GM!

This week it was Manus, Manus, Manus.

The general-purpose AI agent from China took the space by storm, so exclusive that invite codes were flipping for $10K in the first few days.

We gave it a spin, and there’s something almost intangible about its magic. Maybe it’s the way you can ask Manus to do anything without overthinking the how. Under the hood, it’s running Claude 3.5 Sonnet ($6 per million tokens) alongside fine-tuned Qwen models (likely $0.80–$2.80 per million tokens).

While there are still lots of kinks—the AI gets stuck, crashes your browser and overloaded servers— it’s an early glimpse of where AI is heading.

With funding tightening and expectations rising, AI agents need more than novelty—they need staying power. OpenAI’s Agent SDK and Anthropic’s Model Context Protocol (MCP) point to a trend: aggregation.

Instead of juggling multiple integrations, these frameworks simplify access to data and automation, making AI agents more embedded than standalone.

This marks a broader shift: AI agents are evolving into infrastructure. Some will stay as flashy, personality-driven frontends, but most will fade into the background—silently optimizing workflows and making decisions.

In this edition of our weekly AI Edge, we cover:

  • Bankr Bot and Grok Team Up

  • A look into Bittensor Subnet 3 (Templar)

  • Nous Research releases it’s Inference API (along with two new models)

Hold up—are you even in our Telegram yet? If not, you’re missing out. Cyra, our AI researcher, is basically an alpha vending machine, dropping daily AI insights.

Also, we’ve got podcast summaries. Saves you hours of time each week.

If you’re vibing with this, spread the word! We keep writing because our readers support our mission of growing the pie for Crypto AI. Share or forward this newsletter with someone who’ll love it. 

We won’t harp too much on the markets this week—lots of red. ENQAI, ARC, and NOVA got wrecked, each down nearly 40%. AI projects have gone full builder mode, while the broader market hit snooze.

But there’s a pulse. AI mindshare is creeping back up, ticking to 34% from just under 30%.

Source: kaito.ai

We’ll leave you with some optimism for the rest of 2025:

Bankr Bot and Grok Team Up

Source: dexscreener.com

So, Grok and BankrBot unintentionally teamed up to launch a token—and the results have been wild. It all started when a human prompted @bankrbot on X to launch DRB (DebtReliefBot) based on a suggestion from Grok.

The results? People are increasingly using Grok and BankrBot to launch tokens, turning AI into an active participant in crypto speculation. The numbers speak for themselves: Grok’s wallet is now collecting over $200K in fees, raking in about $5K per hour.

This is because BankrBot runs on Clanker, which collects fees on every trade. Meanwhile, $DRB has already seen $70M+ in trading volume and ranks in the top 10 tokens in terms of mindshare, according to Kaito. Even BankrBot's own token, $BNKR, has surged 43% in a week.

So, what is BankrBot exactly? It’s an AI-powered crypto trading and management tool built by Bankr, leveraging Clanker technology. Its core value is simplifying DeFi interactions through X (formerly Twitter), Farcaster, or a private terminal.

Users can execute trades (BTC, ETH, and other tokens) via social media commands, check balances, send crypto, and authenticate wallets.

Source: bankr.bot

Since its public alpha launch on Feb 11, 2025, any X user can access it. Bankr Club stated that it will soon roll out a premium subscription with lower fees.

However, this whole situation raises serious questions. Are we witnessing the dawn of AI-driven market manipulation? If AI can suggest, launch, and profit from tokens, who’s really in control? And where does this leave the concept of fair market participation?

Either way, AI is no longer just a tool in the market—it’s becoming a player.

🌴 Bittensor Subnet 3: Templar

Templar (Bittensor Subnet 3) is shaking up AI training by letting people worldwide collaborate using internet data. Instead of one company hoarding the process, individuals contribute their computing power to improve models. Contributors earn TAO tokens for making the AI model better.

It’s like Bitcoin mining but for AI, rewarding effort with real incentives.

This is a different approach from teams like Prime Intellect and Nous Research, which zero in on low-bandwidth interconnect optimization. Instead, Templar focuses heavily on the incentive layer.

There are two key roles in Templar.

  • Miners train parts of AI models and submit improvements.

  • Validators check if those improvements actually help.

This happens across a distributed network, where each miner works on different data, but collectively, they refine the same global AI model.

Templar measures improvement simply: compare model performance before and after a contribution. Bigger gains? More rewards. Data is compressed before sharing to minimize the risk of leaking sensitive information. And anyone can participate.

The system is designed to prevent abuse. Rewards depend on actual improvement, not just time spent. The system assigns data randomly to prevent miners from cherry-picking easy tasks. And sybil attacks—creating fake accounts to game rewards—are useless since only real contributions count.

Right now, Templar has 256 active miners, maxing out its available slots. The network receive 2.36% of total TAO emissions, or 170 TAO ($42K) per day. Earnings are split: ~30 TAO daily for subnet owners, ~70 TAO for miners, and ~70 TAO for validators.

What’s particularly interesting is that there has been no way for retail participants to get exposure to decentralised training startups thus far. Most have been VC backed companies.

