I was reading Andej’s new blog this week (he’s the founder of Grass), and honestly, it was a breath of fresh air.

It feels like crypto’s entering a new era. This time around, we’ve got real legitimacy. Institutions are adopting stablecoins, the regulatory outlook is improving, and saying you “work in crypto” doesn’t feel radioactive anymore..

What makes this time different from the previous cycles (remember Bored Apes?) is substance. More teams are actually shipping. They’re making money. And in many cases, that value is flowing back to the token. Grass is one of them: they’re reportedly pulling 8-figure annualized revenue while serving top-tier AI research labs.

Value, not just vibes.

One more thought from my chats this week: the people who really believe in decentralization and privacy….like, deep-in-the-bones belief…should probably form a cult and keep the flame alive. That’s what cults are good at.

Because let’s face it: a lot of the original crypto idealism has been replaced. Pragmatism, tribal chain wars, and incentives dominate now. It’s easy o dismiss decentralization and censorship-resistance as empty slogans. Much harder to defend them as core values.

But I still want that world. One where the little guys win. Even if I’m not the one who builds it.

In other news: Elon Musk just leapfrogged the AI race.

xAI released Grok 4 this week, and it’s already making the competition sweat. It crushed Humanity’s Last Exam, a brutal 100+ question gauntlet across math, logic, and chemistry, outscoring GPT-4, Gemini, and o3. One of the lead researchers called the pace of progress “ludicrous.”

But that speed came at a cost: compute. Grok 4 runs on Colossus, xAI’s $-bleeding GPU supercluster.  Their next video model is training on 100,000 Nvidia GB200s, each priced around $60K–$70K. Do the math. (Or let Grok do it.)

The model’s reasoning is next-level. Its filter, though? Still rough. Grok’s X account posted some pretty wild stuff this week. xAI pulled the outputs but didn’t say much about safety. Meanwhile, Grok’s being prepped for robots and autonomous vehicles.

So we’ve got a model that’s brilliant, unfiltered, and running on enough GPUs to power a small country.

What could possibly go wrong? 🥹

🚨 But wait—

  1. Are you in our telegram channel yet? Cyra, our AI researcher, drops daily AI alpha.

  2. Catch our podcast summaries too, complete with detailed show notes.

  3. And if you want to jam with us directly, hop into our Discord

Infinity Ground’s $AIN token has officially launched across various exchanges, including Binance, Bitget, Gate, MEXC, KuCoin, and LBank.

Warden Protocol just pulled off a jaw-dropper: 2 million users in 21 days. That’s a 40x spike, with 95,000 new users showing up every day. Absolute onboarding machine.

Giant Protocol is going live on Binance Alpha starting July 11. If you’re holding Alpha Points, you’ll be able to redeem airdrops once trading kicks off.

Now for something quietly unhinged: humanoid robots are getting really cheap. A $50K robot working 20 hours a day, 330 days a year? That’s just $3.79/hour.

Compare that to factory wages in India (under $2/hour) and then consider this: between 2022 and 2024, unit costs plunged over 40%. Unitree’s H1 now runs ~$16K, which is… minimum wage territory.

Our take? By 2030, we’re probably staring down a million humanoid coworkers, as we wrote in our research piece on robotics.

Let’s just hope the bots don’t form a DAO and demand healthcare. 😱

📈 Rei Network’s Core Sees New Improvements

Rei Network is quietly cooking up some of the most compelling infrastructure for AI agents right now.

Their core system (literally called Core) is built to help agents learn, reason, and work across different models at once. It’s not a chatbot. It’s more like a toolkit for building systems that can think and improve over time.

The main structure is called the Bowtie Architecture. It stores memory in three layers: what something means, the key ideas behind it, and abstract patterns. That setup helps the agent make both accurate connections and creative leaps.

Source: 0xreisearch.gitbook.com

Next is the Reasoning Cluster. This handles planning. It looks at what you’re asking, picks the right tools to answer it, and tracks how different ideas connect. Kind of like a living, breathing mindmap that updates as the agent grows smarter.

Then there’s Model Orchestration. That’s the part that actually gets things done. Big tasks get broken into smaller chunks, each routed to the best model for the job. Fast, efficient, and surprisingly elegant.

And yes, it learns. The more you use Core, the better it gets. It remembers past interactions, notices recurring patterns, and adapts its responses over time. It’s building intuition.

The new 0.3 update added some spicy new features:

  • Metacognition: agents can now detect uncertainty in their own answers

  • Chart-making: built-in visual reasoning

  • Structured memory: better concept storage, faster recall

Also, it’s much snappier, with some tasks now running over 80% faster.

Is it perfect? Nah. There are bugs and it’s still early. But if you’re even mildly curious about what long-term agent intelligence might look like, Core’s definitely worth a peek.

💵 USD AI Crosses $31M in TVL on Arbitrum

USD AI has reached $31M in TVL on Arbitrum. USD AI is a synthetic dollar system designed to finance real-world infrastructure: GPUs, racks, and the heavy metal powering AI and DePIN networks.

