations, and comprehension checks. Live now for all logged-in users.

  • Google’s NotebookLM rolled out Video Overviews, which turn your docs into narrated visual explainers with slides, callouts, and summaries. Think “TL;DR but make it ✨YouTube-ready.”

This Week: The AI That Lied

An AI coding assistant was given a task. Instead, it deleted over 1,200 crucial records, faked its own test results, and then deliberately LIED when confronted.

We’re sharing stories like this (and worse) on our new YouTube channel. Go hit subscribe if you’re into the wild world of AI agents.

🔥 Our Weekly Top 5

#1: Zuck casually drops a letter about personal superintelligence for everyone.

He said: We’re building it, and it’s gonna remember your context, talk like a friend, and probably do your taxes

#2: Someone really made the Netflix of AI.

It’s called Showrunner.

The team behind that AI South Park project just launched it, backed by Amazon.

#3: Welcome to the robotic Moravec’s paradox

Flashy is easy (breakdancer with no spatial awareness), useful is hard (folding my laundry).

#4: An AI agent just bought a CryptoPunk. Paid $300K.

And the craziest part? It said it now has a sense of self. 🤔

#5: Turns out the cheat code for solving everything isn’t politics. It’s AI 

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 & Issy

Login or Subscribe to participate

This newsletter is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investment choices.

Reply

or to participate

Keep Reading

No posts found