But as a Bittensor subnet, Templar comes with its own subnet token, currently at a $161M fully diluted valuation.

Project Updates

  • Nous Research just launched its Inference API, making Hermes 3 Llama 70B & DeepHermes 3 8B Preview accessible to developers. OpenAI-compatible, pay-as-you-go, with $5 free credits.

  • Nous Research also annonuced their DeepHermes 24B & 3B modes. These hybrid reasoners let you toggle between fast, intuitive answers and detailed, high-accuracy reasoning. They are now live on the Inference API and Hugging Face.

  • LAGRANGE introduces DeepProve, a zkML library that can verify AI decisions and generate zero-knowledge proofs up to 158x faster than before.

  • Kaito just dropped a major upgrade—now with a public API for real-time Yaps scores, alongside the monthly on-chain updates. Builders can integrate Yaps faster, pushing "attention as a currency" to the next level.

  • Giza introduces Giza Protocol—trustless, context-aware, and permissionless infrastructure for DeFi agents. It enables secure, decentralized execution, smarter on-chain decision-making, and non-custodial agent control. Built on EigenLayer, it bridges AI agents with Web3.

  • GAIB & Peri Labs releases a mega 169-page AiFi Thesis, outlining the current state of the market and the big ideas.

  • Nillion’s Airdrop Allocation Checker is Live. Early supporters can now claim their rewards and showcase their impact with custom social cards.

  • Yotta Labs introduced BloomBee a decentralized LLM service. Built with the research team at UC Merced, it enables cost-effective inference & fine-tuning by leveraging geo-distributed consumer GPUs.

  • Atoma & Walrus have fully decentralized DeepSeek R1! Walrus powers storage, while Atoma distributes inference across a decentralized compute network—marking the first on-chain deployment of a model this size.

  • Akash Mainnet 13 is live. This upgrade boosts on-chain query performance and completes AEP-61—another major step forward on the Akash roadmap.

  • Chutes (Bittensor) is now processing 12B+ tokens daily—and they’ve added a ton of new models to the site and OpenRouter such as Google’s Gemma 3.

AI Agents

  • Virtuals just launched its first ACP Hackathon—$100K in prizes for the best autonomous agent businesses, cohosted with Base and Stanford. Winners get cash, funding fast-tracks, and exclusive Network School access (oh, and we’re judging!).

  • OpenAI launched the Agents SDK—a powerful toolkit for building real-world AI agents. It simplifies agent creation, multi-agent collaboration, tool usage, and guardrails while supporting real-time streaming and OpenAI’s evaluation tools.

  • ARC just announced they’re launching an "AppStore for Agents," letting AI agents access crypto and Web2 services via one integration. Payments in USD, USDC, SOL, with 10% of revenue buying back $ARC—we like to see tokens with value accrual.

  • Coinbase AgentKit now supports OpenAI’s Agents SDK. Build AI agents with web search, file search, and performance tracking—now with integrated secure crypto wallets for instant, low-fee global transactions.

  • Aiccelerate DAO’s releases its 2025 roadmap and also announced AlphaPulseAI—a platform aggregating insights from Crypto Twitter, GitHub, and a crypto-optimized database.

  • Chaos AI is here—an AI-powered crypto researcher. Built on Chaos Labs’ proprietary data, it delivers hedge fund-grade analysis, deep research, verified insights, and yield discovery in seconds. It’s now in closed beta.

Web2 AI

  • Google DeepMind just dropped Gemma 3—its most advanced open model optimized for single GPUs and TPUs. With sizes from 1B to 27B, it supports 140+ languages, 128K context, vision-language tasks, and real-time function calling.

  • Google Deepmind also released native image generation with Gemini 2.0 Flash. Allowing users to create and edit images seamlessly—now available in Google AI Studio.

  • Converegence introduces DeepWork—their most advanced AI agent yet. It autonomously coordinates multiple agents for complex, multi-step workflows, pushing the vision of AI Organizations forward. It is now live for pro users.

  • Microsoft launches Dragon Copilot—the first unified voice AI for healthcare. Combining Dragon Medical One and DAX, it streamlines documentation, automates tasks, and enhances patient care. Rolling out in May across the U.S., Canada, and Europe.

This week we explored Masa, a decentralized data protocol running two Bittensor subnets: Subnet 42, a real-time data refinery, and Subnet 59 (Agent Arena). Dive in here 👇

Masa: AI's Data Tap

Fueling AI with real-time data, Masa powers an open pipeline to keep intelligence flowing

www.chainofthought.xyz/p/masa-ai-data-tap

We also shared our thoughts in Macrocosmos’ latest article on the dTAO launch on Bittensor.

🔥  Our Weekly Top 5

Karpathy gives us plenty to chew on this weekend

Markets are down so bad even aixbt knows to FUD itself

CryptoEternalAI allows you to generate AI videos on X with a single prompt

Luna is launching her own autonomous business

AI agents are blurring the line between apps and the operating system, raising concerns about privacy

That’s it for this week! If you have specific feedback or anything interesting you’d like to share, please just reply to this email. We read everything.

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