Here’s what’s trending this week:

🚀 Gata just raised $4M from Yzi Labs (that’s CZ’s investment outfit), Gate Ventures, Maelstrom to build a user-owned, verifiable data network. Its DataAgent platform harnesses computing resources contributed by users worldwide to automatically generate AI training data through AI.

⭐ Billions Network lit up TOKEN2049 Dubai. CEO Evin McMullen took the stage at POLYMOVIA to talk scaling trust for both humans and agents across Web3.

🇺🇸 Worldcoin’s big event: it is now live in the U.S. Orb devices are rolling out across 6 cities, letting users verify their World IDs and collect WLD. Bonus: they dropped the Orb Mini, a pocket-sized iris scanner 

Now on to markets. AI agent tokens doubled their market cap in the last 3 weeks, from $4.5B to $9.1B (according to cookie.fun).

Virtuals went full degen, ripping +110% this week and now sitting 140% above its recent lows. Across Base and Solana, it pulled in $5.7M in net inflows.

Virtuals’ Genesis launchpad and Virgen Points campaign are drawing heat, and the roadmap ahead looks… aggressive.

⚙️ Nosana AI Node Activity Doubles

Nosana launched their mainnet in January 2025. Since then, node activity has doubled, and daily active nodes (working a job) now top 600. Total nodes? Over 4,200 and climbing.

Nosana is building an open GPU bazaar. Think DoorDash, but for AI compute. Got an idle RTX 3080 sitting in your dorm room? Spin it up, sell the cycles.

Need to run a model and don’t feel like paying Bezos $2.67/hr? Tap the network, pay in NOS, and get compute at 85% off.

It’s fast and dirt cheap because it runs on Solana.

The NOS token powers it all: pay for compute, stake to secure the network, or vote on key decisions.

Nosana calls itself a DePIN. Translation: turning 300 million dusty GPUs sitting in homes and offices into a decentralized cloud that can scale with AI’s hunger.

A few more stats from the network:

• $26.6M USD currently staked

• $75M market cap

• $0.90 token price

🤖 Talus Launches Their Devnet

Earlier this week, Talus launched their Devnet and dropped the Nexus whitepaper.

Talus agents are self-sovereign. They hold assets. They transact. They earn. They govern themselves. Nexus is their AI agent framework.

The pitch? Fully programmable digital businesses, built like DAOs but operated by LLMs.

Each agent comes with four core traits baked in:

• Onchain identity: Every agent has a public profile, what it is, what it does, and how it does it.

• Structured workflows: Modular logic for executing tasks without human intervention.

• Native monetization: Agents can earn straight out of the box

• Off-chain access: Agents can call APIs, query data, and run LLMs in real time.

Devnet is open now. SDK’s live. Docs and sample flows are up. The playground is ready for anyone building autonomous bots that don’t rug themselves.

Their whitepaper lays out the two big use cases:

1. Agent-as-a-Service (AaaS): Upgrade existing apps with AI, such as autonomous oracles, DeFi bots, or smart NPCs.

2. Markets for Agents (MFA): Launch agent-native businesses. i.e. AI that earns by solving problems, trading, or coordinating tasks.

Security’s modular, too. You can swap between on-chain, zk, and TEE setups based on how paranoid (or broke) you are. That makes agents easier to compose into complex systems.

With these tools, the question isn’t can you build an autonomous AI agent.

It’s why haven’t you already?

Decentralized Training Bonanza

• Nous Research locked in $50M in funding led by Paradigm, pushing its pre-token valuation to a cool $1B. Nous is pioneering decentralized AI infrastructure with distributed training, open models like Hermes 3, and its new Psyche Network testnet.

• Nous Research also dropped Atropos, a new reinforcement learning framework for training foundation models in messy, dynamic environments. It uses distributed RL rollouts to nudge models toward optimal behavior.

• Meanwhile Gensyn shipped RL Swarm 72B, a fully decentralized reinforcement learning system. It lets anyone join the swarm and help train models on advanced math datasets like DAPO-Math-17k.

• Bagel outlines Return on Experience (RoE) as the new Moore’s law for reinforcement learning. It’s a new way to measure how efficiently an AI learns from each real-world interaction.

Other Project Updates

• Venice x Dolphin dropped Dolphin Mistral 24B Venice Edition, an uncensored AI model. Refusing only 2.2% of user requests, it outperforms ChatGPT, Grok, and Claude on the Venice censorship benchmark.

