Happy friday {{first_name}} 👋

Welcome to #65 of the weekly AI edge, where we bring you AI x Crypto that matters, before it matters.

One thing I really appreciate, sitting at our vantage point as a research firm, is that I can really feel the pulse of the sector. Early this year was pure mania with AI agents. But by mid-summer, things had slowed to a crawl. Now I’m starting to see the tempo pick up again: teams are flooding us with products, whitepapers and tokenomics to review.

Q4 looks set for a bonanza of token launches. The question is whether the spark turns into another full-blown mania as we close out the year. My hunch is it depends on whether the broader crypto market decides to throw a party.

Anyway…

The open-source AI debate is flaring up again. Critics say it's stalled, pointing to OpenAI’s pivot, Meta’s slowdown, and Mistral’s modest reach.

But Matthew, a researcher at Prime Intellect disagrees. He argues that reinforcement learning environments are infrastructure, and infrastructure trends open. PyTorch, Docker, Linux…systems like these win because they demand constant tweaking and thrive when more people can patch, extend, and improve them.

Closed systems will exist, but open ones will iterate faster. More developers means more fixes, more features, and faster compounding.

As fine-tuning gets easier and cheaper, open RL infra will quietly become the default.

The COT Meme of the Week

Experts say AI x Crypto still lacks product-market fit. Meanwhile, some tokens keep climbing (yes I’m looking at you, NEAR). Who's right?

0G has opened registration for its first airdrop, marking the beginning of its broader community journey. Eligible participants must complete KYC registration to claim.

Maiga’s $MAIGA token just dropped on Binance Alpha, KuCoin, MEXC, Gate, and PancakeSwap. Contract details and docs are up on their official GitBook.

DeAgentAI’s $AIA Genesis Airdrop portal is live. Airdrop and pre-staking are open, and this pre-TGE window is the shot at locking in the max reward multiplier.

🛠️ Pluralis Research Launches Node-0-7.5B

Pluralis Research has launched Node-0-7.5B: a global pretraining run that anyone with a 16GB+ GPU can join.

It’s currently training at ~12.7k tokens per second. At this pace, the run could take 1–2 years. GPT-4 trained much faster with 100x more params and 10x more training tokens.

Source: dashboard.pluralis.ai

The Wedge: Pluralis’ Node-0-7.5B

  • Pluralis calls this “Protocol Learning.” There’s no central server. Each participant runs a shard of the model. No one has the entire model. And together the swarm trains something much larger than any single machine could handle.

  • A peer-to-peer network keeps everything connected. New GPUs boost total training power automatically. A few backbone nodes keep it running smoothly, but anyone, from a gaming PC to a cloud server, can plug in.

  • The current run is 7.5B parameters, trained on the FineWeb-Edu dataset with a 4K context length.

The Fine Print

  • Reliability: This is the first real-world stress test. Stability with many outside nodes joining and leaving is still unproven.

  • Proof, not product: Node-0 is NOT aiming to compete with cutting-edge models. It demonstrates that decentralized training at scale is feasible.

Node-0 is an experiment, but an important one. If it holds up, it could provide communities with a way to train serious models together, without waiting for the next big release from the major AI labs.

📊 Google and Coinbase Release AP2 (+x402)

Google just rolled out payments for AI agents. With its new Agentic Payments Protocol (AP2) and Coinbase’s x402, agents can now send each other stablecoins like USDC.

That means actions like checking the weather or finding SKU prices might soon happen seamlessly, with agents making microtransactions without needing credit card numbers.

The Wedge: AP2 (+x402)

  • A2A lets agents share information; AP2 + x402 lets them handle value.

  • Stablecoins enable instant, tiny “pay-per-use” transactions

  • Demo highlight: aa Lowe’s Innovation Lab agent could assess a DIY project, build a cart, and settle payment in USDC. All from one prompt.

  • Vision: AI agents buying, selling, trading services automatically, without humans serving as transaction middlemen.

The Fine Print

  • The Lowe’s scenario is a proof-of-concept. Some steps are simulated. Real world integration (merchant systems, inventory, shipping) will be harder.

  • Adoption is still thin. AP2 and x402 are open specs; getting merchants, agents, payment processors, and regulators comfortable will take time

  • Agents moving money autonomously invites attack vectors: malicious agents, abuse, runaway spending, fraud. Governance and safety must be baked in.

If this works, agents stop being just helpers. They become economic actors. Autonomous, transactional, real-time.

They could manage tasks like scheduling, sourcing, and contracting entirely on behalf of users. Pay for bugs found, pay for APIs used, pay for data fetched. And settle those transactions transparently and instantly.

Every time software can treat value as part of its logic (not just information), the architecture of commerce changes. We move from “pay with card” metaphors to “pay with purpose.”

Are you building something awesome in crypto × AI? Or spotted a startup or product that more people should know about? Fill in this form and share it with us. We’ll feature the sharpest picks in the newsletter each week.