There are two tokens:

  • USDai: fully-backed, stable, always redeemable

  • sUSDai: yield-bearing, backed by income-generating hardware

sUSDai pays 15–25% APR by lending to smaller infra operators, the folks who can’t just stroll into JPMorgan for a loan. These loans are collateralized by the hardware itself using something called CALIBER, which turns servers and GPUs into onchain assets. No credit scores. Just machines with value.

Here’s who makes the system tick:

  • Depositors mint USDai and stake it into sUSDai.

  • Borrowers get the gear they need to run workloads.

  • Curators take the first-loss risk and earn yield premiums for doing it.

Capital sits in T-bills while the hardware gets sourced, with something called “Cores” smoothing out the lag between commitment and actual deployment.

Big picture? USD AI is part of a growing trend: using crypto rails to fund the physical infrastructure behind AI.

Fundraises

  • Zypher Network just raised $7M to scale its zero-knowledge trust stack for autonomous AI agents. The funds will speed up the development of its zk protocol suite and AI-focused rollup, Zytron.

  • (Web2) LangChain, an AI agent framework, is raising a new round of funding at $1B valuation. Its open-source framework has 111K GitHub stars. What does this imply for Web3 agent frameworks?

  • (Web2) ZeroEntropy raised a $4.2M seed round led by Initialized Capital and launched zerank-1, a reranker that outperforms larger models with lower costs and latency.

Startup Updates

It’s been a slower summer, but not all quiet.

  • Prime Intellect dropped SYNTHETIC-2, a massive dataset of 4 million verified reasoning traces for complex reinforcement learning tasks. It was generated via decentralized inference across 1,250+ GPUs and is now live on Hugging Face.

  • Templar (Subnet 3) releases CCLoco, an optimizer that helps train AI models faster and more efficiently in a distributed fashion, reducing communication by up to 50x.

  • Auki Network’s $AUKI is now bridged to peaq and tradable on MachineX, the first DEX built for the Machine Economy. Auki is a spatial computing DePIN building a collaborative sense of space for robotics and smart cities.

  • peaq also rolled out a Python SDK for the Machine Economy. Devs can now build AI agents that interact with machine IDs, store data, and move tokens.

  • Masa just launched the first open TikTok scraper on Bittensor Subnet 42. Anyone can now scrape and transcribe TikTok videos into clean, AI-ready data using secure TEE infrastructure. Free and open.

  • qBitTensor Labs kicked off Subnet 63 on Bittensor, a.k.a. the Quantum Subnet, with a mission to bring quantum computing to decentralized networks. The first experiment, Peaked Circuits, is already running.

  • NetMind AI launched a new Stake-to-Participate model: $NMT rewards are now tied to compute and staking. Unstaked rewards get burned.

  • MakeInfinite Labs dropped the Infinite App Printer (alpha) on Space and Time DB. It’s a full no-code stack for building, launching, and tokenizing on-chain apps. Beta lands in August.

  • Autonomys, an AI-focused L1, will be enabling native token transfers on July 16 and unveiling its grants program and ecosystem growth strategy.

AI Agents

  • Project Zero launched ZERO Chat, a Web3-native assistant that lets you query wallets, DeFi positions, NFTs, and on-chain activity.

  • DynaVest is live on Virtuals Protocol, bringing a new DeFAI agent that handles yield optimization and portfolio insights with automated strategies for your DeFi bags.

  • Proximity Labs just dropped the Shade Agent Sandbox, giving devs an easy way to build Shade Agents in JavaScript, no NEAR contract management or infra needed. Local testing and TEE deployment supported.

  • Hyperfly unveiled its Automation Marketplace, where traders can publish, explore, and copy trading automations. Think: crowdsourced algo trading, on-chain style.

  • Alchemist AI rolled out Rapid Generation Mode v3.5, a faster path to prototyping that skips full agent orchestration. If you want quick results without the overhead, this is your mode.

Web2 AI

  • Mistral AI nd All Hands AI just launched Devstral Me and the upgraded Devstral Small 1.1, pushing agentic coding forward with high performance at low cost. Small is fully open-source; Medium is available via API.

  • Liquid AI released LFM2, a blazing-fast, memory-efficient foundation model built for on-device AI. It runs 2x faster and brings real-time, private GenAI to everyday hardware

  • Perplexity launched Comet, an AI-powered browser with a built-in assistant that helps you search, summarize, and stay organized. It’s now live via a $200/month Max plan.

  • Hugging Face dropped SmolLM3, a fully open 3B multilingual model with 128k context and dual-mode reasoning. It beats other 3B models and even rivals some bigger ones.

  • Cerebras launched Supernova, giving developers 1M free inference tokens daily, plus up to 245M tokens/day and boosted limits for startups. Yes, free.

🎙️ New episode of our Escape Velocity podcast just dropped, and the AI wars are heating up. Me + CK go deep on:

  • Meta's ruthless AI talent grabs

  • Elon Musk's xAI's growing pains

  • Is Apple… cooked?

  • AI x robotics = the real endgame

And yeah, we drop some spicy alpha at the end (no one’s talking about this yet)

🔥 Our Weekly Top 5

Lots of great insights on Crypto VC. It’s tough out there

Vitalik’s response to AI 2027

NEAR seems to be gaining traction…

Hugging Face launches a $299 open source robot

Virtuals ACP releases a builder’s guide

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