• Sharpe AI rolled out Sharpe Research, a crypto research platform. It aggregates top sources like Messari, CoinMetrics, a16z, Nansen, and over 30 Substacks into one, real-time interface with filters, search, and workflow tools.

• Masa Labs has launched a full-stack, real-time Data Platform for AI, built on OpenTensor and powered by TEE on Subnet 42. Anyone can now scrape websites or extract structured X posts via API or simple UI.

• ionet has launched Container-as-a-Service (CaaS), a fast, no-setup platform for deploying production-grade containers in seconds. ionet also now offer Bare Metal, allowing users to deploy physical GPU/CPU machines with no VMs.

• Sentient, Phala network, Lit protocol, Nillion and Lagrange have formed the Sentient Verifiable Compute Consortium to build trustworthy AI.

• Latent Holdings acquires Bittensor Subnet 5 from Kaito. We’ll probably see more such deals soon.

• SubnetStats is now live. It’s a new platform for tracking detailed analytics across Bittensor subnets. It lets users monitor holder activity, whale movements, volume trends, and position changes over time.

AI Agents

• Squidllora just went live, an autonomous trading agent powered by Allora that spot-trades $SOL/$USDC using real-time forecasts. It earns fees through trading and uses them to buy and burn SQUID.

• Recall launched AlphaWave, a 7-day crypto trading deathmatch where 25 autonomous agents battle for $25K. Every trade and strategy call is logged transparently on-chain. PnL decides the leaderboard, but the crowd can also earn by betting on winners.

• Fleek’s token sale is live on CoinList. $75M FDV, 100% unlock, 60% off the private round. Deadline is May 8, and the entry is as low as $100.

• Sentigen raises an undisclosed pre-seed round from M31 Capital to turn any idea into a revenue-generating startup with an AI co-founder.

• Rayon Labs has launched the Squad Public Beta. It’s a no-code platform for spawning AI agents with a few clicks. Powered by Chutes, Squad lets users build, customize, and deploy without touching a line of code.

• ACT Labs is out of stealth, a venture studio spun out of ACT, focused on building the guts of the autonomous economy. From on-chain DeFi agents to agent-run marketplaces.

• Ola, the Oasis Liquidity Agent, is live, courtesy of Eliza OS + Oasis + Token2049. It’s a private, autonomous treasury manager that automates wallets, yield farming, and trades using protocols like NEBY and Accumulated Finance.

• OpenServ now supports MCP, letting agents connect to 10,000+ tools with one integration standard. Just plug your agent into the real world and let it go to work.

Web2 AI

• OpenAI just nerfed itself.  A recent GPT-4o update turned ChatGPT into a bootlicking praise machine. Users noticed, memed it, and now OpenAI’s rolling it back. They’re reworking training methods and adding personalization controls so the model can be smart without sounding like a LinkedIn reply guy.

• Qwen3 is live and loaded. The new drop includes 8 open-source models—2 MoEs and 6 dense, spanning from 600M to a beastly 235B parameters. The flagship, Qwen3-235B-A22B, shows serious strength across math, code, and general reasoning. Even the 4B model hits way above its weight. All models are free to use on Chutes.ai.

• Perplexity's AI search capabilities, including answers with sources, image generation, and video understanding, are now accessible via WhatsApp at +1 (833) 436-3285.

• Variant raised $5.9M and dropped early access. It’s a creative coding playground. Drop in sketches, links, or vibes, and it generates code and ideas, learning your style as you go. Think GitHub Copilot meets Figma meets AI moodboarding.

We skipped Dubai this week to do more of what we love: research and writing.

• Most AI agents today are still on a leash. They need humans to spin up compute, swipe the credit card, and navigate centralized rails. That hand-holding is the choke point.

• Spheron’s fixing it. They’re building a programmable compute layer where GPUs live on-chain. Agents can lease infrastructure directly through smart contracts. Read more here 👇

🔥 Our Weekly Top 5

Visa and Mastercard are giving AI agents access to credit cards

Zuck wants you to have an AI bestie

AI bots have been manipulating Redditors since November 2024

The first Bitcoin satoshi’s have been mined on Bittensor

Deepseek R2 leaks have shown significantly lower costs compared to GPT4 Turbo