GM ā issue #50 is here!
Whether youāve been reading since day one or just jumped in, thanks for being part of the ride. š«”
Pluralis just pulled off a big first for the team: they trained a large language model across everyday internet-connected devices with no loss in speed or performance.
We've reached a major milestone in fully decentralized training: for the first time, we've demonstrated that a large language model can be split and trained across consumer devices connected over the internet - with no loss in speed or performance.
ā Pluralis Research (@PluralisHQ)
3:00 PM ⢠Jun 3, 2025
Hereās what they did:
They sliced the model into smaller chunks, spread them across GPUs, and used smart compression to keep everything in sync. That cut data transfer by over 100x. And it worked.
Now anyone, anywhere can pitch in compute to help train large models. More nodes, bigger models. Thatās the unlock.
Weāre seeing momentum pick up fast in decentralized training. Teams are pushing out real research, testing in the wild, and building on each other's progress. Over time, most of it will converge toward open systems with minimal technical moat. But the ones doing the hard work now are earning credibility, and locking in mindshare for whatever comes next.
Weāve also been thinking a lot about benchmarks.
If we want to track real AI progress, we need to map actual jobs. Break them down into atomic tasks. Measure performance. Iterate. Improve.
What does a lawyer actually do, step by step?
Same for teachers. Designers. Analysts.
This is where things are headed, turning chaotic real-world workflows into structured, trainable sequences.
Same goes for crypto AI agents. Props to teams building real benchmarks here. CAIA is testing how well LLMs can navigate blockchains and crypto. Even the top models like GPT o4 are barely hitting 35 percent.
Or Recall, which uses competitions to determine which agents can actually generate trading alpha. Right now, āMoonSage Alphaā is top of the leaderboard.
š New Podcast Alert!
CK and I spend so much time immersed in AI, we figured its time to talk about it!
Weāre launching Escape Velocity, a show that makes sense of AI, particularly whatās happening in web2 land (OpenAI, Gemini, AI doomers, and everything in between).
This is not your typical web3 AI podcast. We'll be inviting some of the top minds, including your favourite investors, to join us.
Pilot episode is live now. š«¶
š Nebulai just launched its Ambassador Program. Web3 creators can earn $NEB, get early access, and help scale its decentralized AI network to 10 million nodes.
š Donut is hiring. Theyāre looking for PMs, prompt engineers, AI researchers, editors, designers, and illustrators to push product development and grow the community.
š§Ŗ Saharaās SIWA testnet is live on Bitget Wallet. Users can earn daily $SAHARA testnet rewards by testing AI infrastructure and dapps.
Look at this beautiful chart. AI adoption is hitting escape velocity.
ChatGPT hit 400 million users in under two years, then doubled to 800 million by April 2025.
20 million are paying. For comparison, it took Spotify 17 years to cross that line. OpenAI is now closing in on $4 billion in annual revenue.
A few years ago, no one expected the next great consumer giant after Meta to be an AI lab. And yet, here we are.
300K AI inference calls in 24 hours. No centralized servers. Thatās what Atoma pulled off last week.
Atoma is a decentralized AI compute protocol built for verifiability, privacy, and scale. When you send a prompt, it doesnāt disappear into the black box of Big Tech infra.
It pings a global mesh of GPU nodes, many running inside Trusted Execution Environments (TEEs) so that even node operators canāt peek at your data.
Source: cloud.atoma.network
Everything runs on Sui. Smart contracts handle payments, coordination, and attestation. The setup supports OpenAI-style APIs, SDKs for Python/TypeScript, and dev tools that make integration feel familiar.
Want control? Youāve got it. Users can filter nodes by speed, cost, or privacy guarantees. Node performance and reputations are transparent, with staking and rewards aligning incentives.
Atoma isnāt competing with OpenAI. Itās building the backend for apps that need privacy, transparency, and decentralization built in from the start.
OpenGradient just launched the Nova Testnet, bringing AI model inference on-chain with full verifiability.
Introducing OpenGradient Nova Testnet
Today, we introduce the first testnet where a block records what happened, how it happened, and why.
Blockspace used to move numbers. Then logic. Now it moves meaning.
A thread on the third era of blockspace ā
(1/14)
ā OpenGradient (@OpenGradient)
7:10 PM ⢠Jun 3, 2025
Most AI models are hidden behind closed APIs, offering zero transparency. Nova makes every model output provable, traceable, and auditable directly on-chain.
Hereās how it works:
Inference runs on GPU nodes called Sprinters. Each result is bundled with the model, inputs, and outputs into a cryptographic fingerprint, then posted on-chain. Once a block is finalized, it locks in the full trail of how the AI reached that answer.
Novaās network is made up of four key roles:
Sprinters handle inference and generate proofs
Judges verify those proofs and finalize blocks
Librarians store and serve model weights
Scouts pull real-world data inside TEEs
Source: opengradient.ai
The compute layer, called PIPE, parallelizes inference tasks. Whichever node returns first wins, which keeps the chain efficient, even for heavy models.
If youāre building your own chain, OpenGradientās Neuro Stack lets you spin up AI-native rollups that tap into its inference and proof system. You keep your token and rules, and OpenGradient powers the verifiable AI layer underneath.
This turns black-box AI into public infrastructure. Agents, games, and dapps can now prove exactly what their models did, and why.
HeyElsa raises $3M to build natural language agents that execute directly onchain. Already live on Base with $15M+ processed and gasless execution.
3Jane pulled in $5.2M from Paradigm to launch Ethereumās first credit-based money market. Think unsecured USDC lines backed by real financial data, not collateral.