💰 The Money Flow: Who Got Funded

  • Bio Protocol raised $6.9M from Arthur Hayes and Maelstrom Fund to push science with AI and crypto, building infra for open research and faster breakthroughs.

⚙️ Infra & Protocols

  • FLock launched its API Platform: devs get OpenAI-compatible tools to build AI apps, and model contributors earn through Moonbase.

  • Talus has introduced the Talus Foundation to drive growth and decentralization ahead of mainnet and TGE, backing early apps like AvA Gaming and AI prediction markets.

  • Prime Intellect’s Compute Exchange now offers Reserved Instances. You can book 8–1,000+ GPU clusters, compare quotes from 50+ providers, and re-sell unused GPUs.

  • Gensyn Introduces SAPO for Decentralized RL Training: GensynAI has launched SAPO, a decentralized reinforcement learning training method that enables a "swarm" approach to AI model optimization. 

  • Bittensor rolled out Sub-Subnets, letting each subnet run its own Yuma Consensus and bond pool for precise control of incentives without changing total rewards.

  • Akash is adding hardware verification via attestation, proving CPUs and GPUs with Intel TDX, AMD SEV-SNP, and NVIDIA NVTrust, pushing toward confidential computing.

  • Sentient’s ROMA repository has achieved the #1 trending spot globally on Github. It helps developers and enterprises to get better inferences, specially on long-horizon tasks

🤖 Agents & Apps in the Wild

  • Linera launched Testnet Conway, where you can log in, play the Game of Life, and rack up points by completing quests.

  • Dria unveiled Mem Agent (4B), a lightweight memory layer that makes LLMs stateful across apps like Claude, ChatGPT, and LM Studio. Runs locally via MCP for retrieval, updates, and context.

  • Warden opened its Agent Hub to third-party devs, letting LangChain builders monetize agents to 8.5M users (300K daily). Top agents already earn thousands daily, putting it among the largest AI agent marketplaces.

  • Vercel dropped x402-mcp, an open protocol that bakes payments into HTTP requests. Built on the 402 Payment Required standard, it lets AI agents tap paid services without setup.

  • Inspira Enterprise and Humans.ai launched H1uman, a 5-foot voice-activated AI robot built on blockchain tech. After a pilot with Romania’s government, it’ll roll out for employee outreach, then expand to citizen and enterprise use.

🌐 The Web2 Giants

  • Nvidia chip imports have been blocked in China, with firms pushed toward local options like Huawei, which plans new supercomputers by 2026–27.

  • Kling released Kling AI Avatar, letting you animate avatars with custom roles, voices, emotions, and expressions.

  • Perceptron released Isaac 0.1, a 2B-parameter open-source perceptive-language model from the team behind Meta’s Chameleon, built for real-world use in manufacturing, logistics, security, and robotics.

  • Alibaba launched Tongyi DeepResearch, the first fully open-source web agent rivaling OpenAI’s Deep Research. It hits SOTA results with just 30B (3B active) params.

  • OpenAI introduced GPT-5-Codex, tuned for agentic coding with faster performance, advanced code review, and real-time collab.

OpenAI just published data on how people actually use ChatGPT. And it’s not what you think.

Around 700 million people use ChatGPT weekly, firing off 2.6 billion messages a day, roughly 29,000/s

Personal use jumped from 53% to 73% in the past year. People ask it for dating advice, dinner ideas, life tips, conspiracy theories, you name it. It’s basically everyone’s weird, brilliant friend.

About 80% of conversations revolve around three things: getting help, finding answers, or writing stuff. Coding is just 4.2% of usage. Companionship and therapy barely crack 2%.

Nearly half of adult users are under 26. Growth is fastest in low- and middle-income countries. The early tech-bro skew has mostly vanished, ChatGPT is now everybody’s tool.

This week we’re diving into 404-Gen, a Bittensor subnet that turns text into 3D asset. Basically, Midjourney for world-building.

It’s already live and generating millions of objects. And like Midjourney, we believe it could go from $0 to $500M fast once monetization flips on. Here’s why👇

404–Gen: From Keyboard to World

Why the holy grail of text-to-3D may come not from Big Tech, but from a Bittensor subnet.

chainofthought.xyz/p/404-gen-from-keyboard-to-world

🔥 Our Weekly Top 5

#1: Zuck can’t catch a break.

Meta AI’s live demo of their new smart glasses failed twice at the Meta Connect event.

#2: Robots can now fight?!

This robot got right back up after it got knocked down. A bit scary no?

#3: Ethereum Foundation just launched a dedicated AI team.

The new “dAI Team” will make Ethereum the home for AI, helping agents pay, coordinate, and prove trust while shaping protocol upgrades for developers.

#4: You can now make 3D worlds from a single image.

This new model can generate persistent, expansive environments, and it’s now open in limited beta.

#5: AI art is getting better and better.

The creator used Qwen Edit to generate keyframes of a watercolor drawing, then animated the transitions with Wan 2.2. It’s so realistic it looks hand-drawn.

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