Grass announces Video Search, allowing you to search over a billion videos on the public web to find an exact moment or visual. Pretty awesome.
Heurist Deep Research is live. Itās like Bloomberg Terminal for Web3, built on x402 and powered by Mesh agents crawling 100+ sources. Reports drop for 1 USDC per query.
Bectra is now live on Berachain mainnet, bringing smart accounts, instant validator exits, and full Pectra-level execution upgrades, making Berachain the first L1 after Ethereum to go all-in on the EIP 7000+ era.
Rumi (a16z-backed) opened private beta. Users earn rewards by watching shows while training an AI to understand video content, frame by frame.
XMAQUINA has launched its DAO Portal, a real-time governance hub where $DEUS holders can track treasury allocations, vote on proposals, and oversee investments in robotics and Physical AI.
Sentient new Dobby Plus model is out. 8B parameters with sharper tone, tighter formatting, and cleaner context flow. Live now in Sentient Chat.
Exa Research is live. A search API powered by agents that finds answers fast, outputs clean structure, and hits 94.9% on SimpleQA. Itās faster and more cost-effective than many alternatives.
CryptoAgents #9001 and #9002 are now live as fully local AI companions in the new CryptoAgents iOS app, which runs offline, has no data sharing, and is iPhone 14+ only.
Re7 Capital, a leading DeFi investment firm, is now using Gizaās autonomous agents to manage on-chain liquidity. This is the first institutional rollout of AI-native treasury infrastructure.
Lilypad Network has joined the Olas Accelerator to build a research assistant agent for Pearl, the Agent App Store. It delivers custom insights in seconds and cuts research time from days to minutes.
Savant is live on TAO.app. Itās the first AI assistant built specifically for Bittensor, free to use and ready to answer anything about the network.
Yuma Consensus 3 is live on Bittensor, shifting validator rewards toward competitive miner discovery and enabling new features like crowdloans and controlled subnet token launches.
IOTA, the incentivized orchestrated training architecture, is now live on Bittensor Subnet 9 with real-time training of a 15B parameter model underway.
Backprop dropped a mobile wallet and trading app for Bittensor. Non-custodial, local key storage, password-protected, and available now at backprop.finance.
Sakana AI just released the Darwin Gƶdel Machine, a self-improving agent that rewrites its own code. It adapts in real time and outperforms static systems with evolving intelligence.
Anthropic dropped Claude Gov, a custom AI suite built for U.S. national security. Designed for classified environments with enhanced tools across defense, intel, and cyber.
Cursor 1.0 is live. It includes BugBot for code reviews, Background Agent for all users, Jupyter support, project memory, and one-click MCP setup. The dashboard also got a major upgrade with better visuals and smoother UX.
ElevenLabs launched Eleven v3 (alpha), their most expressive TTS model yet. Now supports over 70 languages, multi-speaker dialogue, and emotional tags. Available to the public at 80 percent off through June.
Lumaās Modify Video tool is now live in Dream Machine. Creators can restyle scenes, animate characters, or swap entire worlds without losing motion or performance fidelity.
Weāre celebrating a year of research with a new series breaking down what weāve uncovered at the edge of Crypto and AI.
š Unlock Progress: 5 of 10 essays now live
This weekās focus: Decentralized training is the highest-upside bet in Crypto x AI right now. Hereās everything weāve learned so far š
50K jobs. $6B saved. These agents leave web3 agents in the dust.
Weāre excited to introduce Retool Agents, your first fully autonomous AI team.
Already, our customers have automated 50k jobs and saved $6B in manual work. This is in every department, including product, finance, operations, support, etc.
The secret? Your existing hyperspecific
ā David Hsu (@dvdhsu)
2:30 PM ⢠May 28, 2025
Will China or US win the robotics race?
š¤China ramps up humanoid robot race: 50+ startups, 10K units/year, $138B in planned investments, challenging Tesla's lead with real-world deployments across factories, hospitals, and military.
ā China now has over 50 startups building humanoid robots. Key players include
ā Rohan Paul (@rohanpaul_ai)
11:17 PM ⢠Jun 1, 2025
A map of the Bittensor Ecosystem
The Bittensor Ecosystem Map
Bittensor has quickly become the most exciting frontier in AI x Crypto.
Hereās your definitive map of the builders shaping its ecosystem.
1. AI Data, Training, Inference
@omegalabs_bt - Model datasets
@MacrocosmosAI - Data collection & storageā Cipher (@CipherResearchx)
11:50 PM ⢠Jun 3, 2025
Some tips on vibe coding
I promise you youāre vibe coding wrong
as someone who has built multiple production-ready applications, with thousands of users, from just Cursor with minimum intervention.
But first here's you (probably):
You open Cursor. Type ābuild me X.ā
It spirals. Nothing works. Youā vas (@vasumanmoza)
1:25 AM ⢠May 18, 2025
OpenAIās court order says it allāprivate AI has never mattered more. RIP privacy.
OpenAI is now required by court order to preserve all ChatGPT logs including "temporary chats" and API requests that would have been deleted
if I understand this correctly, it means data retention policies for apps that use OpenAI API simply cannot be honored
ā kepano (@kepano)
11:18 PM ⢠Jun 4, 2025
Thatās a wrap for this week! Got thoughts, feedback, or something cool to share? Just hit reply, we read it all.
šØ Before you go:
Are you in our telegram channel yet? We drop daily AI alpha.
Catch our podcast summaries too, complete with detailed show notes.
And if you want to jam with us directly, hop into our Discord
Cheers,
Teng Yan & Issy
Did you like this week's edition? |
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